What ChatGPT Actually Searches When You Ask It a Question
You ask ChatGPT one question. It runs three searches you never see. They don't look like your question, and half of them name vendors you didn't.
When ChatGPT searches the web before answering, it doesn't search for your question. It writes its own searches, runs them, reads the results and then answers. OpenAI's help pages say as much: ChatGPT search "typically rewrites your query into one or more targeted queries."
You never see those searches. They decide which pages get read, and the pages decide the answer.
We record them. Here is what they look like.
TL;DR
How we got this
When ChatGPT searches, the answer carries the searches it ran. We store them next to the prompt and the answer.
This analysis covers 105 answers from the consumer ChatGPT app between August and September 2026. That's 335 searches in total, from 93 buyer-style prompts about B2B software across three brands we track. It's a modest sample from one kind of buyer, and we say where that limits things below. The examples in this post are paraphrased into other categories, with names swapped for placeholders, so no customer's prompts appear here.
One question becomes three searches
| Searches per answer | Share of answers |
|---|---|
| 1 | 14% |
| 2 | 29% |
| 3 or more | 57% |
The median was 3 and the mean 3.2. The three brands in the sample averaged 3.4, 3.0 and 2.8, so this isn't one brand's prompts skewing the result.
Not every answer searches. Across all the app answers we collected in the same period, 31% didn't search at all and answered from what the model already knew. Those answers carry no citations, and in our comparison of the app and the API they almost never named the brand being tracked.
The searches don't look like your question
A typical buyer prompt in our set runs about 12 words and is phrased as a question. ChatGPT's searches run about 9 words and read like a power user's keyword string.
Here's the pattern, paraphrased into a different category:
Prompt: "Why does my [platform] sync keep failing and how do I find the root cause?"
Searches: "[platform] docs sync error codes retry backoff" · "[platform] docs view sync logs failure reason" · "[platform] docs connection timeout troubleshooting"
The user asked one fuzzy question. ChatGPT turned it into three specific lookups, each aimed at documentation.
It decides who's in the running before it searches
This was the finding that surprised us most. Half of all searches, and nearly two-thirds of answers, named a product the user never mentioned.
Prompt: "Best alternatives to [Incumbent] for combined invoicing and expense tracking?"
Searches: "best [Incumbent] alternatives invoicing expense tracking [Vendor 1] [Vendor 2] [Vendor 3] [Vendor 4] 2026" · "site:[vendor-1-domain] invoicing expense tracking" · "site:[vendor-2-domain] expense management invoicing"
Read that first search again. The shortlist is already written into it. ChatGPT isn't searching to find out who's in the running. It's searching to check the vendors it already picked.
That changes what "being found" means. If you're not on the list the model carries into the search, the search is unlikely to go looking for you. Our working theory is that the list comes from what the model already absorbed about your category before anyone asked. The best evidence we have on what shapes that is where AI gets its sources: third-party roundups, comparisons and community threads, far more than vendor sites.
We measured this by matching searches against a list of 932 product names from our tracking data. Anything not on that list went uncounted, so the real share is probably higher.
It reads your docs to check you
A third of all searches asked for "official" pages or "docs". Nearly one in five used site: to search inside a single vendor's website.
Prompt: "Comparing [Tool A] versus [Tool B] versus [Tool C] for handling large teams"
Searches: "[Tool A] large teams permissions official docs" · "site:[tool-b-domain] docs workspace limits enterprise" · "[Tool C] enterprise scaling official documentation"
That looks like fact-checking. Once a vendor is on the list, ChatGPT goes to the vendor's own pages to confirm specifics: limits, permissions, pricing, integrations.
There's a wrinkle. In our citation study, brands' own pages almost never got cited when AI recommended them. Both findings can be true. Reading a page to check a fact isn't the same as citing it as the reason for a recommendation. Your docs decide whether the facts come back right. Third parties decide whether you get recommended.
"Best X 2026" is the wrong target
SEO trained a generation of marketers to chase "best [category] [year]" keywords. ChatGPT's own searches barely use them:
| Word in the search | Share of searches |
|---|---|
| official / docs | 33% |
| site: | 18% |
| 2026 | 15% |
| pricing / cost | 5% |
| open source | 2% |
| alternatives | 2% |
| best | 1% |
| review | under 1% |
| under 1% | |
| vs | 0% |
This is one kind of buyer, B2B software, so don't read too much into the order at the bottom of the table. The gap between the top rows and "best" is too big to be noise, though.
Reading isn't citing
More searches mean far more pages read. Citations barely move:
| Searches in the answer | Results read (average) | Citations (average) |
|---|---|---|
| 1 | 9 | 4 |
| 2 to 3 | 23 | 5 |
| 4 or more | 55 | 8 |
On the biggest answers, roughly one result in seven made it into the citations. Being read is table stakes. Being cited is what shows up in the answer.
The API did something else entirely
Before August, we measured ChatGPT through the OpenAI API. Its searches looked nothing like the app's. Of 95 API answers that searched, almost every one ran exactly one search: usually the prompt squashed into keywords, sometimes almost word for word:
Prompt: "How do I choose a password manager for a small team?"
Search: "how to choose password manager for small team 2024"
That "2024" is not a typo. A quarter of the older API model's searches in mid-2026 added a year that was two years stale. If your tracking tool measures ChatGPT through the API, it may be measuring a search behaviour your buyers never see.
What to do about it
FAQ
What is query fan-out in ChatGPT?
When ChatGPT searches the web, it rewrites the user's question into its own search queries, runs them and answers from the results. In our data from the ChatGPT app, the median answer that searched ran 3 queries, and the most was 10.
Does ChatGPT search for the exact question I type?
No. In our sample none of ChatGPT's searches contained a question mark, and none were near-copies of the prompt. On average a search shared under a quarter of its words with the original question.
How does ChatGPT choose which brands to search for?
It often names them in the search itself. Half of the searches we logged named a product the user never mentioned, which suggests ChatGPT forms a shortlist from what it already knows and then searches to check those vendors.
Should I optimise for "best [category] [year]" keywords for AI search?
They matter less than you'd think. "Best" appeared in about 1% of the searches ChatGPT ran, while about a third asked for official pages or documentation.
Does ChatGPT cite every page it reads?
No. Answers that ran four or more searches read about 55 results on average and cited about 8.
Method: 105 answers from the consumer ChatGPT app, August to September 2026, carrying 335 searches, from 93 English buyer-intent prompts about B2B software across three tracked brands. One brand supplied about 60% of the answers; per-brand averages are reported above. API figures come from 95 answers collected June to August 2026. Product names were matched against a list of 932 names from our tracking data, so that share is a lower bound. Examples are paraphrased into other categories with placeholder names.
Related Articles
Six Questions to Ask Your AI Visibility Data
A score is a thermometer. It tells you there's a fever. Six questions that tell you why, and what to do about it.
Ask Renown From Claude and ChatGPT
Your AI visibility data, answered by the AI you already use. Connect once, then ask in plain language.
How AI Chooses Its Sources: 3,543 Citations, Analyzed
AI doesn't read your website. It reads Reddit, YouTube, Wikipedia, and a handful of independent blogs. We have the data: 3,543 citations across two engines and 60 questions.
