The Competitors AI Names Instead of You
The useful data in an AI answer is not whether you are in it. It is who is there when you are not, and that list rarely matches your market map.
Most people run their first AI visibility check hunting for their own name. They skim the answer, find it or fail to find it, and close the tab. That is the least informative thing on the screen. The interesting part is the rest of the sentence: the companies recommended while you were not.
That list is a competitive map, and you did not draw it. It was assembled by a system that has never seen your pipeline, your price list or the deals you lost last quarter. It knows only how the web talks about your category, and what the web says is usually not what your market map says.
Where a model gets its idea of your competitors
A language model does not store a tidy directory of markets with a vendor list under each one. It stores statistical associations between words, and it retrieves pages at answer time. When a question names a category, the brands that come out are the ones that appear near that category most often in the text it learned from, or in the pages it just fetched.
So the grouping logic is textual, not commercial. Brands land in the same bucket because they get named on the same pages: the same roundups, the same comparison articles, the same forum threads, the same arguments on Reddit. Co-mention is the whole qualification. Nobody checked whether you and the company next to you actually chase the same buyer.
That inverts the usual assumption. You treat your competitive set as a fact about the world; the model treats it as a fact about the corpus.
What that produces, and why it looks wrong at first
Companies you have never considered competitors show up
Often smaller ones. Sometimes from a category next door, solving an adjacent problem for a slightly different buyer. They appear because they have a dense textual footprint in the places that matter: real documentation, an active community, a founder who answers questions in public, and above all a habit of being written about by others. The model reads that density as relevance.
They are not necessarily winning the market. They are winning the paragraph. Those are different achievements, and only one of them is visible from inside your company.
Real competitors go missing
The mirror image is stranger. A large, established rival with a strong sales organisation, a permanent stand at the industry trade show and instant recognition among buyers can simply be absent from the answer. If the open web says little about them in the context of the question being asked, there is nothing for the model to pull on.
This hits enterprise vendors hardest. Their knowledge lives in gated PDFs, sales decks, partner portals and meeting rooms: places the model was never in. Absence from an AI answer says nothing about the strength of a business. It says something about where that business is discussed.
Your competitor changes when the language changes
Ask the buying question in English and you get one set of names. Ask it in German or Spanish and you can get a substantially different set, with local players occupying the seats. The German-language web discusses German suppliers, cites German trade publications, argues in German forums, and the model answers from what that language contains.
The consequence for anyone who exports: your rival in an AI answer is market-specific, and it is frequently a company your sales team has never heard of. Any AI competitor monitoring that runs only in your home language is describing one market and quietly inviting you to read it as the description of all of them.
How to read the list without telling yourself a story
These models are probabilistic. The same question, asked twice, can return two different lineups. So the unit of analysis is never a single answer.
Count frequency, not appearances. Run the question many times, across platforms, and keep a tally. A name that comes back in most runs is a structural association: the model reliably links that brand to that question. A name that showed up once is noise, and rewriting a positioning deck because of a single sample is a good way to chase a random number.
Then look at order. Being named first is not the same as being named last. The first name in a generated answer tends to occupy the position of the default, the safe choice, the thing you mention before you qualify anything.
Then look at characterisation. This is the part almost nobody records, and it carries more information than the mention itself. Read the clause attached to each name. "The industry standard" is one thing. "A cheaper alternative for small teams" is another. Being named third and framed as the budget choice is a completely different result from being named first and framed as the standard, and any measurement that treats both as a mention has thrown the finding away. That includes a visibility score, ours included, if you look only at the number and never at the sentences underneath it.
The only question worth asking about the list
Not "how do I get rid of them". The useful question is: why does the model associate them with this question and not me?
The answer is almost always in the sources. Perplexity shows its citations inline. ChatGPT with browsing enabled links what it fetched. Open those pages, one by one, and you will find the same shape again and again: third-party pages where your competitor is present and you are not. A roundup on a trade publication. A comparison written by someone nobody briefed. A forum answer from a practitioner. A documentation page from a company that integrates with them.
One caution. Do not ask the model why it picked them. It will produce a fluent, confident reason, and that reason is a reconstruction after the fact, not a report from inside the machine. The retrieved sources are real and checkable. The stated rationale is prose.
That reframes the work. The gap is not on your website, where you have complete control and have already spent years. It is on pages you do not own, in conversations about your category happening without you.
What you cannot do about it
You cannot move a model by attacking a competitor. There is no disavow file for language models and no form where you report a recommendation as unfair. There is no positive lever either: no submission endpoint, no ranking factor to tune, no partner tier that buys inclusion. Anyone offering to remove a rival from AI answers, or to swap your name into their slot, is selling something they cannot deliver, and you should assume the rest of what they told you is the same quality.
What is available is knowing: which companies get named in your place, in which language, in which market, how often, in what order and described how. PSentry reports exactly that, per language and per market, across ChatGPT, Claude, Gemini and Perplexity. It counts who is named and who is not. Rewriting the paragraph, silencing a competitor, or guaranteeing that the list looks different next time are not services on offer, from us or anyone.
Whether the list ever changes is a business question, not a software one. It depends on whether your category starts talking about you in public, in the languages you sell in, on pages you do not control. That moves slowly. And you cannot even attempt it while you still believe your competitors are the ones on your slide.
The honest summary
The most valuable line in an AI answer is the one with someone else's name in it. It is a free and unflattering read on how your market is actually described in text. Read the names, count them properly, note how they are framed, and go find the pages that put them there. That is a real to-do list, and it is longer than the one you had.
Frequently Asked Questions
Why does an AI recommend a company smaller than mine?
Because size is not visible to a language model. Textual presence is. A smaller company with thorough documentation, an active community and a habit of being written about by third parties has a denser footprint in the exact text the model draws on than a larger company whose knowledge lives in sales decks.
How many times do I have to ask before the list means something?
More than once, and more than once per platform. There is no magic threshold, but the principle is simple: treat each answer as a sample, not as a verdict, and only trust names that recur across many runs. The single answer that upset you is probably noise.
Can I get a competitor removed from an AI answer?
No. There is no mechanism for it and no vendor who can perform it. The names in an answer are a byproduct of what the web says, and the only thing that shifts them is the web saying something different, which takes time and is nobody's deliverable.
Should I run this in every language I sell in?
If you export, yes, and that is where most of the surprise lives. The competitive set an AI names in a destination market is frequently made of local players, and that answer stays invisible if you only ever ask in your own language.
What should I record each time I check?
The platform, the language, the exact prompt, every brand named, the order they appear in, the clause used to describe each one, and the sources cited if any. The brand list alone is not enough: the ordering and the framing are where the meaning is.