ChatGPT Search Visibility: Discoverability, Source Links, and Measurement

A website becomes discoverable in ChatGPT’s search features when OpenAI’s search crawler, OAI-SearchBot, is allowed to reach it. That is the one access requirement OpenAI documents. Everything after that, whether a page is shown in an answer, which sentence it supports, and whether a reader clicks through, is decided by systems OpenAI has not published. A useful approach to ChatGPT visibility therefore has three parts: remove the access barriers you control, make pages easy to quote accurately, and measure what can actually be measured.

The access layer: OAI-SearchBot, not GPTBot

OpenAI uses separate robots.txt tokens for separate purposes, and its documentation says each setting is independent (OpenAI’s crawler documentation).

• OAI-SearchBot “is used to surface websites in search results in ChatGPT’s search features”. OpenAI states that sites opted out of it “will not be shown in ChatGPT search answers, though can still appear as navigational links”, and recommends allowing both the token and OpenAI’s published IP ranges for sites that want to appear.
• GPTBot collects content that may be used to train foundation models. Disallowing it is a training decision; OpenAI’s example of independent settings is exactly a site that allows OAI-SearchBot while disallowing GPTBot.
• ChatGPT-User visits pages when a user acts inside ChatGPT or a custom GPT. OpenAI says it “is not used to determine whether content may appear in Search” and that, because the actions are user-initiated, “robots.txt rules may not apply”.

The consequence: a site that blocks GPTBot to avoid training has not removed itself from ChatGPT search, and a site that blocks OAI-SearchBot has removed itself from search answers even if GPTBot is allowed. OpenAI notes it can take around 24 hours after a robots.txt change for its search systems to adjust.

An eligibility and troubleshooting checklist

Use this list when a site does not seem to appear in ChatGPT search, or appears less than expected. Each item is something you can check; none of them guarantees appearance.

1. robots.txt. Confirm there is no Disallow for OAI-SearchBot, either by name or through a User-agent: * group that blocks the relevant paths. Remember that a named group replaces the * group for that crawler.
2. Firewall and bot management. CDN or WAF rules that challenge or block unfamiliar bots can stop OAI-SearchBot even when robots.txt allows it. OpenAI’s recommendation to allow its published IP ranges addresses exactly this. Check your security tool’s logs for blocked requests carrying the OAI-SearchBot user agent.
3. Server responses. Pages that return errors, time out or serve an empty shell to non-browser clients give a crawler nothing to work with. Fetch a key URL with a plain HTTP client and compare the HTML with what a browser renders.
4. Content in text. Information shown only inside images, canvas elements or interactive widgets is harder for any crawler to read. Put the facts you want quoted into indexable HTML text.
5. Log evidence. Search server logs for OAI-SearchBot requests and verify the IP addresses against OpenAI’s published list. Requests that only claim the name but come from other networks are not OpenAI.
6. Time. After unblocking, allow for the adjustment period OpenAI describes before concluding the change had no effect.

The source-link layer: what can and cannot be influenced

When ChatGPT search answers a question, it can show links to the pages it drew on. OpenAI has not published a ranking or citation algorithm for which sources are shown, so no checklist can promise a citation. What a site can influence is whether its pages make good sources when they are retrieved:

• One page, one clear answer. A page that answers its main question directly, near the top, offers a passage that can support a statement without distortion.
• Specific, verifiable facts. Dates, version numbers, definitions and conditions written plainly are easier to cite accurately than vague claims.
• Visible authorship and dates. Readers who click through judge credibility on the page itself; an author name, an update date and source links help that judgment.
• Consistency. If a product price or policy differs between two pages on the same site, any system quoting the site can quote the wrong one.

These are the same practices that make content useful in any search product. They are listed here because they are within a site owner’s control, not because OpenAI has documented them as citation factors.

The measurement layer

OpenAI’s crawler documentation describes no site-owner report comparable to Search Console, so measurement relies on three sources you already own.

Analytics referrals. Sessions that arrive from ChatGPT links can carry a referrer from a ChatGPT domain and, in some cases, a campaign parameter in the landing page URL. In your own data, check landing page URLs for a utm_source value naming ChatGPT and check the session source dimension for ChatGPT domains. Building those segments in GA4 is a separate configuration task with its own pitfalls.

Server logs. OAI-SearchBot requests show which URLs OpenAI’s search crawler fetched and when. Crawl activity is evidence of discoverability, not of appearance in answers.

Branded and direct demand. Some readers see a site named in an answer and later visit by typing the address or searching the brand. That traffic lands in direct or branded organic channels and cannot be attributed to ChatGPT with certainty.

Two limits follow. Referral sessions undercount visibility, because an answer can be read without any click. And a cited link does not guarantee a session: the reader may not click, privacy settings may strip the referrer, and some clients open links without passing referral data. Referral data is a subset of AI visibility, not a measure of it.

When a security rule blocked OAI-SearchBot

Suppose a software documentation site allowed GPTBot and OAI-SearchBot until a security change. After the change, ChatGPT referral sessions in GA4 drop to almost none within a week.

1. robots.txt is unchanged and allows OAI-SearchBot.
2. WAF logs show requests with the OAI-SearchBot user agent receiving a challenge page since the security change.
3. IP check: the blocked requests come from addresses on OpenAI’s published list, so they are genuine.
4. Fix: an allow rule for OpenAI’s published ranges is added for OAI-SearchBot only; GPTBot stays blocked because the owner decided against training use.
5. Verification: check that logs show successful OAI-SearchBot fetches after the change, then watch whether referral sessions recover over the following weeks.

The example shows the order of checks: access first, then evidence in logs, then outcomes in analytics. Unblocking does not guarantee recovery; other changes in ChatGPT’s systems could also move the numbers.

ChatGPT visibility claims to avoid

• That allowing OAI-SearchBot guarantees appearance or citation. OpenAI describes it as the crawler used to surface sites; it does not promise inclusion.
• That ChatGPT search ranks pages by a known set of factors. OpenAI has not published one.
• That GPTBot controls search visibility. OpenAI documents it as the training crawler.
• That ChatGPT referral sessions equal ChatGPT visibility. A view can end without a session.
• That a user agent string proves a request came from OpenAI. Verification uses OpenAI’s published IP ranges.

Where ChatGPT search fits

ChatGPT search is one AI answer product among several, each with its own crawler tokens, controls and measurement gaps. Google’s AI features rely on Googlebot and Search Console’s controls; Anthropic documents Claude-SearchBot for its search features (Anthropic’s crawler documentation). Treat each platform’s documentation as the authority for that platform, and avoid carrying a rule from one provider to another.