Competitors
Who gets recommended instead?
You can track it, and the ranking is worth having. What decides whether it is any good is where the list of rivals came from in the first place.
The short answer
Yes. Ask the same prompt 10 times, write down every brand the assistant names in each answer and the order it named them in, and you have a ranking of who gets recommended most often for that question. Repeat it per prompt and per assistant. That count, not a single score, is the thing you are after.
The catch is upstream of all of it. Most competitive tracking is built on share of voice, which is defined as your mentions divided by the mentions across you and the rivals you selected. So the ranking only ever contains names you already typed in. A company taking your answers that you have never heard of scores zero and reads as an absence rather than as the thing beating you.
A competitor leaderboard built from a list you supplied cannot surprise you, and the competitor worth finding is exactly the one you would not have thought to list.
Let the list build itself
The fix is not a longer list. It is recording every brand named on every tracked prompt and ranking whatever comes back, with no seed list at all. Do that and the rivals arrive on their own, which regularly turns up three kinds of name a person would not have entered:
- A company from an adjacent category that solves the same problem differently, which your competitive analysis never covered because it is not in your market on paper.
- A directory, a review site or a forum thread. Not a rival at all, and frequently the thing actually occupying the answer, which changes the job from outranking a competitor to getting listed on a page.
- A product too new or too small to be on anyone's list, which the model picked up from a roundup published last quarter.
The same exercise sometimes returns the more useful negative result: the rival you have been benchmarking against is not being named either, and the two of you have been losing the same answers to somebody else. Beseen builds the list this way, which is what the FAQ means when it says competitors arrive on their own.
Most often is a per-prompt fact
One leaderboard averaged across every prompt hides the structure that tells you what to do. In practice one rival owns the enterprise phrasing of a question, another owns the cheap-and-fast phrasing, and a third is named on almost everything and is always last. Averaged together those three look like one ordered list of competitors, and none of the three facts survives.
So keep the rankings apart. There are four of them and they answer different questions.
| The ranking | What it ranks | What it is for |
|---|---|---|
| Named most often, averaged over every prompt | Who owns the category in general. | A slide. It is the ranking least likely to tell you what to change, because the prompt you lose worst is averaged in with the ones you win. |
| Named most often, on one prompt | Who owns that specific question. | Deciding what to write. This is the unit of work: one prompt, one rival, one page that beat you. |
| Cited most often: whose pages the answer was built from | Which pages the assistant actually read. | Knowing what to change. A named rival whose own site is never cited cannot be beaten by copying their site. |
| Named first against named last | Who is the default pick and who is the afterthought. | Reading progress before the counts move. Going from unnamed to named last is real movement a yes-or-no record throws away. |
The most-named brand is often not the one whose pages won
This is the split that decides your response, and it is the one a single competitor ranking cannot show you. Being named and being cited are separate events with separate causes.
- A rival named in the sentence but with none of their own pages in the sources is being recalled, not read. The assistant knows them from somewhere else, usually a roundup or a review site it did read. Rewriting your pages to look more like theirs changes nothing, because their pages were never in the answer.
- A domain cited in the sources but never named in the sentence is supplying the explanation without getting the credit. If that is a directory or a publication, being listed on it is the shortest route into the answer.
- A rival both named and cited is the only case where reading their page tells you what to change on yours, and it is the least common of the three.
Which is why a ranking is only half a finding. The other half is the page that won, and Ranking is not being named goes through where being ranked on Google and being named by an assistant come apart, which is the same gap from the other side.
Do not average the assistants together
ChatGPT, Claude, Perplexity and Gemini do not answer the same question with the same brands, and the one that leads on one of them is often mid-table on another. A single cross-engine number reports the average of four different markets and matches none of them.
- Rank per assistant, and compare rankings across assistants rather than merging them. Where they disagree is itself the finding.
- Use the same number of runs everywhere. A rival counted over 10 runs on one engine and three on another is not comparable, and the arithmetic will not warn you.
- Throw away a round that failed to reach one of the engines rather than publishing it short. A rate is only comparable against a rate measured over the same set.
- Record the model version where it is shown. A leaderboard that reshuffles the week a model ships is telling you about the model, not about your market.
What the tools built for this do, and where they stop
Both claims below were read off the products’ own live pages on 2026-09-05, and both tools do real work worth paying for.
Citeview
Built for exactly this surface, and its metric set is the closest of the two: share of voice, citation share, and an average brand rank across the models it queries, plus splits by persona and by country that nothing here offers. If what you want is a benchmarking dashboard across many models, it is a reasonable answer.
Mention
A media monitoring platform first, with breadth nothing in this category matches and roughly two years of history to compare against. Competitor tracking there is a dashboard you configure, and AI answers are one surface among the press and social ones it was built for.
Where both leave you
With the ranking, and with the next move still to work out. Knowing a rival is named on six of your ten prompts does not tell you which of your pages to change or what to change about it, and that gap is where the week after the dashboard goes. Beseen closes the same loop it opens: every prompt you lose comes back with the brand that won it, the page cited for it, and the rewrite or the brief for the page of yours that should have been there.
What to do with the ranking once you have it
- Take the prompt you lose worst rather than the rival you dislike most. The prompt is the unit that can be won.
- Open the pages actually cited in that answer and read what they do that yours does not. Usually it is structural: the answer in the first paragraph, a comparison table, the brand name sitting next to the claim.
- Change one thing, and note the date. A ranking that moved after you changed four things tells you nothing you can repeat.
- Re-run the same prompt, same wording, same count, three to four weeks later, and compare who was named then against who is named now.
That last step is a measurement with its own failure modes, and measuring whether a change moved an answer covers which numbers can settle it and what has to stay frozen between the two rounds. Proving a change moved an AI answer has the run counts and the test behind them.
To see the ranking for your own prompts rather than build it by hand, how it works describes what a run records, and the first two checks are free when you start: the competitor list lands on the first one, because it is built from the answers rather than from anything you have to type in.