What happens before the answer is written
When ChatGPT has search on, it does not read your question and then survey the market. It rewrites your question into its own search query, retrieves a small set of pages, and composes the answer from what came back. The brands in that first retrieved set are the ones with a real chance of being named. Everything else is background.
Suganthan Mohanadasan tested this in August 2026 by watching ChatGPT's own first query. Brands that appeared in that first query reached the final answer 68.9% of the time. Brands that were merely retrieved later reached it 2.1% of the time. The author calls the study directional, and it is: one account, a few dozen conversations. But the shape matches what anyone sees when they run the same buying prompt repeatedly. The shortlist forms early, and the answer is written from it.
So the question "why does ChatGPT recommend my competitor" usually has a plain answer. Your competitor has a page that matches the job the buyer asked about, in words the model's own query uses. You have a homepage, a features page and a blog post with a different title. On query one, the competitor is retrieved. You are not.
Named is not the same as cited
Before you diagnose, separate two things the answer can do with a brand. It can name you, which means your brand appears in a sentence. It can cite you, which means one of your URLs is linked as a source. These move independently. A brand can be named from the model's memory with no link at all, and a URL can be linked while the brand goes unmentioned. Semrush's June 2026 study of 3,981 domain appearances found that 61.7% of AI citations were exactly that second case: the source was linked, the brand was never said.
When you look at a competitor's answer, note which of the two you are seeing. A competitor that is named and cited on the same prompt owns that job. A competitor that is only cited has a page in the shortlist but no entity strength. A competitor that is only named is living on memory, which is the easiest position to displace with a page that answers the job.
How to check it on your own prompts
- Write down the question a buyer actually types. Not "best AI visibility platform", but the job: "tool to see why ChatGPT recommends a competitor for our category".
- Run it in ChatGPT with search on. Record three things: which brands are named, which URLs are cited, and whether your domain is in either list.
- Run the same prompt again, at least a day apart. AI answers move between runs. In a frozen study of 30,504 production responses, the mention state of a prompt flipped between consecutive runs 9.7% of the time. One run is an anecdote.
- Open the cited URL for the competitor that keeps winning. Read its H1 and first paragraph. That page is what the model retrieved on query one. Ask whether you have any URL that answers the same job as directly.
- Repeat for the four or five prompts that matter most. You now have a table, not a screenshot.
| What you observe | What it means | First move |
|---|---|---|
| Competitor named and cited; you absent on every run | You are not in the retrieved set for this job | Ship a page whose H1 is the job. Nothing else matters until you are retrievable. |
| Competitor cited but not named; you absent | Their page is in the shortlist, the entity is weak | A direct job page can enter the same shortlist. Named follows retrieval. |
| You named on some runs, competitor on all | You are on the edge of the shortlist | Strengthen the one page the model retrieves when it does name you. Do not spread the work across five. |
| You named on a brand prompt, absent on the job prompt | The model knows who you are, not what you do | Read how to get the work cited, not just the brand. |
A first-party example, with the date attached
On 31 August 2026 we ran eight unbranded category prompts through ChatGPT with search on, the kind that begin with "best" or "top". AirPulse was named on 0 of 8. On a branded prompt the same day, it was named. That is the pattern this guide is about: the model can know a brand and still leave it off a buying shortlist, because the shortlist is built from the pages retrieved for that job, not from what the model remembers. We are publishing the count with its date so it can be re-run rather than believed.
What not to do
- Do not treat the competitor as the problem. They wrote the page the buyer's question retrieves. The gap is a missing URL on your side, not a trick on theirs.
- Do not ship another "best tools" list with yourself at the top. Category prompts are already won by whoever is retrieved for the category. Job prompts are open.
- Do not expect a percentage lift. Nobody can promise one. What you can do is run the same prompt on the same engine before and after a change and report what moved.
- Do not change five pages at once. If the prompt flips, you will not know which page did it.