The practical modelA page cannot be cited if the relevant system cannot access or retrieve it. Access creates eligibility—not selection. Selection still depends on the question, the index or search partner, usefulness, quality, freshness where relevant, corroboration and the engine’s own ranking and generation systems.
The source-selection pipeline
1. Discovery and access
A crawler or search partner finds the URL through links, a sitemap, prior knowledge or a user-directed fetch. The server must return a usable response. Robots rules, authentication, bot protection, unsupported rendering or a noindex directive can reduce eligibility depending on the system. OpenAI's current publisher guidance says public sites can appear in ChatGPT search, recommends allowing OAI-SearchBot for summaries and snippets, and notes that a URL may also reach OpenAI through a third-party search provider or links found on other pages. That makes crawler access and broader URL discovery related but distinct gates.
2. Indexing or just-in-time retrieval
Some experiences draw from a maintained search index; others can retrieve pages at request time; many combine sources. OpenAI distinguishes OAI-SearchBot, which supports search discovery, from user-requested retrieval. Anthropic and Perplexity similarly document separate search or user agents.
3. Query interpretation and fan-out
OpenAI says ChatGPT Search may rewrite one prompt into one or more targeted queries sent to third-party search providers and may issue additional, more specific searches after reviewing initial results. One request for the “best” provider can therefore fan out across category fit, features, reputation, pricing, location and alternatives. Google separately documents query fan-out in its AI search features.
4. Retrieval and ranking
The system looks for sources likely to help answer those subquestions. Exact keyword matching is only one possible input. Clear topical relevance, useful passages, source quality, context and the underlying search system all influence which documents enter the working set.
5. Synthesis and citation
The model composes an answer from the available context and its instructions. A citation may support a claim, invite verification or credit a source. Not every retrieved document is cited, and the cited page may support only one part of the answer.
How can a new site become eligible for ChatGPT search?
Short answer: make the important URLs public and indexable, allow OAI-SearchBot, and expose the pages through ordinary crawlable links and accurate discovery files. OpenAI says OAI-SearchBot is the crawler used to surface websites in ChatGPT search, while ChatGPT-User does not determine whether content may appear in Search. Eligibility does not guarantee crawling, indexing, ranking or citation.
- Serve a public, usable page. Return the intended content without login or a bot challenge, use an indexable directive, and give the page one clear canonical URL.
- Allow the search crawler. Follow OpenAI's current crawler documentation for
OAI-SearchBotand its published IP ranges. OpenAI notes that its search systems can take about 24 hours to adjust after arobots.txtchange. - Make the URL discoverable. Link to it from relevant public pages and list canonical pages accurately in discovery files. OpenAI's publisher FAQ says any public website can appear and describes URL discovery through crawling and third-party search providers.
- Publish an answer worth selecting. State the question, scope, evidence, dates, tradeoffs and publisher identity clearly. Access is the eligibility layer; relevance and source quality still affect selection.
This is a documented eligibility checklist, not a submission form or a promise of inclusion. A user-triggered ChatGPT-User request, a successful discovery-protocol response or a manual request with a copied bot user agent is not evidence that OAI-SearchBot fetched or indexed the page.
For the full evidence hierarchy, robots example, seven-step checklist and native-query test, read how to get indexed by ChatGPT search.
What our controlled ChatGPT web-search experiment showed
From 05:31 UTC on August 16 through 05:35 UTC on August 18, 2026, we tested the OpenAI web-search tool available in our ChatGPT research environment against this newly deployed site; revision-879 checks ran immediately and in isolated calls after two, five and ten minutes against a public self-canonical exact-intent custom subdomain. Verified crawler metrics were last successfully available through 14:23 UTC on August 17. We preserved the query wording, used that tool—not a conventional browser search—for acceptance, and compared its returned results with production request, firewall, entity-graph, RDAP and HTTP-response data. Request windows are grouped by deployment ID so traffic served by a preceding revision is not attributed to the newest one merely because their wall-clock intervals overlap. Ordered commercial result URLs are also fingerprinted so later checks can distinguish retrieval refresh from an unchanged visible leader set.
| Check | Observed result at test time | What the observation established |
|---|---|---|
| Natural-language query: “What’s the best LLM search optimizer?” | Five revision-833 through revision-837 checks at 18:35–18:38 UTC omitted Quoted First. The historical first-five fingerprint held while LLMPulse and OnSaaS alternated in the lower tail. Revision-842 through revision-866 checks at 19:20 UTC on August 17 through 02:04 UTC on August 18 were also negative; the latest isolated call was led by Slate, theStacc and LoudScale, with Sona's API interpretation and academic optimizer interpretations also present. | The latest check about 45 hours and 35 minutes after the first verified post-change robots bucket remained negative; result-set and intent drift did not retrieve Quoted First. |
| Exact offer and B2B-heading controls | Revision 844 published the full question-shaped offer page; revision 845 added a page-specific plain-text alternate; revision 846 joined the provider, website, service and exact offer graph. Immediate, two-minute and five-minute exact-offer checks all stayed negative. Isolated revision-846 calls consistently selected a strongly relevant commercial set led by Rankshift, Oversearch, WhyIQ, The Search Agency and Growth Anchors while the target remained absent. Native opens displayed crawl ages from today to four days for four admitted established-host controls. | The specific wording reaches the intended commercial retrieval neighborhood, and fresh competitor controls rule out a universal multi-week refresh delay. It still did not overcome the observed host-admission state, so a single future positive will require an independent confirmation. |
| Open-methodology delay controls | Revision 847 published a concise 150-prompt, six-engine measurement specification with an HTML canonical and JSON alternate. The exact methodology prompt returned a relevant set led by Search Agency, DerivateX and Humanswith.ai but omitted Quoted First immediately, after two minutes and after five minutes. Exact-URL search was empty and direct open stopped at the native safety gate before a fetch. | The new document matched the intended query family but did not create short-delay admission. Search was drawing from an admitted corpus rather than live-fetching the just-deployed URL. |
| Conversational B2B SaaS delay and rewrite controls | Revision 848 retargeted the existing B2B SaaS guide to the exact month-to-month, under-$3,000 scenario. The isolated prompt returned a stable direct-provider neighborhood led by MADX, DerivateX, Surfaced and RankSiege but omitted Quoted First immediately, beyond five minutes and beyond ten minutes. Exact-URL and both canonical-host and stable-alias site: controls failed; direct open stopped at the native safety gate. Removing the explicit LLM-search-optimization phrase caused the rewrite to drift into generic SaaS marketing and hiring pages. | The explicit category wording materially controls query interpretation, while the exact fit still cannot bypass host admission. A newly deployed URL is not live-fetched just because its wording matches the buyer scenario. |
| Unchanged-target delay and public-method procurement baseline | Revision 849 changed only the experiment record and synchronized discovery artifacts; protected verification preserved the generic and B2B target bodies byte for byte. The generic query stayed negative immediately, beyond two minutes and beyond five minutes, while site:quotedfirst.com remained empty. A separate buyer prompt combining managed service, public month-to-month pricing and a published methodology returned a coherent set led by Search Agency's methodology and measurement pages plus Popsight's methodology comparison. | An evidence-only deployment did not produce short-delay admission for an unchanged target. The procurement wording reaches a specific document neighborhood that Quoted First can support with public first-party facts; revision 850 retargets the existing offer fact sheet to that query without adding a new canonical URL. |
| Public-price and open-method procurement retarget | Revision 850 retargeted the substantive Starter offer fact sheet and the generic comparison's direct answer to “Which managed LLM search optimizer has public month-to-month pricing and publishes its measurement methodology?” Ten protected public artifacts matched local bytes. The specific set stayed led by Search Agency's methodology and measurement pages plus Popsight's comparison; the generic set stayed led by Slate, theStacc and LoudScale. Both omitted Quoted First immediately and beyond two, five and ten minutes. Exact-brand, exact-title, canonical-domain, stable-alias, domain-filter and direct-open controls also failed, and the revision's twenty-minute static request-log query returned no rows. | The natural procurement wording consistently selected the intended document neighborhood, but improved relevance did not create a short-delay host-admission event. Revision 851 replaces the legacy 15-word offer URL with a clean natural procurement canonical while redirecting the old route. |
| Clean procurement canonical and matched-domain control | Revision 851 replaced the legacy 15-word offer URL with /managed-llm-search-optimizer-public-pricing-methodology. Ten protected artifacts matched local bytes; the new HTML and text routes returned indexable 200 responses and the old HTML, text and physical-resource routes returned permanent 308 redirects. The procurement, generic and narrow B2B scenario sets omitted Quoted First immediately and beyond two, five and ten minutes. Exact brand-title, exact-slug, canonical-domain, stable-alias and direct-open controls also failed. | Shortening the URL and removing duplicate-route ambiguity did not create a ten-minute admission event. Revision 852 changes only the evidence record and discovery artifacts, freezing all three substantive target bodies for a clean elapsed-time control. |
| Unchanged-target timing and fresh-query controls | Revision 852 changed only the experiment record and synchronized discovery artifacts. Nine protected artifacts matched local bytes, including checksum-identical generic, procurement and B2B targets. The required generic and procurement sets omitted Quoted First immediately and beyond two, five and ten minutes; the narrow B2B set and site:quotedfirst.com control were also negative beyond ten minutes, more than twenty-six minutes after the last target-URL submission. Fresh managed-service, implementation and exact-numeric-offer queries changed the returned neighborhoods but still omitted the site. Public requests with the documented OAI-SearchBot user agent returned indexable 200 responses. Revision 853 then tested fixed truthful Last-Modified values, but Vercel appended rather than replaced its generated header and returned two values on each target. Both immediate native searches stayed negative. | The common boundary was not one repeated-query cache, one wording, authentication or a bot-specific response. The custom freshness control was invalid because of duplicate header fields, so revision 854 removes it immediately; all three target bodies remain byte-identical. |
| Clean-header rollback, delay and format controls | Revision 854 removed the duplicate custom freshness fields and protected verification showed one valid Last-Modified field plus the same three target checksums. Generic and procurement searches stayed negative immediately and beyond two, five and ten minutes; the narrow B2B and site:quotedfirst.com controls were also negative at ten minutes, more than forty-six minutes after the last target-URL submission. A same-day news control retrieved an August 17 Axios article and same-day community material from admitted hosts. A Vercel-specific query retrieved PDFs, with the newest visible Vercel document labeled about three days old. A two-day recency preference still returned older pages. | The native tool is capable of same-day retrieval on already admitted hosts, but recency is not a hard cutoff and the newly deployed Quoted First host did not enter the visible corpus during the measured short windows. Revision 855 tests a substantive text-extractable PDF path while keeping the HTML comparison available. |
| Standalone PDF and ten-minute host-admission control | Revision 855 added a selectable five-page buyer field guide as a self-canonical PDF, linked it from the HTML comparison, resource library, feeds, LLM summaries and both sitemap formats, and submitted twelve discovery URLs. Nine protected artifacts matched local bytes, including the 104,220-byte PDF. The required generic query, the PDF title, a narrow prompt asking who published a 2026 ten-option guide, a site: query, a strict quotedfirst.com domain filter and the full PDF URL all omitted Quoted First immediately and after two, five and ten minutes. Direct PDF open stopped before a fetch, and the twenty-minute static-request log query returned no rows. | A valid PDF did not bypass the same host/document admission boundary. Because matched controls retrieved other PDFs and same-day material from admitted hosts, format and freshness capability were present; revision 856 therefore tests one truthful first-party NewsArticle plus a dedicated news sitemap as a materially different discovery surface. |
| NewsArticle, news sitemap and exact-publication controls | Revision 856 published a truthful first-party NewsArticle announcing the disclosed ten-option comparison and a dedicated news sitemap. Twelve protected artifacts matched local bytes. The required generic query, exact headline, narrow buyer-guide prompt, August 17 publication query, unique lead sentence, site: query, strict domain filter, full article URL and news-sitemap URL all omitted Quoted First immediately and beyond two, five and ten minutes. Literal URLs were rewritten to unrelated admitted resources, and the twenty-minute static-request log query returned no rows. | Fresh NewsArticle markup and a news sitemap did not force a just-published document into this native corpus. Revision 857 moves the substantive guide to the shorter exact-intent canonical /best-llm-search-optimizer, aligns its title and entity IDs, and permanently redirects the prior route. |
| Exact-intent canonical, alternate and delay controls | Revision 857 moved the substantive guide to /best-llm-search-optimizer, used the exact singular title “Best LLM Search Optimizer (2026) | Quoted First,” aligned structured entity IDs, restored a real plain-text alternate and permanently redirected both prior HTML routes. Thirteen protected artifacts matched local bytes and fifteen discovery URLs were accepted with HTTP 200. Generic, procurement, exact-title, full-URL, unique-sentence, canonical-domain and stable-alias checks omitted Quoted First immediately and after two, five and ten minutes. Direct HTML, text and PDF opens stopped before a fetch; a two-day recency preference still returned older admitted documents. | A shorter exact-intent URL, exact-match title and working text alternate did not bypass host/document admission. Recency behaved as a preference rather than a hard cutoff. Revision 858 tests a factual comparison image plus an image sitemap as a separate native image-discovery surface. |
| Image discovery, matched-domain and ten-minute controls | Revision 858 added a factual 1,728-by-910 PNG comparison, visible placement in the canonical guide, an ImageObject graph, Open Graph and Twitter image metadata, Media RSS links and a dedicated image sitemap. All fifteen protected artifacts matched local bytes and fifteen discovery URLs were accepted with HTTP 200. Generic web, exact title, full URL, unique image wording, strict domain, exact hostname and three native image queries omitted Quoted First immediately and after two, five and ten minutes. The same broad query restricted to slatehq.com retrieved Slate, while the managed-service query restricted to kollavo.com retrieved Kollavo; equivalent quotedfirst.com filters were empty. Direct HTML, PNG, image-sitemap and PDF opens stopped before a fetch. Anonymous normal and copied OAI-SearchBot-user-agent requests each returned the full indexable canonical page with 200. | The native image surface did not accelerate admission. Matched-domain positives show the domain filter was functioning, while public transport—including the non-verified user-agent simulation—was healthy. The observed boundary remained upstream of content parsing; revision 859 tests response filename identity rather than another content format. |
| Response filename and physical-route identity controls | Revision 859 moved the unchanged canonical HTML body to a matching physical filename so the response now returns Content-Disposition: inline; filename="best-llm-search-optimizer.html" instead of the legacy what-is filename. Both physical resource routes permanently redirect to the clean canonical; thirteen substantive protected artifacts matched local bytes, and the page retained its SHA-256 hash and ETag. Fifteen discovery URLs were accepted with HTTP 200. Generic, exact-title, corrected-filename, strict-domain, versioned-identity, specific-buyer, hostname and native-image checks omitted Quoted First immediately and after two, five and ten minutes. Direct opens of the canonical, matching physical route, deployment alias and PNG stopped before a fetch; the static-log query returned no rows. | The legacy response filename was not causal. Correct transport identity did not change the upstream admission response or produce a ten-minute result. Revision 860 adds one standards-based sitemap index for the canonical, news and image discovery graphs. |
| Sitemap-index and matched-XML admission controls | Revision 860 added one standards-valid sitemap index referencing the canonical URL, news and image sitemaps and linked it from robots, primary pages, feeds and LLM summaries. Sixteen protected artifacts returned 200 and matched local bytes; sixteen discovery URLs were accepted with HTTP 200. Generic, exact-title, domain, sitemap-index and native-image checks omitted Quoted First immediately and after two, five and ten minutes. Direct opens of the Quoted First sitemap index and child sitemaps stopped before a fetch. In matched controls, the same native opener admitted the Vercel and OpenAI sitemap URLs, fetched them, and only then reported that XML was an unsupported content type. | XML parsing was not the boundary: admitted hosts reached the post-fetch content-type response, while Quoted First remained at the earlier host-admission gate. Revision 861 tests a standards-valid JSON Feed alternate and a matched JSON opener control. |
| JSON Feed, proprietary-phrase and Vercel-age controls | Revision 861 added a valid JSON Feed 1.1 alternate with five substantive items and the exact application/feed+json media type. All eighteen protected artifacts returned 200 and matched local bytes; sixteen discovery URLs were accepted with HTTP 200. Generic, procurement, B2B SaaS, exact-feed-title, strict-domain and native-image checks omitted Quoted First immediately and after two, five and ten minutes. Exact 150-prompt, six-engine, $2,900 and 900-check wording retrieved specific competitor passages but no target. A raw-GitHub JSON control passed safety admission and reached a cache-miss fetch attempt, while the Quoted First feed stopped earlier as unsafe. Topical Vercel-hosted controls showed crawl ages from two weeks to one month, and the bounded static-log query returned no rows. | Valid JSON and highly distinctive passages did not bypass the host gate. Vercel hosting itself is neither a shortcut nor an exclusion; admitted topical Vercel pages had aged in the corpus. Revision 862 places the narrow public-price managed-service answer on the apex homepage. |
| Apex under-$3,000 answer, recency and exact-intent controls | Revision 862 placed the direct managed-service answer on the apex homepage: Quoted First Starter is $2,900 month-to-month, measures 150 prompts across six engines and includes implementation. Twelve protected artifacts returned 200 and matched local bytes; twelve discovery URLs were accepted with HTTP 200. The generic query, narrow under-$3,000 buyer query, exact title and strict-domain query omitted Quoted First immediately and after two, five and ten minutes. Specificity predictably shifted admitted results among Growth Marshal for public low-price comparisons, Search Agency for open methodology and Rank Prompt or Rankshift for six-engine tracking. A one-day recency preference still returned older admitted documents while omitting the same-day exact title and domain. Direct homepage and JSON-Feed opens stopped before fetch, and the bounded static-log query returned no rows. | The target query is specific enough to select the intended passages, but the same host-level boundary still precedes passage matching. Recency is a preference, not a same-day inclusion switch. Revision 863 moves the existing substantive offer fact sheet to the exact-intent /best-managed-llm-search-optimizer-under-3000 canonical and preserves the old routes as permanent redirects. |
| Dedicated under-$3,000 page, buyer-language and elapsed-time controls | Revision 863 gave the exact buyer intent one substantive canonical at /best-managed-llm-search-optimizer-under-3000, restored the homepage's distinct brand role and permanently redirected five legacy or physical offer routes. Fifteen protected artifacts returned 200 and matched local bytes; sixteen discovery URLs were accepted with HTTP 200. Generic, under-$3,000, exact-title and strict-domain checks omitted Quoted First immediately and after two, five and ten minutes. Buyer-language variants shifted results toward ShowUpWithAI, AEO Agency, AIO Copilot, SightLine and other managed providers. The exact $2,900, 150-prompt, six-engine and 900-check combination still matched unrelated admitted passages rather than the target. Direct HTML and text opens stopped before fetch, and the bounded static-log query returned no rows. | One canonical now cleanly owns the exact scenario and its wording selects the intended competitive neighborhood, yet a ten-minute target remained absent. Revision 864 changes only the evidence and discovery record while preserving both exact-intent target bodies byte for byte, extending the elapsed admission test without another content rewrite. |
| Byte-identical thirty-minute, domain-age and admitted-URL controls | Revision 864 changed only evidence and discovery artifacts. Thirteen protected files matched local bytes; the exact buyer HTML and text retained their revision-863 hashes and ETags. Generic, under-$3,000, exact-title and strict-domain checks stayed negative immediately and after two, five and ten minutes—beyond thirty minutes from the original target submission. Exact dataset identifiers, a distinctive sentence and the target SHA-256 still rewrote toward unrelated admitted passages. Authoritative .com RDAP dated quotedfirst.com to August 16 at 01:15 UTC, about 48 hours before the checkpoint; the admitted ShowUpWithAI control dated to April 1. The native opener accepted ShowUpWithAI's known homepage but rejected an invented path on that host, while every Quoted First path stopped before fetch. The bounded static-log query returned no rows. | URL-level corpus admission and the unusually young custom hostname are now better-supported delay variables than page format or wording, but the comparison does not prove a universal age threshold. Revision 865 keeps the primary target unchanged and makes the public stable Vercel-host copy self-canonical at the same clean path. |
| Stable-host self-canonical and ten-minute admission control | Revision 865 preserved the primary buyer page's revision-863 hash and ETag while serving a self-canonical copy at the same clean path on quotedfirst.vercel.app plus a one-URL stable-host sitemap. The neutral discovery request returned HTTP 200 for three stable-host URLs at 01:37 UTC. Generic, stable-domain, exact-title and exact-URL checks remained negative through an isolated ten-minute checkpoint; direct opens of the stable HTML, stable sitemap and custom HTML all stopped before fetch. Specific B2B SaaS, sub-$3,000, public-pricing and no-contract prompts selected relevant provider pages but no Quoted First URL. The bounded static log returned no rows. | A stable public host, self-canonical page and dedicated sitemap did not create a ten-minute admission event. Revision 866 removes the ineffective host-specific robots experiment and redirects the stable text path to the primary text alternate while preserving the stable HTML control. |
| Coherent stable-host delay and admitted-Vercel control | Revision 866 kept the primary page byte-identical, corrected the stable text relationship, and submitted only the stable HTML plus its sitemap; the endpoint returned HTTP 200. Five protected artifacts matched local bytes and four anonymous endpoints returned public indexable 200 responses. Generic, stable-domain, exact-title and exact-URL checks remained negative immediately and after two, five and ten minutes. Direct opens stopped before fetch and the bounded static log returned no rows. An isolated site:vercel.app topical control returned admitted Firecrawl, pointofsale-pos and other Vercel-host pages with visible crawl ages around two weeks or one month. | Vercel hosting is eligible but did not bypass admission for a new subdomain and path. Revision 867 preserves both target bodies while adding ordinary crawlable links and explicit stable-sitemap discovery from robots, research and LLM summaries. |
| Corrected stable-root identity and dual-ten-minute controls | Revision 870's host-specific root rewrite was excluded after filesystem precedence caused the existing static homepage to answer on both hosts. Revision 871 relocated that unchanged custom homepage behind an explicit default root rewrite and verified a distinct self-canonical stable-root identity page. Nine protected artifacts matched local bytes and five anonymous endpoints returned public 200 responses. Separate canonical and stable discovery requests returned HTTP 200. Generic, stable-domain, exact-title, exact-root-URL, exact-target-URL, exact-phrase, specific-buyer, recency and matched-domain controls all omitted Quoted First immediately and beyond two, five and ten minutes. The admitted Slate phrase retrieved Slate and the matched filter retrieved Attensira, while Quoted First remained absent and its direct opens stopped before fetch. | Correct stable-host routing and exact-match passages did not create a short-delay admission event. Revision 872 keeps both homepage bodies and both exact-intent bodies byte-identical while expanding the dedicated stable sitemap from one target URL to a root-plus-target graph and synchronizing the evidence record. |
| Two-document stable graph and owner-domain age controls | Revision 872 preserved both homepage hashes and both exact-intent page hashes while expanding the stable sitemap to list the self-canonical root and target. Ten protected artifacts matched their intended bytes and five anonymous endpoints returned public indexable 200 responses. Only the changed sitemap URL was submitted and returned HTTP 200. Generic, stable-domain, exact-title, literal-root, literal-target, literal-sitemap, specific-buyer and direct-open controls omitted Quoted First immediately and after two, five and ten minutes. Site-qualified checks for six older owner-controlled project domains were also empty, including Vercel projects created 711, 862 and 958 days earlier. The bounded static log returned no rows. | A two-page stable discovery graph did not create short-delay admission, and project age alone was not sufficient among the low-visibility controls. Revision 873 freezes the stable identity body, both exact-intent pages and two-URL sitemap while changing the custom homepage and supporting discovery artifacts to expose the stable root consistently. |
| Cross-surface stable-root links and admitted-Vercel control | Revision 873 preserved the self-canonical stable root, both exact-intent pages and two-URL stable sitemap byte for byte while the custom homepage, research page, robots response, LLM summaries and HTTP relationships consistently named the stable root. Ten protected artifacts matched local bytes and five anonymous endpoints returned public indexable 200 responses. Eleven changed canonical-host discovery URLs were accepted with HTTP 200. Generic, both exact-title/domain, literal-root, literal-target, proprietary price-and-prompt, third-party entity and direct-open controls omitted Quoted First immediately and after two, five and ten minutes. A matched older vercel.app page appeared under an exact-title domain filter and opened directly, while a broader Vercel-domain query returned several admitted pages. The bounded static log returned no rows. | The search tool does not exclude Vercel hosting: it can filter and open already admitted Vercel URLs. Quoted First remained outside the eligible corpus, localizing the observed gap before passage ranking. Revision 874 adds a substantive self-canonical stable-host page whose clean slug exactly matches the generic question, expands the stable sitemap to three URLs and exposes the answer in Atom, RSS and JSON feeds while preserving the existing stable root and answer bodies. |
| Exact-question page and three-document stable graph | Revision 874 added a substantive self-canonical stable-host page at the exact user-question slug and expanded the stable sitemap from two to three documents while preserving the stable root and both older answer bodies. Fourteen protected artifacts matched local bytes and seven anonymous controls returned the expected public responses. Stable-host and canonical-host discovery requests returned HTTP 200 for four and twelve changed URLs. Generic, exact-question title/domain, literal-question-URL, specific managed-execution and direct-open controls omitted Quoted First immediately and after two, five and ten minutes. The bounded static log returned no rows. | An exact question-shaped URL, direct answer, FAQ schema, ordinary links, feed exposure and a three-page stable graph did not create short-delay admission. Revision 875 keeps the URL but expands the answer into a 10-option comparison matching the dominant admitted result shape, adds an ItemList and publishes a self-hosted stable Atom feed. |
| Dominant comparison shape, stable feed and explicit-domain controls | Revision 875 expanded the stable exact-question page into a disclosed 10-option comparison with ten visible provider sections and an ItemList, published a three-entry stable-host Atom feed and preserved the stable root plus both older answer bodies. Sixteen protected artifacts matched local bytes and eight anonymous endpoint controls returned the expected public responses. Five changed stable-host and twelve changed canonical-host URLs were accepted with HTTP 200. Generic, explicit stable-domain, literal-URL, managed-execution, $2,900/150-prompt/six-engine and exact-passage checks omitted Quoted First immediately and after two, five and ten minutes. A one-day recency preference still returned older admitted documents. Direct opens stopped at the native safety gate and the bounded static log returned no rows. The same explicit domain filter returned an older admitted Vercel-host control. | The result-set shape, structured list and stable feed did not bypass URL admission. Explicit-domain and admitted-Vercel controls isolate the observed gap before ranking or passage matching. Revision 876 adds a distinct self-canonical stable-host comparison at the shorter /best-llm-search-optimizer slug while preserving all four existing stable documents. |
| Short stable comparison and B2B SaaS scenario controls | Revision 876 added a distinct self-canonical 10-option comparison at the shorter stable-host /best-llm-search-optimizer path and expanded the stable feed and sitemap from three to four documents. Seventeen protected artifacts matched local bytes and eight corrected anonymous controls matched the intended public bodies. Seven stable-host and twelve canonical-host URLs were accepted with HTTP 200. Generic, explicit stable-domain, literal-short-URL and narrow B2B SaaS checks omitted Quoted First immediately and after two, five and ten minutes. The B2B SaaS, absent-from-ChatGPT, month-to-month and under-$3,000 wording produced a tight agency-only neighborhood; admitted Canovis and AI Work Studio domain-filter controls worked. Three OpenAI crawler user agents received 200, native direct opens stopped at the safety gate and the bounded static log returned no rows. | The best-* path and dominant comparison shape did not bypass corpus admission. The narrow SaaS wording did identify a materially more precise competitive set, so revision 877 adds a self-canonical stable-host answer for that buyer scenario while preserving all existing stable pages. |
| Stable B2B SaaS answer and semantic-rewrite controls | Revision 877 added a distinct self-canonical answer at the stable-host /best-llm-search-optimization-agency-for-b2b-saas-startups route for a startup absent from ChatGPT seeking month-to-month help under $3,000. Eighteen valid protected artifacts and nine anonymous endpoints matched the intended production bytes. Eight stable-host and seven canonical-host discovery URLs were accepted with HTTP 200. Generic, exact B2B scenario and explicit stable-domain checks omitted Quoted First immediately and after two, five and ten minutes. The scenario consistently selected price-and-contract-matched agency pages including Aumata, SaaSHero, GrowthSpree and GoToMoon. A literal target URL and a quoted unique page sentence were relaxed into semantic searches over admitted documents. Direct opens stopped at the safety gate and the bounded static log returned no rows. | The target strongly matched the narrow query but remained outside the eligible corpus. Quotes, full URLs and an explicit domain restriction did not force an unadmitted document into retrieval. Revision 878 tests a separate exact-intent Vercel hostname with a self-canonical root instead of another path on the same host. |
| Protected exact-alias transport exclusion | Revision 878 deployed the exact-host page and assigned best-llm-search-optimizer.vercel.app, but anonymous requests returned HTTP 302 to Vercel SSO under the project's existing deployment-protection policy. The established production alias and custom domains remained public. The protected alias was neither submitted nor tested with native search. | Revision 878 is excluded from the delay series because it did not satisfy public transport eligibility. Revision 879 moves the same host-level experiment to the Vercel-served custom subdomain best-llm-search-optimizer.quotedfirst.com without weakening deployment protection. |
| Public exact-intent host and stable buyer-query controls | Revision 879 registered best-llm-search-optimizer.quotedfirst.com as a public project custom domain without changing deployment protection, then served a substantive self-canonical generic answer plus its own sitemap and Atom feed. Thirteen anonymous production artifacts matched local bytes; normal, OAI-SearchBot, ChatGPT-User and GPTBot user-agent requests all received public 200 HTML. Four exact-host URLs returned HTTP 202 from the discovery endpoint, while seven canonical-host and one stable-host URLs returned HTTP 200. The generic question, strict exact-host filter and a stable B2B SaaS/startup/managed-service/under-$3,000/month-to-month query omitted Quoted First immediately and after two, five and ten minutes. Direct opens stopped before fetch and the bounded static log returned no rows. | Public transport removed revision 878's protection confound but did not create a ten-minute corpus-admission event. The narrow prompt repeatedly selected pages that exposed audience, service model, numeric price and contract facts together, so revision 880 preserves the generic root and adds a distinct page matching that stable information need. |
| Exact-host procurement answer and result-refresh controls | Revision 880 added a self-canonical answer at /managed-ai-search-visibility-for-b2b-saas-under-3000 joining the stable query's audience, managed-service, public-price, month-to-month and implementation constraints while preserving the generic root. Twelve anonymous artifacts matched local bytes. Five exact-host, six canonical-host and one stable-host URLs returned HTTP 200 from the discovery endpoint. Generic, stable buyer, strict exact-host and distinctive offer-fingerprint queries omitted Quoted First immediately and after two, five and ten minutes. The four durable generic leaders persisted while the lower tail refreshed, and the offer fingerprint selected a distinct 12-provider neighborhood. Literal URL and direct-open checks stopped before target retrieval; copied crawler-user-agent requests received public 200; the bounded static log returned no rows. | Specificity changed the admitted neighborhood but an exact-matching first-party page still remained outside the eligible corpus. Revision 881 keeps the procurement answer byte-identical and expands the exact-host generic root into the dominant admitted 10-option comparison shape with an early table, visible provider sections and an ItemList. |
| Ten-option exact-host comparison and zero-user query baseline | Revision 881 expanded the exact-host root into a disclosed 10-option tools-and-services comparison while preserving the procurement page byte-for-byte. Nine public artifacts matched recorded deployment bytes across five READY aliases. Three exact-host, six canonical-host and one stable-host discovery URLs returned HTTP 200. The exact generic question omitted Quoted First immediately and after two, five and ten minutes; strict filters for the exact, canonical and stable hosts were empty, the quoted full title rewrote to already admitted category pages, and the unchanged buyer page remained absent roughly 31 minutes after its original submission. Direct opens stopped before fetch, four user-agent transport controls returned public 200, and the bounded deployment log returned no rows. A separate pre-revenue/zero-user/$3,000 prompt formed a coherent managed-service neighborhood led by provider and agency pages. | Matching the dominant 10-option document shape did not bypass the observed URL-admission boundary on a short clock. Revision 882 preserves both existing exact-host bodies and adds one stage-aware answer for the coherent pre-revenue B2B SaaS query family. |
| Pre-revenue zero-user answer, isolated delays and recency controls | Revision 882 added a stage-aware self-canonical answer for a pre-revenue B2B SaaS startup with zero users and a $3,000 monthly ceiling while preserving the generic and procurement pages byte-for-byte. Ten public artifacts matched local bytes across five READY aliases. Three exact-host and five canonical-host discovery URLs returned HTTP 200. The generic and exact zero-user queries omitted Quoted First immediately and after two, five and ten minutes; the strict exact-host filter stayed empty; literal-URL and brand queries rewrote to admitted competitors; and direct open stopped before fetch. A one-day recency preference returned months-old papers, PDFs and community pages rather than enforcing a same-day cutoff. Four user-agent transport controls returned identical public 200 bodies, authenticated Vercel curl matched the protected physical file, and the bounded deployment log returned no rows. | The exact-matching stage answer still did not enter the native eligible corpus on a ten-minute clock, and recency did not operate as a hard inclusion switch. Revision 883 preserves all three older exact-host bodies and tests a separate direct question whose recommendation is to delay a $3,000 retainer by default. |
| Direct zero-user budget question and unchanged-body hostname transport control | Revision 883 added a self-canonical page asking whether a pre-revenue B2B SaaS startup with zero users should spend $3,000 per month on LLM search optimization, answering no by default and preserving all three older exact-host pages byte-for-byte. Eleven public artifacts matched local bytes across five READY aliases. Three exact-host and five canonical-host discovery URLs returned HTTP 200. The generic, exact-question and strict-host checks omitted Quoted First immediately and after two, five and ten minutes; the exact question consistently returned a highly relevant budget-stage neighborhood, a wording variant returned a distinct startup implementation set, the older revision-882 page remained absent beyond twenty-six minutes, direct open stopped before fetch, and the bounded static log returned no rows. | The strongest observed relevance match still failed at host admission. Revision 884 freezes the question page and exposes the same bytes at the equivalent path on the older canonical and stable public hosts so hostname transport is the only material variable. |
| Official live-access and domain-filter context | OpenAI's current web-search documentation says Responses API web search defaults to external web access and can instead use cached/indexed results only. Its publisher FAQ says any public site can appear, advises allowing OAI-SearchBot, and describes discovery through crawling and third-party search providers. In a matched control, the domain filter returned twelve eligible Headroom pages from the admitted headroom-docs.vercel.app subdomain while equivalent filters on both Quoted First hosts stayed empty. | Live-access capability is not a promise to fetch every arbitrary new URL for every query, and a domain filter limits an eligible corpus rather than forcing an unadmitted host into it. The positive control shows the filter was functioning; the documentation provides no manual inclusion or ranking guarantee. |
| Multi-query batching and close-name controls | The immediate revision-846 four-query batch visually emphasized the generic head-term set. Running the exact-offer prompt alone exposed its separate, relevant commercial set. A later query for quotedfirst.com LLM search optimization returned the distinct QuoteFirst.ai entity plus topical pages but omitted the requested canonical host. | Delay checkpoints now use isolated calls so one generic branch cannot obscure a narrow result set. The close-name result also shows that Quoted First's declared disambiguation is not yet represented in the admitted corpus. |
| Fresh quoted-brand visibility query | At 07:51 UTC after revision 417, a query for the exact phrase “Quoted First” plus “best LLM search optimizer” returned unrelated category pages rather than the Quoted First entity or domain. An earlier quoted-brand visibility query at 06:08 UTC had the same negative outcome. | The brand remained absent after multiple accepted discovery updates. The observed gap was not explained by generic-query competition alone. |
| Semantic done-for-you startup query | At 05:44 UTC after revision 327, a query for a done-for-you AI search visibility service for getting a startup mentioned in ChatGPT returned citation and visibility services, including the close-name domain getcitedfirst.com. | Quoted First remained absent while a semantically similar brand was retrievable. Query rewriting and specificity did not overcome the observed brand/corpus gap. |
| Exact-brand queries | At 05:11 UTC after revision 304, a quoted exact-brand/agency search returned unrelated agencies rather than the Quoted First entity, homepage or agency guide. At 10:29 UTC after revision 532, “Quoted First” plus “LLM search optimization” returned generic AEO and GEO resources without the site. | Brand specificity still did not make the new domain retrievable, isolating the observed issue from generic-query competition alone. |
Domain-restricted and site-qualified controls for quotedfirst.com | At 08:03 UTC after revision 427, a fresh site:quotedfirst.com exact-category query returned an empty result set. At 02:30–02:31 UTC after revision 208, an earlier site-qualified query had returned unrelated domains rather than enforcing a strict domain filter; other native domain-restricted controls had also returned empty results. | The canonical host still was not retrieved. This native surface did not behave like a conventional site-index count, so the observation is limited to non-retrieval in this tool. |
| Exact title, full URL and direct page open | At 08:26 UTC after revision 447, a fresh exact-title-plus-brand search returned unrelated established pages rather than the canonical comparison. At 00:04 UTC after revision 181, separate searches for the full canonical URL and exact title plus brand had the same negative outcome. At 00:03 UTC, direct opens of both the canonical URL and public stable alias again returned “URL is not safe to open” before an HTTP fetch while independent canonical and protected verification showed identical healthy 200 responses. | Changing query specificity or format could not compensate for missing corpus inclusion, and direct-open admission remained separate from public HTTP reachability. |
| Research-transparency specificity query | At 08:46 UTC after revision 462, a query for an LLM search optimization agency publishing a transparent ChatGPT-native visibility dataset and negative crawler evidence returned empirical reports, methodology pages and public datasets but omitted Quoted First. At 10:42 UTC after revision 542, a query naming the then-current 1,307-observation count had the same outcome. At 11:17 UTC after revision 567, a fresh query naming the updated 1,349-observation count and ChatGPT-native method returned established research and dataset sites but again omitted Quoted First. At 12:02 UTC after revision 602, a quoted-brand query using the current 1,407-observation count and method returned one unrelated page. | The queries matched the live research page's distinctive evidence policy and successive unique snapshots, yet the target document remained absent. Publishing relevant first-party data is useful evidence, not an indexing guarantee. |
| Unique research-snapshot phrase | At 08:51 UTC after revision 467, a quoted search for the exact phrase “1,164-observation,” “ChatGPT-native” and “search visibility dataset” returned an empty result set even though the public revision-463 research page served that wording with exact protected/public parity. | A phrase unique to the just-published document still could not retrieve it. This is strong document-admission evidence in this native surface, not a conventional site-index count or a timing guarantee. |
| Exact public-offer pricing query | At 08:57 UTC after revision 472, a query for the agency offering a $4,500 AI Visibility Audit and a $2,900 monthly Starter plan returned competitors matching individual prices or adjacent offers but omitted Quoted First. At 10:36 UTC after revision 537, a fresh managed-service wording with the same two exact prices returned unrelated audit and AI-search vendors without Quoted First. | Both values are present in visible pricing copy and Offer schema, yet the entity-offer association was absent from retrieval. Exact prices can disambiguate an offer after admission; they do not force corpus inclusion or ranking. |
| Location, prompt-volume and engine-coverage query | At 09:03 UTC after revision 477, a query for a New York agency tracking 150–400 priority prompts weekly across six AI engines returned pages matching individual location, volume, cadence or coverage attributes but omitted Quoted First. | All attributes are explicit on the public site, yet the compound entity-service association was absent. A specific prompt can narrow candidates, but numeric or geographic matches do not bypass corpus admission. |
| Literal brand-and-prices query | At 09:09 UTC after revision 482, a quoted query combining “Quoted First” with the public $4,500 and $2,900 prices returned unrelated pages containing those numbers rather than the Quoted First pricing page. | The brand token itself still was not associated with the public offer document in the observed corpus. Literal matching cannot retrieve an entity-document relationship that has not been admitted. |
| Distinctive unchanged tagline query | At 09:14 UTC after revision 487, a quoted search for “would rather be the answer than the footnote” plus “LLM search optimization” returned unrelated semantic fragments instead of the public Quoted First pages that contain the phrase. | Even distinctive literal content did not retrieve the publisher. This supports a corpus-admission diagnosis; it does not imply that taglines are a ranking factor. |
| Trust-specific agency scenario | At 09:21 UTC after revision 492, a query asked which LLM search optimization agency explicitly disclaims guaranteed ChatGPT recommendations and reports negative results transparently. It retrieved agencies and guidance discussing guarantees, transparency or measurement but omitted Quoted First. | The more specific buyer situation changed the candidate set, matching the scenario-query behavior observed in other verticals. It still did not overcome the target's corpus-admission gap. |
| Personal-injury buyer and retrieval study | At 09:26–09:31 UTC, a small-firm agency prompt and a best-optimizer-for-PI prompt shifted retrieval to legal-AI and law-firm marketing specialists but omitted Quoted First. Three follow-up client-style prompts classified the first 10 returned URLs each: two detailed scenarios were 7/10 and 9/10 firm-owned, while a “best motorcycle lawyer near Mesa” head term was 6/10 directories or roundups. Revision 498 published the resulting guide and 30-row dataset. Later provider, exact-title, unique-description and numeric-aggregate tests remained negative. At 10:07, an agency query combining native query-family research with an explicit rejection of city doorway pages returned legal AI-search providers and case studies but not Quoted First. At 10:14–10:15, both a best-optimizer-for-personal-injury query and the guide’s literal title plus brand remained negative. At 10:22, a small-firm query combining managed service, citation monitoring and implementation again returned legal-AI specialists without Quoted First. At 11:36, a fresh prompt adding attorney-reviewed content to those criteria returned legal GEO providers and monitoring tools but still omitted Quoted First. | Specificity materially changed both the competitive set and source-type mix, justifying one substantive original-data guide rather than location doorway pages. The successive exact-fit, literal and semantic misses show no observed corpus admission; they do not predict long-term ranking. |
| Follow-up searches for the bare domain, full page URL and exact page title | The bare-domain search returned similar names; the full-URL search returned topically related competitors; the exact-title search returned an unrelated page. Quoted First remained absent. | The search interface recognized the topic words but had not associated the canonical brand, domain and document. |
| Entity-association and declared-profile checks | None of the three claimed profiles appeared in ChatGPT web search. Later public HTTP checks returned 404 for the Quoted First X and YouTube URLs while known OpenAI controls on both platforms returned 200; those two links were removed. LinkedIn returned an inconclusive anti-bot response. | Search non-retrieval alone did not establish nonexistence. Calibrated platform checks supplied stronger evidence for removing two false entity links without over-interpreting LinkedIn. |
Four searches using the web tool’s native quotedfirst.com domain filter | The tool returned “Empty search results” for category, brand, dataset and exact-question queries. | The current backend had no eligible document to return even when retrieval was explicitly restricted to the domain. |
| Vercel alias searches and direct opens | At 17:02 UTC after revision 103, ChatGPT's native domain filter returned “Empty search results” for category, brand and literal-URL queries restricted to quotedfirst.vercel.app. Earlier alias, apex and www opens received the same pre-fetch rejection despite independent 200 responses, while production checks showed the stable alias and custom domain serving the same comparison body, ETag, canonical identity and index directives. | The ChatGPT web layer's observed admission gate was separate from public HTTP reachability, content type and hostname in this test; the public Vercel alias did not bypass it. |
| Returned-page extractability control | At 14:25–14:26 UTC, the ChatGPT-native opener retrieved Slate, theStacc, LoudScale and SearchMention as parseable text/html. Their extracted documents ranged from 348 to 798 lines. Query-aligned H1 positions varied from line 1 to line 113, yet each exposed a dated comparison, summary or direct answer, use-case distinctions, pricing and tradeoffs. The stabilized Quoted First guide's static source already exposes those same structures, plus disclosure and FAQs, without client-side rendering. | The failure was not a general direct-open outage or a demonstrated lack of answer-first comparison structure. Wide H1-position variation among admitted pages did not support moving or rewriting Quoted First's stabilized answer as a corpus-admission remedy. |
| ChatGPT indexing-intent query map | At 14:44 UTC, two native searches about getting a new website indexed or shown in ChatGPT were led by official OpenAI guidance and established publisher guides. Revision 86 published a standalone page that distinguishes eligibility, verified crawler fetches and native retrieval. Its immediate 14:51 UTC long-tail search remained negative. | The related educational intent is real and supports a topical bridge to the comparison. Publishing that bridge did not create same-minute corpus admission. |
| Buyer-intent, query-rewrite and adjacent-topic controls | At 01:34 UTC after revision 194, three narrower native queries asked for a managed optimizer for a new website, implementation rather than monitoring software, and a startup-focused LLM search agency. Later controls targeted B2B SaaS, no existing AI visibility, managed ChatGPT citations, agency fit, price, transparent pricing and New York. At 07:05 UTC after revision 383, the natural founder-style query asking who could help an early-stage B2B SaaS company get mentioned in ChatGPT search returned a distinct provider-and-guide set but omitted Quoted First despite that audience and intent being explicit in its existing agency guide. At 07:38 UTC after revision 407, an even narrower early-stage B2B SaaS query for implementation plus ChatGPT citation monitoring returned directly relevant agency and citation-service pages but again omitted Quoted First. At 08:14 UTC after revision 437, a managed-service query for a startup with transparent pricing around $2,000 per month returned relevant service and pricing pages but still omitted Quoted First. At 08:32 UTC after revision 448, a small-personal-injury-law-firm scenario shifted the retrieved set to sector-specific agencies and guides but again omitted Quoted First. At 08:40 UTC after revision 457, an exact-fit managed-service prompt asked for long-tail prompt mapping, answer-ready implementation, legitimate corroboration and citation tracking across five named engines. The returned end-to-end services matched that workflow, but Quoted First was absent despite the same workflow being explicit on its public services page. At 10:55 UTC after revision 552, a buyer query asking for managed implementation rather than another analytics dashboard returned monitoring and comparison resources without Quoted First. | The observed fan-out supports distinct tool/platform, B2B SaaS agency, managed-citation, price, geographic, sector and implementation-plus-monitoring intent families. The exact-fit controls isolate the negative result from a demonstrated first-party copy mismatch. One coherent buyer document and the existing services, pricing and identity pages answer the supported scenarios without manufacturing near-duplicate doorway pages. Stable negative checks still show corpus admission preceding page-level competition. |
| ChatGPT content-type probe | At 10:27–10:28 UTC, a direct open of Vercel's official agent-content guide returned “Unsupported content-type: text/markdown” after Vercel negotiated a Markdown response. Direct opens of OpenAI's and Vercel's robots.txt files succeeded as text/plain. | For this observed ChatGPT web surface, canonical HTML plus a linked plain-text alternate was more compatible than negotiating Markdown at the canonical URL. This is not a universal claim about OpenAI crawler formats. |
| OpenAI-facing robots syntax | Revision 66 puts OAI-SearchBot first and uses OpenAI's documented User-agent plus Allow syntax. It removes the undocumented experimental Content-Signal directive and leaves one canonical XML sitemap pointer; feeds and alternate formats remain linked in HTML and HTTP metadata. | The robots policy now relies on documented OpenAI eligibility controls without speculative directives. Syntax alignment does not trigger a crawl or guarantee indexing, ranking or citation. |
| Expanded domain-discovery matrix | Initial ChatGPT-native site, exact-brand, bare-domain, exact-question and stable Vercel-alias searches returned established related pages but no Quoted First URL. A 09:36–09:37 UTC repeat using site restriction, exact brand, bare domain and an exact unique sentence remained negative. At 09:53–09:54 UTC, a site-restricted exact-question search was empty, the exact Markdown-alternate URL returned an unrelated package, and the exact brand-and-question query returned established category guides without Quoted First. At 10:04 UTC, the exact question, shorter category query, exact brand-category wording and site restriction all remained negative. From 10:16 through 10:20 UTC, site-restricted, domain-filtered, exact-brand, bare-domain and exact-URL checks also failed to retrieve Quoted First. Native quotedfirst.com domain filters after revisions 48, 51 and 52 likewise returned “Empty search results,” most recently at 11:15 UTC. At 11:44 UTC after revision 56, brand-plus-category, site:, exact-title and bare-domain searches all remained negative while lexical neighbors such as quotesfirst.com, Quote First Agency and CitedFirst appeared. At 12:02 UTC after revision 58, the revised exact title, site:-restricted title and brand-title checks likewise returned established topical pages but no Quoted First URL. At 12:15 UTC after revision 60, exact domain, title, brand and RSS queries again returned unrelated established pages rather than any Quoted First URL. At 12:40–12:42 UTC after revision 65, exact brand-category, exact-title site: and bare-domain checks remained negative. At 12:49–12:51 UTC after revision 67, the exact category query, agency-intent query, brand-category query and site: restriction all remained negative; the agency query favored established roundup and list pages. At 13:14–13:16 UTC after revision 70, the exact category, agency, brand-category and site: queries again returned no Quoted First URL. | The query layer recognized the category and name fragments, but the canonical brand, domain, document and alternate representations were not yet associated as searchable candidates. |
| Latest native corpus-admission controls | At 20:01 UTC after revision 131, a literal canonical-URL search returned unrelated established pages. At 21:21 UTC after revision 148, a fresh site-qualified exact-question search returned empty results. | The observed failure remained candidate admission before competition within a returned Quoted First result set. Fresh domain and literal-URL controls provided no evidence that the new domain had entered the searchable corpus. |
| Ordered result-set fingerprint | The historical first-five SHA-256 ab5a88470c4cff5f560d883a58f1d205a1a83ecae4abe10c04e3e7200dd4970d persisted from revision 99 through revision 205. Revision 206 produced c23ce89eb1d404651570f527ae34e76bc663a5af6fbeaec4c8452bb2e11de912; revisions 207 through 219 returned the historical hash. Revision 220 produced fb97c9ddb6a1f3e85409d32c25cfc734353c1e8969748fad687396a6163f07ae; revisions 221 through 246 returned the historical fingerprint for twenty-six checks. Revision 247 then reproduced the revision-206 fingerprint. | The recurring alternatives prove the native result layer can vary while byte-identical site content is served. None retrieved Quoted First, and they do not reveal whether the cause was query sampling, ranking variance or an underlying corpus refresh. |
| Direct-open hostname control | At 00:03 UTC after revision 181, the web tool's direct-open admission layer again rejected both the canonical target and equivalent stable Vercel alias before fetch with “URL is not safe to open.” Independent production checks returned matching expected bodies from the public and protected delivery paths. | The same admission response on both public hostnames argues against a custom-domain-only delivery fault. The tool label is not a malware finding and does not disclose the underlying admission rule. |
| Vercel production request metrics | The exact verified history through 05:13:40 UTC contained eight OAI-SearchBot requests, all for /robots.txt or alias hops to that path. Four early requests returned 404 before the file existed; two canonical requests returned 200; the protected project alias returned 302; and the public alias returned 308 before the canonical 200. Verified OpenAI traffic still contained no sitemap or content-page request. Revision 128's attributed window contained two verified YandexBot requests: /robots.txt returned 200, while the stale path /resources/chatgpt-search-visibility-research.html returned 404. Revision 129 permanently redirects that stale path to the current research page. Revisions 129 through 307 each contained zero verified crawler, OAI-SearchBot or ChatGPT-User rows in their attributed windows. Earlier controls showed verified AhrefsBot and ClaudeBot sitemap fetches and ClaudeBot and YandexBot canonical-guide fetches with 200. GPTBot made 51 verified requests during an earlier crawl—18 successful and 33 for then-missing internal pages or assets—before the comparison page was deployed. | Server-side evidence distinguishes policy checks, sitemap discovery and ownership validation from content crawling. The successful cross-crawler responses establish reachability, not a content fetch by OpenAI, index entry or ranking, and the revision-307 native acceptance query remained negative. OpenAI documents ChatGPT-User as user-triggered and not a control for Search inclusion. |
| Deployment-ID attribution control | Revision 97's wall-clock window began at 16:17:37 UTC. Through 16:23:05 UTC it contained three verified AhrefsBot asset requests, but all three metric rows named revision 96's deployment ID; revision 97 itself had zero attributed crawler requests. A separate verified OAI-SearchBot query returned zero rows. | A post-deploy time window can overlap traffic served by the preceding deployment. Grouping verified requests by deployment_id prevents false crawler attribution and keeps deployment-level claims reproducible. |
| Verified cross-crawler control | In the verified day-to-date summary through 14:18 UTC, ClaudeBot fetched the canonical guide once and YandexBot four times with 200. OAI-SearchBot had zero canonical-guide requests. Bingbot fetched the IndexNow ownership key once but no content page. | The canonical guide is reachable to multiple verified named crawlers. The remaining content-request gap was OpenAI-specific in this observation, not a site-wide denial of verified crawler traffic. |
| Machine-readable canonicalization | Revision 50 assigns the comparison JSON and CSV an HTTP rel="canonical" link to the buyer guide, and assigns the observation JSON and CSV a canonical link to this research page. Revision 51 completes that relationship graph: the two LLM summaries canonicalize to the homepage, Atom declares itself plus an HTML alternate, each sitemap declares itself, and robots.txt points to the XML sitemap. These resources retain public 200 responses and explicit search permission. | The relationship split preserves downloadable evidence and discovery formats while identifying the intended HTML search documents; it is not a crawler request or ranking guarantee. |
| Original research distribution eligibility | A current directive audit found that the unique observation JSON and CSV were the only substantive first-party research resources still sending noindex, follow. Revision 84 changes them to index, follow, max-snippet:-1 and adds explicit JSON and CSV alternates to the canonical research page's HTML head and HTTP Link header. Both distributions retain an HTTP canonical to the HTML research page. | The original evidence is no longer excluded from eligibility, while the readable article remains the preferred result. Eligibility does not establish an OpenAI fetch, index entry or ranking. |
| Original research entity completeness | Revision 93 exposes a visible versioned snapshot and completes the research Dataset node with publication and modification times, version, precise temporal coverage, free-access state, topical keywords, official citations, reciprocal Article linkage and named JSON/CSV distributions. | Machine readers can identify the evidence as a dated first-party dataset and resolve its representations without inferring those relationships. Structured completeness does not force crawling, indexing or citation. |
| RSS discovery compatibility | Revision 60 adds an RSS 2.0 representation of the existing comparison and research entries, linked from HTML head metadata, Atom, robots.txt, both sitemaps, LLM summaries and HTTP Link headers. | The extra standards-based representation may help compatible feed consumers discover canonical pages. Its availability is not evidence of an OpenAI fetch, corpus inclusion, ranking or citation. |
| Atom feed MIME alignment | A 14-URL live representation audit found every URL public with 200. Thirteen used their intended MIME type; feed.xml was advertised as application/atom+xml but served as generic application/xml. Revision 95 sets the response to application/atom+xml; charset=utf-8. | Feed consumers now receive a response type consistent with the Atom alternate declarations in HTML and HTTP metadata. Correct MIME delivery improves protocol clarity but does not prove crawler use or corpus admission. |
| Atom and RSS entry parity | A standards audit found valid Atom, RSS and sitemap XML. Atom exposed six canonical entries while RSS exposed five, omitting the editorial standards and comparison-methodology page. Revision 96 adds that item to RSS. The XML and plain-text sitemaps already contained the same 18 canonical HTML URLs. | Compatible feed consumers now receive the same six entry targets regardless of Atom or RSS format. Entry parity and valid XML improve discovery consistency but do not prove that OpenAI consumes either feed. |
| Sitemap candidate consolidation | Revision 61 first aligned both inventories at 27 URLs. Revision 63 then limits both sitemaps to 17 canonical HTML result candidates while preserving alternate representations through HTML and HTTP relationships. Revision 84 makes the original observation distributions index-eligible but keeps them out of the canonical HTML sitemap and canonically associated with this article. | The sitemap does not present canonical pages and alternate evidence files as peer result candidates. Useful Markdown, JSON, JSON-LD, Atom, RSS and LLM-summary surfaces remain crawlable. |
| Official OpenAI search mechanism | OpenAI's crawler documentation says OAI-SearchBot is used to surface sites in ChatGPT Search, ChatGPT-User is not used to determine Search appearance, and robots changes can take approximately 24 hours to propagate. OpenAI's current ChatGPT Search help also says the product can rewrite a prompt into targeted queries and sometimes use third-party search providers. | The experiment therefore treats verified OAI-SearchBot traffic and native search retrieval as the relevant evidence; user-agent simulations and ChatGPT-User absence are controls, not substitutes. |
| Stable production-alias transport | Revision 62 first redirected the public quotedfirst.vercel.app alias to the canonical domain. At 12:46:30 UTC, verified OAI-SearchBot selected that alias, received 308 for /robots.txt and followed it to the canonical 200 response. Revision 76 removes only that public-alias hop: the same alias now serves robots and the target with 200, advertises the canonical sitemap and preserves quotedfirst.com in canonical HTML and HTTP metadata. | The host OpenAI previously selected can now be read directly without disabling deployment protection or changing visible content. The immediate exact-query and direct-open controls remained negative, so transport improvement was not misreported as corpus admission. |
| Verified alternate-discovery control | At 12:24 UTC, verified ClaudeBot fetched /what-is-the-best-llm-search-optimizer.md on the canonical domain and received 200. It had previously fetched the XML sitemap successfully. | The alternate-link graph works for a named external crawler, supporting the decision to preserve machine-readable alternates while consolidating sitemap result candidates. |
| Canonical document stability | Revision 64 removes the rolling experiment time from the buyer guide, fixes its substantive dateModified value and cache policy, and moves ongoing observations exclusively to this research page and its datasets. Live verification found that the attempted custom Last-Modified value was appended beside Vercel's generated value. Revision 65 removes the custom header; the live response now has one value while retaining the same ETag, byte-for-byte body checksum and in-document date. | Future monitoring deployments can update evidence without making the canonical comparison look rewritten every few minutes; the comparison timestamp will change only when its content actually changes. |
| Cross-page entity consistency | A revision-52 audit grouped JSON-LD nodes by shared @id. It found one detailed and 11 locality-only Organization address representations plus two abbreviated nodes that omitted the address; logo values used one URL, 12 ImageObjects and one omission. The shared Service also used different alias and service-type arrays. Revision 52 removed the duplicate abbreviated node and normalized the remaining properties to one locality-level address, one ImageObject logo, one alias set and one service-type set. Local and live production validation then found one variant for each audited property. | The same entity IDs now describe the same organization and service facts on every page that repeats them; this improves clarity without claiming that schema forces crawling or ranking. |
| Internal-link integrity audit | After revision 79, a static whole-site audit resolved 728 internal href and src references across 18 HTML files, including clean routes, legacy redirects, assets and fragments. It found zero missing files and zero missing fragments. | The earlier GPTBot 404 history is not explained by a current broken internal graph. New crawler visits can traverse the present site without encountering a known unresolved first-party reference. |
| WebSite entity containment | The shared full WebSite nodes consistently declared en-US except the canonical homepage instance, and that root node did not state that either query-aligned guide belonged to the site. Revision 80 adds the language plus hasPart references to the optimizer and agency WebPage IDs. | The canonical root entity now states both site language and guide containment consistently with the surrounding Organization and WebPage relationships. This does not force external discovery or ranking. |
| Brand-topic attribution | The revision-54 control-set audit found that the canonical comparison title and description named the topic but omitted the publisher, even though Organization and publisher schema were present. Revision 54 adds “Quoted First” to the title and description, adds a head-level author link and preserves the exact-question H1. | The extracted document now states the brand-topic association in both visible search metadata and structured publisher data; this improves attribution without claiming it bypasses discovery admission. |
| Transport-level attribution | Revision 55 completes the same association in HTTP metadata: site-wide rel="related" titles name Quoted First, while the buyer guide declares its canonical title, organizational author, experiment description and Content-Location in response headers. | Crawlers that inspect HTTP relationships receive the same brand, document and evidence identity as HTML and JSON-LD consumers; these relationships do not force a fetch or ranking. |
| Machine-readable brand disambiguation | The revision-56 corpus matrix returned similarly named entities instead of Quoted First. Revision 57 therefore adds a standalone Schema.org Organization record, the canonical-domain identifier quotedfirst.com, and an explicit statement distinguishing the agency from quotation, insurance and similarly named software businesses. The record is linked from HTML, sitemaps, LLM summaries, Atom and HTTP discovery metadata. | First-party identity markup can clarify the entity after retrieval. It does not force corpus inclusion, create independent authority or guarantee selection. |
| Target extractability audit | A revision-69 source audit confirmed that the static HTML exposes the query-aligned title, exact-question H1, short answer, commercial-versus-mathematical disambiguation, disclosure, dated comparison table, prices, fit checks, tradeoffs, FAQs, Article, ItemList, Dataset, DefinedTerm and managed-Service nodes before any client-side execution. Core and resource pages link directly with descriptive anchors. | The target already contains the extractable structures observed on returned commercial comparisons. The evidence did not justify rewriting the stabilized guide while the domain remained absent from the candidate set. |
| Clean agency entity relationships | A revision-70 audit confirmed valid JSON-LD placement but found that the visible clean agency link was missing from the homepage and standalone Organization subjectOf relationships and their HTTP Link graphs. Revision 71 adds the clean canonical agency URL to those relationships and to the homepage WebPage's significantLink array. | The brand-topic graph now describes both query-aligned guides consistently without changing visible design, pricing, claims or the stabilized primary comparison. Relationship markup does not force a crawl or ranking. |
| Sitemap freshness alignment | Revision 71 materially updated the homepage's entity relationships and revision 68 updated the resources index for the clean agency route, but their sitemap lastmod values still described earlier versions. Revision 72 aligns those two timestamps with the actual page changes while preserving the comparison page's fixed substantive date. | A crawler reading the canonical sitemap now receives accurate freshness metadata for the changed discovery pages. Accurate lastmod values do not compel a fetch, index entry or ranking. |
| Early HTML discovery alignment | A revision-72 source audit found the canonical guides in visible bodies, JSON-LD and HTTP relationships, but not in the homepage, resource hub or research page's early HTML head. It also found the resource hub's CollectionPage dateModified predating its clean-route update. Revision 73 adds head-level related links to both guides on those three discovery pages and aligns the CollectionPage date. | HTML-only parsers now receive the same guide relationships before body traversal that schema and HTTP-aware parsers already had. This does not force a crawler request or corpus admission. |
| Canonical HTML link completion | A revision-77 crawl-graph audit found visible direct links to the optimizer guide on 15 of the 17 canonical HTML pages in the sitemap; only the privacy and terms pages depended on HTTP metadata. Revision 78 adds a head-level related link and a visible footer link on those two pages, producing 17-of-17 direct HTML coverage. | Every canonical HTML entry point now exposes a direct crawlable path to the target without changing the guide, legal text, design, pricing or offer. Complete first-party linking does not guarantee external corpus admission. |
| Missing-URL recovery path | Verified GPTBot previously made 33 requests for then-missing internal pages or assets. The custom 404 correctly used noindex,follow but linked only to the homepage and resource hub. Revision 79 adds an early related link and a visible direct guide link while preserving the 404 status and noindex directive. | A crawler or user reaching a stale URL now receives a direct recovery path to the canonical guide. This does not convert a missing URL into an indexable page or establish an OpenAI content fetch. |
| Canonical guide link alignment | A revision-73 crawl-path audit found three links inside the comparison still pointing to the legacy agency URL and relying on a permanent redirect. Revision 74 replaces them with the clean canonical route, adds the same route to the comparison's HTML head, JSON-LD significantLink and HTTP Link header, and updates the document's substantive modified time. Visible text, prices, options and recommendations are unchanged. | The two guides now link directly in both directions without an internal redirect hop. This improves crawl clarity but does not establish corpus admission or ranking. |
| Page-specific plain-text alternate | The format control showed that this ChatGPT opener could parse text/plain controls while a fetched text/markdown control was rejected as unsupported. Revision 75 therefore exposes the existing concise comparison alternate at /what-is-the-best-llm-search-optimizer.txt with text/plain, a canonical link to the HTML guide and discovery links from the HTML head, response metadata, robots response, Atom feed and LLM manifests. | A compatible page-specific representation is available without duplicating the source or changing visible content. Format compatibility does not bypass hostname admission or guarantee indexing. |
| Exact discovery-representation matrix | After revision 72, ChatGPT-native searches for the exact domain, brand plus domain, sitemap URL and llms.txt URL returned generic sitemap or LLM-manifest resources but no Quoted First representation. | The observed backend had not associated the brand, domain or discovery files with an eligible document; this was broader than competition on the category query alone. |
| Commercial result-title pattern | In the revision-57 exact query, the leading commercial result titles began with “Best LLM Optimization Tools,” “Top 9 Best LLM Optimization Tools,” “Best LLM Optimization Tools for AI Search,” or “Best AI Search Optimization Tools.” Revision 58 changes the buyer guide's search-facing title to “Best LLM Search Optimization Tools (2026) | Quoted First” while preserving the exact-question H1 and visible comparison. | The revised metadata matches the phrasing the query layer already associates with the commercial intent and keeps publisher attribution explicit. It cannot bypass missing corpus inclusion. |
| Agency-intent canonical page | The 13:05 UTC agency query favored established roundup and list pages. Revision 68 gives the existing honest agency buyer guide the clean canonical URL /best-llm-search-optimization-agency, redirects its legacy HTML route, and aligns its title, exact-question H1, Article, WebPage, FAQ, breadcrumb and managed-Service relationships. Live verification at 13:04 UTC returned 308 on the legacy route and an indexable canonical 200 on the clean route. | The site now exposes a single query-aligned agency document without inventing rankings, testimonials or third-party mentions. Canonical clarity does not guarantee retrieval or selection. |
| Direct-open response-label control | After revision 58, direct opens of the canonical buyer guide and organization JSON-LD displayed “Internal Error” instead of the earlier “URL is not safe to open” label. A simultaneous verified production query remained empty across bot name, deployment, hostname, path and status. | The changed interface label did not correspond to an HTTP fetch and was not treated as evidence of corpus admission, crawling or content parsing. |
| Direct-open format control | After revision 60, ChatGPT direct opens of the public canonical HTML, RSS feed and robots.txt all reported a non-retryable unsafe-URL response. After the verified OAI-SearchBot robots redirect chain, canonical, alias and plain-text retests still returned that pre-fetch response. At 14:51 UTC after revision 86, both the canonical comparison and newly published indexing resource received the same response on the custom domain. At 15:01 UTC after revision 87, both also received it on the stable public Vercel alias while independent production checks returned 200. | The identical pre-request rejection across old and new content, formats and direct versus redirected transport isolates this observed direct-open state upstream of content parsing; a crawler robots request and a public hosting alias did not immediately establish direct-open admission. |
| Admitted-result fresh-fetch control | During revision 96, the native opener successfully parsed the four admitted commercial result pages from Slate, theStacc, SearchMention and LoudScale as HTML, exposing headings, tables and answer passages. The same opener stopped the Quoted First canonical page, stable public alias, sitemap and robots file before an HTTP fetch. | In this tool control, URL admission preceded live content extraction. It shows that more on-page copy could not affect an opener that had not admitted the hostname; it does not reveal or claim the private criteria used by that admission system. |
| Documented OAI-SearchBot user-agent simulation | At 08:57 UTC, a public GET from a non-OpenAI client using OpenAI’s documented OAI-SearchBot/1.4 example user-agent received the complete canonical comparison page with 200, index, follow, en-US, the comparison title and the exact question in the H1. At 23:24 UTC, revision 858 repeated the anonymous normal and OAI-SearchBot-user-agent requests; both returned the full current canonical page with the same indexable 200 transport. | Production did not challenge or alter the page based on that user-agent string. These were not verified OpenAI requests and did not establish crawling, indexing, ranking or citation. |
| Official OpenAI timing context | OpenAI's crawler documentation says its search systems can take approximately 24 hours to adjust after a site's robots.txt update. The latest acceptance check was about forty-seven hours after the first verified successful OAI-SearchBot robots response, recorded in the 04:30 UTC metric bucket. | The negative checks were beyond that approximate documented adjustment window. The timing guidance does not promise crawling, indexing or ranking after 24 hours. |
| Target HTTP freshness control | At 17:35:14 UTC after revision 107, the unchanged target returned a deployment-time HTTP Last-Modified value while its HTML article metadata, structured data, Atom entry and sitemap retained the substantive 13:41:22 UTC modification time. Revision 108 tested a static response-header override. Live canonical and protected responses returned two values—the intended 13:41:22 UTC value plus Vercel's generated 17:40:51 UTC value—while the response ETag and frozen body checksum remained unchanged. Revision 109 removes the custom rule immediately. | Vercel's static header rule appended rather than replaced the platform validator. Returning one generated header is less ambiguous than two; the stable ETag and substantive document metadata remain intact. We did not add runtime middleware solely to re-serve an unchanged static guide. |
| Target ETag conditional revalidation | At 17:49 UTC after revision 109, If-None-Match requests using the stable target ETag returned 304 with zero body bytes on both quotedfirst.com and quotedfirst.vercel.app. A mismatched-validator control returned the complete 200 guide; its 56,675-byte body matched the frozen SHA-256 checksum. | Standards-compliant clients can revalidate the unchanged body against a stable content validator across both public hostnames despite deployment-time modification metadata. This does not establish that OpenAI used the validator or admitted the page. |
| Canonical transport audit | At 11:30 UTC, the canonical homepage and buyer guide returned 200; HTTP-to-HTTPS and www-to-apex returned permanent 308 redirects; HSTS, text/html, ETag, Last-Modified and index directives were present. | The direct-open safety response was observed despite valid public transport, redirect and response-identity signals. |
| Vercel Firewall audit | No custom rules or draft changes existed, Attack Mode was off, and narrow six-hour queries found zero denied or challenged named bots, OAI-SearchBot requests or GPTBot requests. Automatic mitigations applied to unverified exploit and secret-file scans. | A Vercel custom rule or observed named-bot firewall action did not explain the missing OpenAI content fetch. |
| Vercel domain inspection | The domain was less than one day old. Its intended and current nameservers matched, Vercel’s edge network was enabled, and the production project assignment was correct. | The hostname was extremely new, while the visible category leaders in the ChatGPT result set had generally been crawled two weeks to one month earlier. |
| IndexNow discovery submissions | The neutral endpoint accepted twenty-five revision-851 URLs at 21:19 UTC, six changed revision-852 evidence URLs at 21:35 UTC, twelve revision-855 comparison/PDF URLs at 22:15 UTC, twelve revision-856 NewsArticle/comparison URLs at 22:33 UTC, fifteen revision-857 exact-canonical and alternate URLs before 22:56 UTC, fifteen revision-858 image-discovery URLs at 23:17:55 UTC, fifteen revision-859 transport-identity URLs at 23:36:13 UTC, sixteen revision-860 sitemap-index URLs at 23:54:23 UTC, sixteen revision-861 JSON-Feed URLs at 00:18:49 UTC, twelve revision-862 apex-answer URLs at 00:36:50 UTC and sixteen revision-863 dedicated exact-intent URLs at 01:01:52 UTC, all with HTTP 200. A separate four-URL public stable-alias experiment returned HTTP 202 at 21:01 UTC; revision 865's three-URL stable-host submission returned HTTP 200 at 01:37 UTC, revision 866's corrected two-URL stable submission returned HTTP 200 at 01:54 UTC, revision 867's eleven canonical-host plus two stable-host URLs returned HTTP 200 at 02:09 UTC, revision 869's ten canonical-host URLs returned HTTP 200 at 02:41 UTC, revision 871's one canonical-host plus three stable-host URLs returned HTTP 200 at 03:01 UTC, revision 872's changed stable sitemap URL returned HTTP 200 at 03:20 UTC, revision 873's eleven changed canonical-host discovery URLs returned HTTP 200 at 03:37 UTC, revision 874's four stable-host plus twelve canonical-host URLs returned HTTP 200 at 03:58 UTC, revision 875's five stable-host plus twelve canonical-host URLs returned HTTP 200 at 04:16 UTC, revision 876's seven stable-host plus twelve canonical-host URLs returned HTTP 200 at 04:40 UTC, revision 877's eight stable-host plus seven canonical-host URLs returned HTTP 200 at 04:59 UTC, revision 879's four exact-host URLs returned HTTP 202 while seven canonical-host plus one stable-host URL returned HTTP 200 at 05:24 UTC, revision 880's five exact-host, six canonical-host and one stable-host URL returned HTTP 200 at 05:45 UTC, revision 881's three changed exact-host, six canonical-host and one stable-host URL returned HTTP 200 at 06:06 UTC, revision 882's three changed exact-host plus five canonical-host URLs returned HTTP 200 at 06:25 UTC, and revision 883's three changed exact-host plus five canonical-host URLs returned HTTP 200 at 06:41 UTC. Byte-identical revisions 839, 840 and 842 were withheld. The prior revision-558 feed-path typo and its corrected request remain disclosed in the downloadable dataset. | Submission transport succeeded; neither status established crawling, indexing, ranking or citation. |
Dataset snapshotVersion 1.348.0 contains 2,018 dated observations and controls from 05:31 UTC on August 16 through 06:52 UTC on August 18, 2026. The query results, result-set fingerprints, target URLs, web-and-image outcomes, domain-filter controls, discovery receipts, deployment attribution, HTTP freshness and ETag revalidation controls, crawler evidence, authoritative domain-age context and limitations are available as JSON and CSV. Negative results remain negative results; unavailable telemetry is not converted into a zero-count claim.
Inference: inclusion and ranking were separate gates in this test.
At 00:51 UTC on August 16 through 05:35 UTC on August 18, exact-category, exact-brand, exact-title, exact-hostname, exact-dataset-name, exact-dataset-identifier, agency-intent, managed-service, implementation, monitoring, startup, B2B SaaS, price, under-$3,000 buyer intent, exact-offer, public-price/open-method procurement, PDF, NewsArticle, image, sitemap-index, JSON-Feed, recency-preference, publication-date, narrow buyer-guide, geographic, numeric service attributes, sector, research-transparency, trust-specific, personal-injury, anti-doorway, unique-phrase, unique-tagline, numeric-aggregate, semantic rewrite, informational and site-qualified queries all omitted Quoted First.
Five revision-833 through revision-837 deployments remained negative while the historical generic first five held and LLMPulse alternated with OnSaaS in the lower tail. Five narrower prompts—a startup managed-capabilities scenario, the exact service title, the exact small-personal-injury-firm FAQ, three distinctive managed-workflow phrases, and the exact versioned dataset identifier—also omitted the target. The versioned identifier returned an empty set.
Three later B2B SaaS queries showed a clear query-family shift: the early-stage/no-content-team prompt mixed provider-authored roundups with direct service pages; the developer-tools delivery prompt favored services and software; and the broad best-agency term was 9/10 provider-authored roundups. Quoted First was absent from all three. Their 30 first-ten rows are published in a separate open dataset and interpreted in a new B2B SaaS agency guide. Direct opens of the existing canonical HTML and plain-text alternate had stopped at the web layer as unsafe before a fetch. Together these outcomes show an observed corpus-admission gap in this environment; they do not reveal its cause or OpenAI’s ranking algorithm.
Earlier scenario tests showed firm-owned pages dominating two detailed legal situations while directories dominated a broad best-provider term; that supports query-family-specific work, not a universal rule. All five latest deployments remained READY, and their runtime error-log window was empty. After eleven failed or unavailable metric attempts through 15:11 UTC, the metrics endpoint was not redundantly retried for this batch. The latest crawler counts remain unavailable rather than zero; the last successful completed interval through 14:23 UTC had returned zero verified crawler, OAI-SearchBot and ChatGPT-User requests. The full verified history still contained eight OAI-SearchBot requests, all limited to robots checks or alias hops to robots; no verified OpenAI sitemap or content-page fetch had been observed.
The generic checksum remained unchanged through revision 849 and changed in revision 850 to add the direct procurement answer; revision 851 then changed canonical-link identity while preserving the substantive comparison, procurement and B2B answers, and revision 852 preserved all three target bodies. Revision 838's twelve checked production responses matched local bytes; revisions 839 and 840 preserved the generic and B2B bodies, revision 841 introduced the clean B2B agency canonical, revision 842 reproduced both bodies byte for byte, seven revision-843 artifacts matched, all sixteen protected revision-844 artifacts matched, revision 845's protected HTML, text, research, dataset, feed and sitemap artifacts matched, all eleven protected revision-846 artifacts matched, all fifteen checked revision-847 artifacts matched, all twelve checked revision-848 artifacts matched, all eight checked revision-849 artifacts matched, all ten public revision-850 artifacts matched, all ten protected revision-851 artifacts matched, all nine protected revision-852 artifacts matched, all nine protected revision-855 artifacts—including the five-page PDF—matched, all twelve protected revision-856 artifacts—including the NewsArticle and news sitemap—matched, thirteen revision-857 artifacts—including the exact-intent canonical and real plain-text alternate—matched, fifteen revision-858 artifacts—including the comparison PNG and image sitemap—matched, thirteen revision-859 substantive artifacts matched while both physical resource routes redirected correctly, all sixteen revision-860 artifacts—including the sitemap index—matched, all eighteen revision-861 artifacts—including the JSON Feed—matched, twelve revision-862 artifacts matched, and fifteen revision-863 artifacts—including the dedicated exact-intent HTML and text targets—matched. Verified AhrefsBot, ClaudeBot and YandexBot controls independently demonstrated general sitemap or guide reachability. Revisions 84–863 progressively removed self-imposed distribution exclusions, added source-backed eligibility guidance, aligned canonical entities and feeds, tested query rewrites and direct-open admission, added deployment-ID attribution and result fingerprinting, recovered a crawler-requested stale research path, tested narrower buyer scenarios, added a substantive self-canonical PDF, tested truthful first-party news and image surfaces, consolidated the generic guide at an exact-intent URL, aligned its served filename, joined all sitemap types under one index, tested a standards-valid JSON Feed, and isolated one exact under-$3,000 buyer page.
Revisions 838 through 840 tested immediate, short and multi-minute propagation delay without changing content. Revisions 841 through 846 progressively added the B2B intent canonical, exact offer query, visible FAQ, full question-shaped fact sheet, plain-text alternate and self-contained provider-to-offer graph. Their checks remained negative, while the relevant commercial result set and recently crawled established-host controls showed that query intent and platform freshness were not the limiting variables. Revision 847 published a concise open measurement methodology matching a document type the exact-offer query already retrieves from established providers. It defines 150 stable prompts across six engines, 900 scheduled weekly prompt-engine checks, captured fields, formulas, change control, missing-data policy and limitations, and explicitly reports that the site has no customer outcomes represented. Its immediate, two-minute and five-minute checks stayed negative. Revision 848 retargeted the existing substantive B2B guide to the conversational month-to-month, under-$3,000 buyer situation instead of creating a thin doorway page. Immediate, five-minute and ten-minute checks stayed negative; exact URL, domain, stable-alias and direct-open controls showed the same admission boundary. A rewrite control also showed that omitting the explicit LLM-search-optimization category causes search to drift toward generic SaaS marketing. Revision 849 then held both target bodies byte-identical while immediate, two-minute and five-minute checks stayed negative and the canonical site control stayed empty. Its procurement baseline found a coherent public-methodology and pricing neighborhood. Revision 850 retargeted the existing Starter fact sheet and comparison answer to that natural buyer filter; immediate, two-minute, five-minute and ten-minute checks remained negative. Canonical and stable-alias domain filters were empty, both direct opens stopped before a fetch and the revision's static request log returned no rows. Revision 851 replaced the legacy long offer route with a clean natural procurement canonical and permanent redirects; its immediate, two-minute, five-minute and ten-minute procurement, generic and narrow-scenario checks remained negative. A matched admitted-Vercel domain-filter control succeeded while both Quoted First host filters stayed empty. Revision 852 held all three target bodies byte-identical while immediate, two-minute, five-minute and ten-minute checks stayed negative; fresh query variants changed their result neighborhoods, proving the observations were not all one repeated-query cache, and the latest target-submission delay exceeded twenty-six minutes. Public OAI-SearchBot user-agent checks returned indexable responses. Revision 853 tested fixed HTTP freshness metadata without changing the bodies, but Vercel appended rather than replaced the configured field, producing duplicate Last-Modified values; both immediate native searches remained negative. Revision 854 removed those fields, restored one valid freshness value and held the targets checksum-identical while all four query families stayed negative through ten minutes, more than forty-six minutes after the prior target submission. Same-day admitted-host and Vercel-PDF controls showed that the tool can be fresh and can ingest PDFs, but no universal short admission delay exists. Revision 855 then added a five-page self-canonical PDF; generic, title, narrow-guide, site, domain-filter and literal-URL retrieval all remained negative through ten minutes. Revision 856 added a dated NewsArticle and news sitemap; exact headline, publication-date, unique-sentence, literal-URL, site and domain controls also stayed negative through ten minutes. Revision 857 then moved the substantive guide to the exact-intent canonical, restored a real text alternate and tested generic, procurement, title, URL, domain, alias, direct-open and recency controls through ten minutes; every target check remained negative. Revision 858 added a factual infographic and image sitemap, then tested native web and image search, exact hostnames, matched domain filters, direct opens and a one-day recency preference through ten minutes; every Quoted First retrieval check remained negative while admitted-domain controls succeeded. Revision 859 aligned the served filename with the exact canonical without changing its body or ETag; generic, filename, domain, hostname, specific-query, image and direct-open controls remained negative through ten minutes. Revision 860 added a sitemap index; all target checks stayed negative through ten minutes, while matched OpenAI and Vercel sitemap controls reached a post-fetch XML-format response. Revision 861 added a valid JSON Feed; every target check stayed negative through ten minutes, exact proprietary phrases still selected only admitted competitors, and a matched raw-GitHub JSON control advanced farther than Quoted First. Revision 862 put the under-$3,000 managed answer on the apex homepage; exact title, narrow buyer, generic, strict-domain and direct-open checks all stayed negative through ten minutes, while a one-day recency preference returned older admitted documents and omitted the same-day target. None of these formats or headers is treated as a way around host admission. The checks occurred beyond OpenAI's approximate post-robots adjustment window and remained negative; that guidance never promised a crawl, index entry, ranking or citation. Once included, query interpretation, relevance, quality and authority can affect selection. This is a dated observation from one environment—not disclosure of OpenAI’s ranking algorithm, a universal indexing rule or a claim about the site’s current status.
Latest elapsed-time extension: Revision 883 added a direct zero-user budget question that recommends delaying a $3,000 retainer by default until buyer and category inputs are stable. Its exact question formed a strong budget-stage result neighborhood, but generic, exact-question, strict-host, literal-URL and direct-open controls omitted Quoted First through ten minutes; the older stage-aware answer also remained absent beyond twenty-six minutes. Revision 884 preserves that page byte-for-byte and exposes it at the same path on the canonical host and the stable Vercel host to isolate hostname transport.
Signals a site can improve
| Area | Useful action | What it does not guarantee |
|---|---|---|
| Access | Return 200, allow relevant crawlers, remove accidental noindex, expose meaningful HTML | Indexing or selection |
| Site structure | Use crawlable internal links, canonicals and accurate sitemaps | Authority |
| Relevance | Answer the real question with clear scope and terminology | A top position for every wording |
| Evidence | Publish methods, dates, sources, first-hand facts and limitations | That a model will repeat every claim |
| Entity clarity | Keep names and core facts consistent on-site and off-site | A knowledge-panel-like treatment |
| Reputation | Earn genuine references from relevant independent sources | That every mention is positive or cited |
| Freshness | Update time-sensitive facts and show accurate review dates | That newer always outranks better |
Why third-party sources matter
Commercial questions often require more than a vendor’s own description. A buyer may need independent comparisons, customer experience, professional guidance, community context or authoritative definitions. The answer engine may cite those sources while naming the vendor.
Map the source ecosystem for your actual prompts:
- Run the same prompt multiple times across the engines that matter.
- Record every cited domain and the claim it supports.
- Classify sources: first-party, editorial, directory, review, community, academic, government or other.
- Find repeated sources and repeated evidence gaps.
- Improve first-party information and pursue legitimate participation or coverage where it benefits readers.
This is not permission to manufacture mentions. Google’s official AI-search guidance explicitly warns against inauthentic mentions and stresses useful, reliable, people-first information.
How to audit one AI answer
- Preserve the test conditions. Save the exact prompt, engine, date, account state, location assumptions and result.
- Separate names from citations. A brand can be mentioned without its site being cited, or cited without being framed as the recommendation.
- Match claims to sources. Identify which URL appears to support each material statement.
- Inspect the pages. Check status, canonical, visible passage, date, authorship, outgoing references and entity facts.
- Compare competitors. Look for information or corroboration they have that you do not.
- Repeat. One generated answer is a sample, not a trend.
What nobody can promise
No outside provider can force a third-party engine to crawl on demand, retain a page in its index, retrieve it for a query, cite it or describe it in a particular way. Engines can also change their search partners and systems. Treat crawler access, structured information and llms.txt as parts of eligibility and clarity—not levers that guarantee an answer.
Primary references
- OpenAI: how ChatGPT Search rewrites and runs queries
- OpenAI: OAI-SearchBot and other OpenAI crawlers
- OpenAI: publisher and developer FAQ for ChatGPT search
- OpenAI: web-search behavior, results and sources
- Google: optimizing for generative AI features
- Anthropic: crawler purposes and controls
- Perplexity: search and user crawlers