ChatGPT searches for its own query, and it often has names in it already
OpenAI's help page on ChatGPT search says it "typically rewrites your query into one or more targeted queries" before sending them to its search partners. Its own example turns a question about cancer drugs into "CCR8 immunotherapy drug development 2025". The buyer never typed that year. ChatGPT added it.
We wanted to see what those rewrites look like for buying questions. Between October 4 and 13, 2026, we collected 2,502 answers from ChatGPT as a logged-out US visitor, across 1,530 buyer questions in 54 workspaces, with the searches each answer ran (6,428 in all). In 2,315 of them ChatGPT named at least one brand.
In 1,462 of those 2,315 answers (63%), one of ChatGPT's searches already contained a brand that the answer went on to name and the question never mentioned. In 1,107 (48%) it was in the very first search, the one ChatGPT runs before it has read any results. A made-up example of the shape: the buyer asks for the best project tool for a small agency, and the search reads "best project management software small agency 2026 Brand A Brand B Brand C".
A brand in the first search was named 89% of the time. A brand in no search, 7%
To test what that means for a single brand, we took every tracked competitor in every answer where the question didn't name it, and asked two things: was it in ChatGPT's searches, and did the answer name it?
- In the first search: named in 1,334 of 1,505 answers, 88.6% (95% interval 86.9% to 90.1%).
- In a later search only: named in 454 of 581, 78.1%.
- In no search at all: named in 1,151 of 16,427, 7.0% (6.6% to 7.4%).
Most brands tracking themselves never made it into the query
The brands workspaces track as their own appeared in ChatGPT's searches in 78 of 1,969 answers, 4%. When the first search held their name, the answer named them in 65 of 67 cases. When no search held it, 174 of 1,891, about 9%.
The gap held in both model versions that the ChatGPT page reported during the period. It says the shortlist is mostly settled before the search runs. The search then decides which of those names survive, and adds the odd newcomer from what it reads.
So "ChatGPT SEO" means ranking for a query ChatGPT writes
The searches had a recognisable style. ChatGPT added the current year to 1,386 of the 2,502 answers' searches (55%) when the question had no year in it. It added the word "official" in 725 (29%), and in 404 (16%) it ran a site: search limited to one website, in a sample we read usually a company's own. Reddit turned up in the searches for 21 answers, and an added "review" in 99.
That suggests two jobs, and they need different work:
- Be a name it brings with it. This is memory, and it builds slowly: describe your category in the same plain words everywhere, and get covered on the comparison and alternatives pages where the names it already lists are discussed. We can't see how much any single page moves this.
- Answer the query it writes. A page dated this year that names your category in plain words and compares you honestly with the brands ChatGPT already lists is the page that matches its search. Your own site needs to answer factual questions plainly, because site: searches go straight to it, and OpenAI says a site must allow its OAI-SearchBot crawler to be eligible at all.
Read the searches before you change a page
For each buyer question you care about, look at what ChatGPT searched, then sort the question into one of four boxes. In the search and named: hold your position. In the search and not named: the model expected you, then something it read talked it out of it, so read the cited pages. Named without being in the search: the reading worked, keep those pages current. In neither: that's the memory problem, and no single page fixes it quickly.
Proofsource shows the searches each engine ran next to every answer, where the engine reports them, alongside the brands named and the sites cited. ChatGPT, Perplexity and Claude report searches; Google AI Overviews doesn't.
Where this could be wrong
This is association, not cause. The same model writes the search and the answer, so a name in the search may simply show what it had already decided to say. Getting yourself into the query by some trick wouldn't guarantee the mention, and we haven't tested whether it would help.
We matched names as plain text, so a brand with a generic name could match a word it shouldn't. We left out names our system flags as ambiguous, and a hand check of 45 random matches found one doubtful case. The answers come from 54 workspaces across many categories and are pooled; answers to the same question aren't independent, so the true intervals are wider than the ones shown. The figure for workspaces' own brands rests on 67 answers from four workspaces.
We sampled ChatGPT logged out, from the US, over ten days. OpenAI says signed-in users' saved memories can shape the rewritten query, so a signed-in buyer may see different searches. We also only see the searches the page reports.
