ChatGPT Search

Quick definition

ChatGPT Search is the web search capability built into ChatGPT. When a question needs current or specific information, ChatGPT searches the web, reads the results, writes an answer, and attaches links to the sources it relied on. People can also start a search on purpose from the interface.

For SEO, ChatGPT Search matters because it is a separate discovery surface. It has its own crawler, its own way of turning a conversation into search queries, and its own referral traffic. Rankings in Google do not transfer to it automatically, and there is no Search Console for it.

A short history

OpenAI tested the idea in 2024 under the name SearchGPT, a prototype with a limited waitlist. On October 31, 2024, search launched inside ChatGPT for paid plans. It was extended to logged-in free users in December 2024 and, from February 2025, became available without an account in the regions where ChatGPT is offered.

In April 2025, OpenAI added improved product results for shopping questions, with images, prices, reviews, and links to buy. At launch, OpenAI stated that these product results were selected independently and were not ads. Features, availability, and policies keep changing, so check OpenAI's current help pages before you document details for a client.

How ChatGPT Search answers a question

OpenAI's help documentation describes the process in a few steps that are useful to separate:

1. Deciding to search. ChatGPT decides whether a question needs the web, for example because it asks about recent events, prices, or local options. A user can also trigger search directly.
2. Rewriting the prompt. ChatGPT usually turns the request into one or more targeted search queries. It may issue more searches after reviewing the first results. Context such as an approximate location or, where enabled, saved memories can shape those queries.
3. Retrieving results. OpenAI says it works with third-party search providers and also uses its own crawler, OAI-SearchBot, to surface pages.
4. Writing the answer. The response combines the retrieved information with the model's own knowledge. Inline citations and a sources list point to the pages used.

Step 2 is the one most SEO teams underestimate. The query that reaches the search provider is often not the words the person typed. Splitting a question into several narrower searches is a form of query fan-out, and it decides which pages get considered at all.

An example of query rewriting

OpenAI's own documentation gives a simple example. A person asks for good restaurants nearby, ChatGPT infers an approximate location, and the search it sends looks more like "top restaurants" plus the city name.

A hypothetical B2B example shows the SEO consequence. Someone asks: "Which project management tool works for a 12-person agency that bills hourly?" The searches behind the answer might target comparisons, time-tracking features, pricing pages, and reviews, each as its own query. A page that clearly answers "does this tool support hourly billing?" has more chances to be retrieved than a generic feature page that never states it.

You cannot see the exact rewritten queries for other people's conversations. Sometimes the interface shows what was searched, which helps when you test your own prompts. Treat those observations as samples, not as a complete keyword list.

The OpenAI user agents and robots.txt

OpenAI documents several user agents. Their jobs differ, and each one is a separate decision in robots.txt.

OpenAI notes that it can take about 24 hours for its systems to adjust after a robots.txt change. A typical configuration for a site that wants search visibility but no training use looks like this:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Many accidental blocks do not come from robots.txt at all. Bot protection at a CDN or firewall can reject crawler requests that robots.txt allows. Check server or CDN logs for blocked requests from OpenAI's published IP ranges before you conclude that ChatGPT ignores the site.

What makes a page usable as a source

OpenAI states that any public website can appear in ChatGPT Search and that there is no separate application process. Beyond crawler access, OpenAI does not publish a ranking guide. The following points are reasonable engineering assumptions, not documented ranking factors:

• Public and reachable. Content behind a login, a paywall without accessible text, or a consent wall that hides everything cannot be summarized.
• Key facts in the delivered HTML. OpenAI does not document in detail how its crawler handles JavaScript. Putting the main content in the initial response through server-side rendering or static output is the safer choice.
• Self-contained passages. Answers are assembled from specific statements. A paragraph that names the subject, the condition, and the answer is easier to quote than one that depends on three earlier sections.
• Checkable details. Dates, versions, prices, and scope statements help a system and a reader judge whether the information is current.
• Consistent entity facts. If your own site, your profiles, and third-party sources describe the product differently, the answer may repeat the wrong version.

ChatGPT Search is not Google

Some practices carry over: crawlable pages, clear structure, accurate content, and a solid reputation. Other things do not:

• There is no official ranking report or impression data from OpenAI for site owners.
• The search providers and the source mix behind an answer are not fully disclosed and can change.
• An answer may satisfy the question without a click, a pattern already familiar from zero-click search.
• A page that ranks well in Google can be missing from an answer, and a page that ranks modestly can be cited.

Do not take a Google ranking report as a proxy for ChatGPT visibility. Measure each surface on its own terms.

Measuring ChatGPT Search traffic

OpenAI's publisher FAQ states that ChatGPT adds utm_source=chatgpt.com to referral links. In analytics, that parameter, together with chatgpt.com as referrer, identifies most clicks from search answers. Some visits still arrive without a usable referrer, for example from apps, copied links, or privacy settings, so the measured number is a lower bound.

Useful reports show landing pages, engaged sessions, and conversions for this source, compared over time. Look at which pages receive the visits. They often reveal the questions in which your content was cited.

Measuring visibility inside answers

Traffic tells you about clicks, not about mentions. To learn how often a brand appears in answers, you need a versioned set of prompts, repeated runs, and a written coding rubric for mentions, citations, and recommendations. Generated answers vary between runs, so one screenshot proves very little.

If you compare brands across such a panel, you are calculating a kind of AI share of voice. Keep that figure separate from organic share of voice and state its numerator, denominator, run count, and date range.

A practical visibility check

1. Confirm in robots.txt that OAI-SearchBot is allowed for the sections you want cited.
2. Decide separately whether GPTBot may crawl the site for training.
3. Check CDN, firewall, and bot-protection rules for blocked OpenAI requests.
4. Make sure important pages return status 200, are public, and carry the main content in the delivered HTML.
5. Rewrite key passages so each one answers a specific question with its conditions.
6. Align facts about your company and products across your own site and major profiles.
7. Set up an analytics segment for utm_source=chatgpt.com and the chatgpt.com referrer.
8. Test a fixed set of real customer questions in fresh sessions, repeat them, and record the cited sources.
9. Review the results monthly and note every change in product, prompt, or method.

Common mistakes

• blocking OAI-SearchBot by accident while trying to block training crawlers;
• assuming a GPTBot block removes the site from ChatGPT Search;
• relying on robots.txt while a CDN rule silently rejects the crawler;
• treating one favorable or unfavorable screenshot as a trend;
• expecting Google rankings to predict citations in ChatGPT;
• hiding key facts in tabs, images, or scripts that may not be processed;
• reporting ChatGPT referral traffic as complete when some visits lose their referrer;
• buying "ChatGPT ranking" promises that no one outside OpenAI can verify.

The Crawl Foundry perspective

Crawl Foundry does not control or predict what ChatGPT Search cites. It helps with the parts you can check and with the work around them. The free AI crawler checker on the site tests how your robots.txt treats GPTBot, OAI-SearchBot, and ChatGPT-User. The free AI traffic checker finds visits from ChatGPT and other assistants in an analytics export you provide.

For ongoing work, you can connect an AI agent such as Claude, Codex, Cursor, or VS Code to Crawl Foundry's SEO MCP server. With Crawl Foundry for AI agents, the agent signs in through OAuth, works in exactly one workspace per session, stays within a budget with hard ceilings, and can only publish after your explicit approval. It can research keywords, check site health, and draft content from saved evidence. Whether an answer engine later cites that content remains an observation to measure, not a promise.

Related terms

• LLM visibility
• Generative engine optimization
• Answer engine optimization
• Cited source optimization
• Retrieval-augmented generation
• robots.txt
• llms.txt
• AI Overviews
• Google AI Mode
• Crawling

Review sources

• OpenAI: Overview of OpenAI crawlers
• OpenAI Help Center: ChatGPT search
• OpenAI Help Center: Publishers and developers FAQ
• IETF: RFC 9309, Robots Exclusion Protocol