# ChatGPT SEO: how ChatGPT search picks the sources it cites

> In short: When a question needs current information, ChatGPT searches the web: it writes its own queries, often longer and more specific than the user's, reads some of the pages that come back, and cites them. To be among them, your pages must be readable by OAI-SearchBot, rank for the searches ChatGPT writes, and state the answer plainly enough to be used.

Guide from Keats, published 2026-11-13, updated 2026-11-13. Canonical: https://getkeats.com/guides/chatgpt-search-sources

"ChatGPT SEO" sounds like a new discipline. Most of it is an old one applied to a new reader: ChatGPT's search is a reader that writes its own queries, skims the results and quotes what it can use. Understand that reader and the work is familiar.

## When ChatGPT searches, and when it doesn't

ChatGPT doesn't search for every question. Definitions and well-settled facts often come from the model's memory, with no citations. Questions that need current or specific information, such as products, prices, recent events and comparisons, usually trigger a search, and the answer then shows its sources. In our own sample of eight buying and how-to questions, seven answers cited pages and one, a definition, cited none ([the sample, as a CSV](https://getkeats.com/data/chatgpt-sample-2026-11-12.csv)).

Perplexity is different: it searches on nearly every question, as [our Perplexity guide](https://getkeats.com/guides/perplexity-citations) explains. For marketers, ChatGPT's split is convenient: the questions that matter most commercially are the ones answered from the web, which means they can be influenced by what's on the web.

## Step one: it writes its own searches

ChatGPT doesn't simply forward the user's question to a search engine. It writes one or more searches of its own, called [fan-out searches](https://getkeats.com/glossary/fan-out-searches). They tend to run longer, and to be more precise, than the user's own words, and they can include brand names, years and qualifiers. In our sample, a general question about AEO tools produced one search that already listed five products and the year.

This matters in two ways. The pages that rank for those specific searches are the ones that get read. And if the searches already name rivals, the model has an association you'll need third-party pages to break.

## Step two: it reads a few results

From the results, ChatGPT reads a handful of pages, not dozens. Which ones depends on how they rank for its searches and how useful they look. Being readable is a precondition: OpenAI says OAI-SearchBot is used "to surface websites in search results in ChatGPT's search features", and that sites opted out of it "will not be shown in ChatGPT search answers". Past that, it picks from search results, so a page that ranks well in ordinary search for those long queries is more likely to be in front of it.

## Step three: it writes, and cites

The answer is written from what it read, and the pages it drew on appear as citations: inline links and a list of sources. A page can be cited without the brand being named (an explainer quoted for a fact), and a brand can be named without its own page being cited (named from a roundup). Track both.

## What decides whether your page is used

- **Readable**: OAI-SearchBot allowed, content in the HTML, not behind a login or a script the crawler can't run. Our guide to [AI crawlers and robots.txt](https://getkeats.com/guides/ai-crawlers-robots-txt) has the checks.
- **Found**: ranking for the searches ChatGPT writes, which are long and specific. A page that answers the specific question ranks for the specific search.
- **Usable**: the answer is near the top, stated plainly, with specifics a model can repeat.
- **Corroborated**: for buying questions, other pages (roundups, reviews, threads) say the same thing about you.

## Which kinds of page it cites

The pattern in our sample, and in most categories we've looked at:

| Question | Pages it tends to cite |
|---|---|
| Which product for X | Roundups, review sites, vendors' product pages |
| Alternatives to a named product | Vendors' comparison pages, roundups |
| A named product's features or price | That vendor's own pages and help docs |
| How to do X | Explainers from publications, documentation, guides |

The third row is the easiest win: when ChatGPT looks you up by name, it often reads your own site. Make sure your product, pricing and help pages state the facts you want repeated.

## What doesn't change ChatGPT's choice

- **Blocking GPTBot.** It's the training crawler; it doesn't affect search answers either way.
- **A file promising AI visibility.** An llms.txt file can help agents navigate your site; it isn't what ChatGPT search ranks by.
- **Repeating the question twenty times.** Stuffing doesn't make a page more useful to a reader that skims for the answer.

## How to see it in your own analytics

Clicks from ChatGPT arrive with chatgpt.com as the referrer, and links ChatGPT shows often carry a `utm_source=chatgpt.com` tag. Our guide to [tracking AI traffic in GA4](https://getkeats.com/guides/ai-traffic-ga4) sets up a channel for it. Remember that many readers never click, so the traffic undercounts how often you're cited.

The work that follows from all this, in order, is in our guide to [getting named in ChatGPT](https://getkeats.com/guides/get-named-in-chatgpt). The bigger picture, including where the model's memory fits, is in [how AI chooses which brands to recommend](https://getkeats.com/guides/how-ai-assistants-choose-brands).


## Your agent does this

Ask your agent: "For each of our buying questions, show me the searches ChatGPT ran and the pages it cited, and which of those pages we rank for." It reads `explain_prompt` and `list_citations`, and turns pages that cite rivals but not you into cards.

## Sources

- [OpenAI: Overview of OpenAI crawlers (OAI-SearchBot)](https://developers.openai.com/api/docs/bots), read 2026-11-13
- [Keats: eight ChatGPT answers to buyers' questions about AEO tools (CSV)](https://getkeats.com/data/chatgpt-sample-2026-11-12.csv), read 2026-11-13