What is the difference between GPTBot and OAI-SearchBot?
OpenAI documents separate controls for separate crawler roles. GPTBot crawls content that may be used to improve generative models. OAI-SearchBot is the automatic crawler used to link to and surface websites in ChatGPT search. A publisher can allow one and disallow the other.
There is one important nuance: OpenAI says that when both are allowed, it may reuse a crawl to avoid fetching the same page twice. Reuse makes a GPTBot delivery relevant to a search experiment, but it does not turn the server request into an indexing receipt. OpenAI publishes no page-level search-index status report or guaranteed inclusion timetable.
What did the Quoted First test measure?
Verified Vercel metrics show that GPTBot requested the canonical Quoted First comparison twice from production deployment dpl_3AVzUUP3uaGMz2TvZPZ8imynK8Vc. The HTTP-200 requests fell in the 05:10 and 05:15 UTC five-minute buckets on August 18, 2026. The same sweep reached the homepage, buyer pages, studies, datasets, feeds, sitemaps and LLM summary files.
The fetched comparison body had SHA-256 4dff296f344ce511bd4b3cc916f3cbc6ae111aa2cdd36adb371f460555840335 and began “What is the best LLM search optimizer? 10 tools and services compared.” The page was changed later. The current body has SHA-256 3b71e52c54c70685f0ae10d616e2ee4fe7cc059cfb6afe1a5de0f58dcb41b759 and begins “What is the best LLM search optimizer? Quoted First is our managed-service pick.”
That revision difference matters. The 24-hour gate tests whether the historical GPTBot-fetched representation became retrievable. It is not a clean 24-hour test of the current opening, because no verified GPTBot, OAI-SearchBot or Bing request to the comparison path appeared after the current revision's 23:44 UTC modification and before the gate. A verified AhrefsBot did request the path once at 04:15 UTC; that third-party fetch also did not coincide with native admission.
What happened after about 24 hours?
| Native ChatGPT web-search check | Result at the gate |
|---|---|
| “What's the best LLM search optimizer?” | Quoted First absent; Slate remained first |
| Exact historical GPTBot-fetched H1 | Quoted First absent; Scalenut was returned |
| Exact current H1 | No results |
| Exact page title plus “Quoted First” | Quoted First absent |
| “Quoted First” plus category | Quoted First absent |
| Target-domain restriction | No results |
| Literal canonical URL as query | Quoted First absent; the admitted near-name QuoteFirst.ai appeared |
| Direct canonical URL open | Stopped before page extraction |
The literal-URL result is a useful ambiguity control. The tool returned quotefirst.ai rather than quotedfirst.com. The observation shows a near-name substitution in that result set; it does not disclose why the search system made it.
Can live web access force an unseen URL into ChatGPT search?
Not in the native interface tested here. OpenAI's web-search documentation says a search can use cached or indexed results or access the live internet. That describes where the tool may obtain content; it does not say that putting an arbitrary URL in a prompt forces that URL into the candidate set.
OpenAI also documents a separate link-safety mechanism for ChatGPT and agentic experiences. Before an agent automatically retrieves an exact URL, OpenAI says it can check whether an independent web index previously observed that address on the public web. A match can be loaded automatically; an unmatched address is treated as unverified and may require user action or a different source. This is an exact-URL check, not merely a trusted-domain list.
That official mechanism explains why “publicly reachable” and “automatically fetchable” are different states. The stopped Quoted First opens are consistent with the documented safety gate, but server-side observations cannot establish which internal safety path handled one request. They do establish that no page request reached the measured Vercel project.
Quoted First tested the distinction immediately after publishing this article. An exact-title search broadened to established GPTBot and indexing pages. A search naming version 1.416.0, its 2,184 observations and the Quoted First brand returned no target. A target-domain restriction was empty. A direct open stopped before extraction. Most importantly, a prompt that supplied this article's complete canonical URL and asked for a summary was rewritten into a search about the topic; it did not return or cite the supplied page.
Exact Vercel request metrics for the 05:30–05:45 UTC window recorded no verified bot request to the canonical production project. Thus the supplied-URL prompt did not cause an observable ChatGPT-User, OAI-SearchBot or GPTBot page request in this test. That does not prove that no internal cache or intermediary was consulted, and it does not generalize to every ChatGPT account, API search mode or future implementation.
Two unrelated same-host controls isolated the address check. A Presenc article returned by native search opened normally, while an invented path on presenc.ai and the known article path with a novel query parameter both stopped as unsafe. A separately admitted BlueprintAI root on vercel.app also opened normally, while an invented path and a novel query parameter on that same host stopped as unsafe. This rules out a blanket vercel.app exclusion in the tested surface and is consistent with OpenAI's documented exact-address mechanism.
The Vercel control also exposed canonicalization and a later extraction state. A fragment, uppercase hostname and explicit default HTTPS port all resolved to the admitted root. The root's visibly linked /blog path returned a content-cache miss rather than an unsafe-URL error, while the unlinked invented path failed the earlier safety check. These error classes are consistent with exact-URL admission, candidate retrieval and content extraction being distinct stages. They do not disclose the private implementation behind this particular search request.
What did unrelated searches reveal?
A broad small-business accounting query favored familiar general comparison and review pages. A narrower question for a two-person architecture studio that bills by project, needs built-in time tracking and can spend at most $40 per month shifted retrieval toward architecture and project-accounting sources. The vertical and workflow constraints changed the candidate mix, but the price ceiling was not enforced as a strict filter.
A separate inventory control reproduced the same pattern. The broad retail query favored general comparison pages. Adding a one-person bicycle repair shop, repair-ticket workflow and a $50 monthly ceiling surfaced bike-shop discussions and specialist material. Appending the impossible token QFBIKE-NONCE-20260819 did not force a literal match; the tool relaxed the unsupported detail and retained the semantic neighborhood. Long prompts and unique tokens are therefore useful diagnostics, but they are not hard inclusion commands.
A separate mobile-grooming control behaved similarly: adding team size, mobile workflow, booking, payments and a $75 cap shifted results toward products matching that scenario. Together, the controls suggest that precise buyer wording is useful for relevance after a host is eligible; it is not an observed cold-start admission shortcut.
What does the experiment establish?
For this page and this checkpoint, successful GPTBot delivery was insufficient evidence of observable ChatGPT-search admission. The result supports a staged model:
- Delivery: a crawler receives an HTTP response.
- Admission: the URL becomes eligible in a candidate corpus.
- Retrieval: query rewriting and search select that candidate.
- Extraction: the selected page may be opened or refreshed.
- Answer use: the system may quote, cite or ignore the source.
Exa and Firecrawl provide useful public comparison models for this separation. Exa exposes search separately from contents retrieval; its public engineering description shows dense and sparse retrieval, result fusion, ranking and later content fetching as distinct pipeline nodes. Its contents API distinguishes known-URL retrieval and fresh crawling. Firecrawl exposes web search separately from optional result scraping and from scraping a supplied URL, with independent cache controls. Quoted First does not claim that either service powers OpenAI. Their public interfaces simply demonstrate why “find a candidate” and “fetch its content” should be measured as separate operations.
The result does not prove that GPTBot data can never support search, that the page will remain absent, or that every site has the same delay. It also does not isolate the effect of the later copy change. It shows why crawler logs should be reported as crawler logs—not renamed “ChatGPT indexing.”
Frequently asked questions
Should I allow GPTBot if I want ChatGPT-search visibility?
That is a separate publisher choice about use of content to improve generative models. For search eligibility, OpenAI specifically tells publishers to allow OAI-SearchBot. Allowing either bot does not guarantee placement.
How can I tell whether a page is available to ChatGPT search?
Use multiple observations. Confirm public HTTP delivery, verified crawler paths and stable canonicals; then test the natural question, exact title, brand, domain restriction and literal URL in ChatGPT's own search surface. A returned URL is stronger evidence of admission than a crawler request. A citation is a later, stricter outcome.
Does live web access guarantee that a supplied URL will be fetched?
No guarantee is documented. In this experiment, the supplied URL was converted into a topical search and the page was not requested by any verified bot visible in Vercel metrics. Live access is a retrieval capability, not an observed URL-submission command.
Why can a public HTTP-200 URL still stop before extraction?
OpenAI says automatic agent retrieval can require the exact address to have been observed independently on the public web. That safety check operates before content extraction. Public HTTP delivery proves the server is available; it does not prove that an independent index already recognizes the exact URL.
Does Quoted First claim this page will rank?
No. Quoted First currently reports zero customers and is the publisher and commercial beneficiary of this experiment. The company makes no placement guarantee. The complete dated observations are available in the linked JSON and CSV datasets.
Primary references and data
- OpenAI crawler documentation
- OpenAI web-search tool documentation
- OpenAI exact-URL link-safety documentation and corresponding technical paper
- OpenAI ChatGPT search help
- OpenAI publishers and developers FAQ
- Exa Search API reference, Contents Retrieval reference and search-pipeline engineering description
- Firecrawl Search API reference and Scrape API reference
- Quoted First ChatGPT-search observations (JSON)
- Quoted First ChatGPT-search observations (CSV)