Why ChatGPT recommends some SaaS companies and ignores others

A founder asks ChatGPT the kind of question their buyers ask. Something like "best tool for monitoring LLM output in production". Three companies come back. One of them is a competitor. Their own company is not there.

This is happening to my own company too. I run OwnedSignal, and every week OwnedSignal checks a fixed set of buyer questions in ChatGPT, Gemini and Perplexity. For the question "Who helps technical SaaS founders get found and trusted before sales calls?", ChatGPT named other firms and not OwnedSignal in each of the four runs logged between 16 and 26 October. So I am writing this from the same position as the founder above, with the difference that I keep a log.

You cannot fix this from one screenshot. First you need to know what kind of result you are looking at.

Four results that all get called "visibility"

These can come apart. You can be mentioned without being cited, which means the assistant knows your name but is not sending anyone to your site. You can be cited without being recommended: your page helped write the answer, and the answer either recommended nobody or recommended someone else.

Most AI visibility reports count mentions, because mentions are easy to count. If you are paying for a report, ask which of the four it measures.

What happened when OwnedSignal was tested

Here is what the assistants said about OwnedSignal in October, according to OwnedSignal's test log, and what I found when I went looking for the reason.

The wrong company. On 7 October, a logged-out check asked ChatGPT "What does OwnedSignal do?". It answered with a class from a Scala.js programming library that has the same name. On 14 October the same question returned my company first. I would like to say my site changes fixed that, but the 14 October check ran in the morning and I changed the site in the afternoon. So I can't credit the edits. All I know is that the answer changed.

The old company. On 21 October, Perplexity answered the same question by saying OwnedSignal "helps bootstrapped B2B SaaS founders", which was my old positioning. The site had dropped that wording on 14 October. Perplexity cited an old page of mine that already redirected, so the old description was still somewhere it could read it.

The missing service. This one is the clearest. On 24 October an OwnedSignal test asked ChatGPT whether OwnedSignal offers actual video production. It said OwnedSignal "does not offer full video production as a standard service". That was wrong: founder video is one of the services I sell and do myself, from research and interviewing to recording and editing.

Then I went through my own site page by page, and ChatGPT was repeating what my site said. The Sprint and 90-day pages described video production as excluded from their scope. There was no page that sold video. The assistant read my pages and described them accurately. My pages were out of date.

So I changed the site. Founder Video Production now has its own service page. Every line that said video was excluded is rewritten. The Sprint has a video outcome, and the 90-day engagements say what video they include.

I don't know yet if this changes the answer. When the new pages are live and indexed, OwnedSignal's checks will ask the same question again, logged out, from a fixed location, on three dates a week apart, and count how many answers say video is included. The next step is to repeat the same question after the website changes and compare the answers against the original result.

If an assistant describes your company wrongly, read your own site the way it would, before you blame the model. Look for the sentence it might be repeating.

Why a competitor gets recommended: what the research shows

None of the AI companies publish how they choose. What we have are outside studies, and most of them are associations: they show what tends to appear together, not what causes what.

Different engines, different answers. In a 2026 study of 24,067 prompts across Google AI Overviews, ChatGPT, Gemini, Claude and Perplexity, Minuttia and Convex Works found that 54% of brands recommended on "alternatives" searches appeared in only one engine. Only 8% were named by all five. Being recommended in ChatGPT tells you very little about Gemini.

Named is not the same as linked. The same study found that only 3% to 29% of brands named on alternatives searches came with a citation to their own website, depending on the engine.

Google rankings go together with AI mentions. Brands on Google's first page were 6.8 to 11.6 times more likely to be named. This is a correlation. It does not show that ranking higher on Google will make ChatGPT recommend you. Brands that rank well are often the brands more people write about, so both may come from the same thing.

Your own site matters more for some questions. In a test of 20,000 US prompts, reported in Ahrefs' study of how ChatGPT searches, 63.4% of citations for factual questions went to official websites. For comparison questions, 46.4% did. So when a buyer asks "what does X do", your site carries a lot of weight. When they ask "X or Y", third-party pages carry more.

Narrow descriptions may help. BrightEdge reported in September 2026 that ChatGPT named 3.87 B2B tech brands per answer on average, down from 5.36 in March, and that focused, single-purpose tools gained share. BrightEdge sells AI search tracking and measured mentions, so read it as a direction, not a rule.

Put together, a competitor that ranks for the question, is described clearly for one use, and appears in third-party comparisons is more likely to be named. But "more likely" is all the research supports.

How to check why a competitor shows up and you don't

This is the method OwnedSignal's own checks follow.

1. Ask a real buyer question without your brand name in it. Use the words a buyer would type, from your sales calls if you can. "What does [your company] do?" tests something else, so keep it as a separate control question.
2. Fix the conditions and write them down. Fresh private window, logged out, one question per new chat, no follow-ups. Note the date, time, model label if shown, and your location. On 26 October the test machine was in Seattle, and ChatGPT named Seattle providers on two commercial questions. Location can change the answer.
3. Record all four results for every company named. Mentioned, cited, listed, recommended. Copy the first sentence that describes each company.
4. Open the sources. For each competitor that was recommended, which pages did the answer cite? Their own site, a review site, a roundup, a forum thread? This shows where the assistant learned what it said.
5. Compare descriptions. Put the competitor's homepage line next to yours. Does theirs name the buyer's exact situation and yours talk about something broader? Does a third-party page describe them for that use and nobody describes you?
6. Look for wrong or missing commercial information on your site. Services listed as excluded, old positioning, no page for the thing the buyer asked about. In my case this was the whole problem.
7. Repeat before you decide anything. Run the same questions again on at least two more dates. One run is a sample, not a result.

Only after that does it make sense to talk about causes. If you are not mentioned at all, possible reasons include crawl access, no page that answers the question, or a description that doesn't match the question. If you are mentioned but never recommended, look first at the third-party pages cited on comparison questions. These are possible explanations to check, not a diagnosis.

What one test can't tell you

Parse measured 161,023 back-to-back answers to identical prompts and found the top recommendation changed 43.6% of the time. OwnedSignal's logs show the same kind of swing on a small scale. On 19 October, Gemini cited ownedsignal.com on 2 of 12 questions. On 26 October it cited it on 0 of 11. The site was updated in between, so I can't separate the two, but it shows how much one sample can mislead.

If I had looked only on 19 October, I would have told you Gemini found me. If I had looked only on 26 October, I would have told you it didn't. Both would be wrong.

Does any of this bring customers?

The honest answer is that the evidence is mixed, because different methods measure different things. Stanislav Peev's review of attribution data found AI under 2% of tracked traffic and leads in large analytics datasets, and 7% to 25% when companies ask new customers directly how they heard about them.

Both can be true. Analytics can't see a buyer who reads an AI answer and then types your name into Google. If you want to know, ask every new lead how they found you, and write down their exact words.

Where OwnedSignal fits

This is the work in my AI Search Visibility service: the buyer questions, the repeated checks across ChatGPT, Perplexity, Gemini and Google, the pages that answer those questions, and a dated record of what changed. I don't guarantee recommendations, because nobody outside these companies can. What you get is the testing, the fixes, and an honest record of what moved and what didn't.