How ChatGPT Search Works: OpenAI's 'Searching the Web With ChatGPT' Explained
99Visibility Research · Updated October 21, 2026 · 6 min read
Short answer
OpenAI's help article 'Searching the web with ChatGPT' says a site is eligible for ChatGPT search only if it allows OAI-SearchBot and its host or CDN allows traffic from OpenAI's published searchbot IP addresses. ChatGPT then rewrites the user's prompt into one or more targeted queries, sometimes sent to partner search providers. Saved memories can shape that rewrite. Eligibility is a gate, not a ranking.
Key takeaways
- ChatGPT search eligibility is a two-part condition: robots.txt must allow OAI-SearchBot and your host or CDN must let OpenAI's searchbot IPs through.
- ChatGPT does not search the exact prompt a user types; it typically rewrites it into one or more targeted queries.
- Saved memories can change the rewritten query, so logged-in tests with memory on are not reproducible.
- OpenAI's only ranking disclosure is 'multiple factors', and it says placement is not guaranteed.
- Check robots.txt and the firewall together, because either one alone can silently remove you from ChatGPT search.
What the article is and who published it
'Searching the web with ChatGPT' is a help-center article published by OpenAI. It explains, for users and site owners, how ChatGPT's search feature finds and shows web pages.
It is first-party documentation, not a study. OpenAI describes its own system. Nobody outside OpenAI tested the claims, and the article does not report measurements.
That makes it useful for one thing: knowing what OpenAI says the rules are. It pairs with OpenAI's bot documentation, which defines the crawlers the help article refers to. Our digest of OpenAI's crawler documentation covers that page.
How we read it
99Research read the article as part of a verified evidence base built on 2026-08-22. Every claim we quote was checked against the live page at that date.
Two limits apply to this kind of source:
- No fixed date. The article shows only a relative update time (it read 'Updated: 20 hours ago' when fetched). You can't tell from the page when a sentence changed.
- Hard to monitor. help.openai.com returns a 403 to non-browser clients, so automated change tracking can't read it directly.
So treat everything below as accurate as of our verification, and re-read the article before acting on a detail. As of October 2026, we know of no change to the points below.
What the article says
Eligibility is a conjunction
The key sentence for site owners: 'To make a website eligible for inclusion, allow OAI-Searchbot to crawl the site and confirm that the website host or content delivery network allows traffic from OpenAI's published searchbot IP addresses.'
The word that matters is 'and'. A robots.txt allow is not enough if your firewall blocks the crawler. OAI-SearchBot is the search crawler. GPTBot is the training crawler, and OpenAI's bots page says 'Each setting is independent of the others.'
Prompts get rewritten
OpenAI states that 'ChatGPT search typically rewrites your query into one or more targeted queries that it sends those providers.' This is query rewriting. The search that runs is not the sentence the user typed.
Other search providers are involved
The article says ChatGPT search 'sometimes partners with other search providers', naming Microsoft. The wording is hedged. It does not say how often, or for which queries.
Memory shapes the rewrite
'If memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query.' For product results, OpenAI says ChatGPT considers 'your query and context (such as Memory or Custom instructions).'
Ranking is undisclosed
OpenAI's full ranking disclosure is: 'ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed.' No factor is named.
What the article does not show
The article tells you how to become eligible. It does not tell you how to be chosen.
- Eligibility is not retrieval. Allowing OAI-SearchBot and opening your firewall puts you in the pool. It doesn't make ChatGPT fetch or cite you for any query.
- Search doesn't always run. ChatGPT searches at the model's discretion. Some answers come from the model's training data, with no retrieval at all.
- The real queries are hidden. Because prompts are rewritten, you can't see what was actually searched. Tracking the prompts you chose measures a proxy. See whether AI searches for the exact question.
- No provider share. 'Sometimes partners' gives no frequency. The named-provider language traces to 2024-era documentation, so don't treat Bing presence as a measured route into ChatGPT.
- No ranking factors. 'Multiple factors' rules out any honest claim about what ChatGPT rewards.
- Answer-time fetching is separate. OpenAI's bots page says robots.txt rules 'may not apply' to ChatGPT-User, which acts on a user's request. The help article's eligibility rule is about OAI-SearchBot.
What it means for testing ChatGPT visibility
Memory and custom instructions change the rewritten query. So two people asking the same question can trigger different searches and get different sources.
If you test how ChatGPT answers buyer questions, control what you can:
- Test logged out, or with memory off and custom instructions cleared.
- Record whether the answer showed citations. No citations usually means you can't tell whether search ran.
- Repeat each question more than once. One answer is a sample, not a verdict.
- Compare like with like. A logged-in, memory-on result is not comparable with a clean session.
Say you run a hypothetical accounting firm in Leeds. Your own ChatGPT account, full of saved notes about your firm, is the worst place to check whether strangers see you. A clean session is closer to what a prospect gets.
What to do: check robots.txt and the firewall together
The help article gives you two gates. Check both, because failing either one is silent.
Gate one: robots.txt
Make sure OAI-SearchBot is not disallowed. A common error is a blanket AI block that removes search eligibility while the owner meant to stop training. To block training but stay in ChatGPT search:
User-agent: GPTBot
Disallow: /
User-agent: OAI-SearchBot
Allow: /
Gate two: host, CDN and firewall
Request a real content URL with the OAI-SearchBot user agent and look at the status code. A 200 passes. A 403, a 429, or a JavaScript challenge page fails. OpenAI's own guidance notes that bot-protection systems 'can mistakenly block legitimate crawlers, often returning 403 Forbidden errors.'
Things to know before you change firewall rules:
- OpenAI publishes its search crawler ranges at openai.com/searchbot.json. When we checked on 2026-08-22, it listed 35 IPv4 prefixes and no IPv6 prefixes. An origin that only answers over IPv6 may be unreachable; see whether ChatGPT's crawlers can reach an IPv6-only site.
- OpenAI also warns against IP-only allowlisting. Read whether to allowlist OpenAI's crawler IPs before you hard-code ranges.
- From 2026-09-15, Cloudflare blocks Training and Agent crawler categories by default on ad-displaying pages for new domains and existing free-tier customers, with Search left allowed. Recheck if you're on Cloudflare.
Our page on OpenAI crawlers, robots.txt and firewalls walks through both gates. The AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away.
Passing both gates makes you eligible. What ChatGPT does next is outside your control.
Frequently asked questions
Does allowing GPTBot help me appear in ChatGPT search?
No. OpenAI's bots page describes GPTBot as a training crawler only and asserts no search benefit. ChatGPT search eligibility depends on OAI-SearchBot. OpenAI says each setting is independent, so you can block GPTBot to opt out of training and still allow OAI-SearchBot to remain eligible for search answers. Blocking GPTBot does not remove you from ChatGPT search.
What happens if I block OAI-SearchBot?
OpenAI's bots page says sites opted out of OAI-SearchBot 'will not be shown in ChatGPT search answers, though can still appear as navigational links.' Blocking it at the firewall has the same practical effect as a robots.txt disallow, except you may not know it happened. Check your server logs and response codes for the OAI-SearchBot user agent.
Can I see the queries ChatGPT actually ran for my brand?
Not from ChatGPT. OpenAI says it rewrites prompts into one or more targeted queries but exposes none of them to site owners. The rewritten queries can also vary with each user's memory and custom instructions. Any tool that tracks prompts is measuring answers to questions it chose, which is a proxy for the real retrieval input.
Does being in Bing's index get me into ChatGPT?
The help article says ChatGPT search sometimes partners with other search providers, naming Microsoft. It gives no frequency, and nobody has published a measured effect. Being in Bing's index is reasonable hygiene and it backs Microsoft Copilot directly. Don't treat it as a proven path into ChatGPT answers.
Why does my own ChatGPT account show different results from my colleague's?
Memory and custom instructions. OpenAI says ChatGPT may use saved memories when rewriting a search query, and product results consider context such as memory or custom instructions. Different accounts can produce different searches and sources for the same question. Test in a logged-out session or with memory off to get comparable results.
Sources
This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.