SIGNALS
Buying AI visibility

How do you tell whether your category is winnable in AI search?

The short version

Share of voice tells you where you stand. Winnability is a different reading and it takes three checks. First, count the names: a category where the same two or three firms are returned on every phrasing is concentrated, and citation concentration is measured: Qi Liu and colleagues found the top 20 domains taking 20.5% to 70.8% of citations per platform across 10 AI search platforms (arXiv 2610.11932, 2026). Second, read the citation list next to the recommendation. If the engine names firms whose pages it does not cite, the names are coming from brand memory rather than from pages, and page work will not reach them. Third, look at what kind of page the engine retrieved. A retrieval set made of third-party comparison pages is one you can get into; a set made of three incumbent vendor sites is not, at least not this quarter.

How do you tell whether your category is winnable in AI search?

To tell whether your category is winnable in AI search, run your buying questions repeatedly and read three things in the answers rather than one: how many distinct firms get named across the set, whether the firms named are the same as the sources cited, and what kind of page the citations point at. Those three readings separate a category you can enter with page work from one where the answer is being assembled out of reach of anything on your own site.

Most competitive checks in this category stop at the first reading. That reading is share of voice: how often you are named against how often your rivals are. Share of voice is worth knowing and it is not a winnability judgement. It tells you the current standing and says nothing about whether the standing can be changed. A buying decision turns on the second question, not the first.

The honest version of the exercise also admits what cannot be read from outside. You can see which pages were cited. You cannot see why one was chosen over another, and no tool can tell you that, because the ranking happens inside a retrieval step you do not observe. What a careful reading gives you is the shape of the problem and therefore the kind of work that could move it. That is enough to decide whether to spend money and on what.

The sections below work through each reading, what the published evidence says about how concentrated these answers are, and a check you can run yourself in an afternoon.

What does a crowded category look like in an AI answer?

A crowded category shows up as a short, stable list of names and a short, stable list of cited domains, repeated across every phrasing of the question you can think of. The stability is the signal rather than the length: a category still in motion returns different firms when you reword the prompt, and a settled one does not.

Citation concentration has been measured. A study mapping citations across 10 AI search platforms covered 17,211 citation instances across 6,356 unique source domains and found that the top 20 domains captured 20.5% to 70.8% of citations on a given platform, according to Qi Liu and colleagues in "From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search" (arXiv 2610.11932, submitted 8 October 2026). The spread across platforms is the useful part. On some engines a handful of domains supply most of the answer and on others the field is far wider, which means a category can be closed on one engine and open on another.

That study also found that 15 of 22 publication platforms tied to cited source domains had low or medium barriers to both account creation and posting, and that one article on a platform the engines prefer carried more citation impact than more than 20 matched posts elsewhere. Read that as a map of where retrieval sets are actually built, not as a tactic: the same paper's methods are now prohibited by platform policy, which the next section but one covers.

Practically, what you are looking for in your own category is the ratio of distinct firms named to prompts run. Twenty prompts returning five firms is a concentrated category. Twenty prompts returning eighteen firms is an unsettled one, and an unsettled one is where a well-built page gets into the set. Our guide on tracking AI visibility across the four engines covers how to run the set so the counts mean something.

Why does the shortlist not change when you fix your own pages?

The shortlist often does not change when you fix your own pages because the page is the smaller of the two levers, and because the names in a recommendation do not all come from pages at all. The best-identified measurement of this split found that which pages the engine's own search retrieves explains about 32% of the variation in citations while page properties explain 2.8%, in a study of 158,099 answers and 468,037 citations by Mahe Chen and Nick Koudas, "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization" (arXiv 2610.15033, submitted 10 October 2026).

The second half of that study is the part that decides winnability. The authors report that ChatGPT includes brands it already knows in its searches, favours sites it recognises, and names brands with no supporting page at all, and that in controlled tests changing only the brand name moved product rankings although the underlying facts were identical. Their summary sentence is that pages are cited when searched and brands are also named when remembered.

That gives you a diagnostic you can run from the outside. Put the recommendation beside the citation list in the same answer. Where the two agree, the answer was assembled from pages, and page work is the lever. Where a firm is named and nothing of theirs is cited, that name is coming from model memory, and nothing you publish on your own domain reaches it directly. A category where the leaders are consistently named without being cited is one where the winnable move is off-site presence rather than on-site structure.

Both halves of that study are worth saying plainly, because 2.8% is a small number next to what this whole category sells. The single page-level lever it finds that survives within a site is the match between the page's language and the question actually asked, which is the thing our own methodology weights highest. Page work is how you become retrievable. Being recommended is a brand outcome.

Does publishing your own best-of page win the category?

Publishing your own best-of page does not reliably win the category, and there is a measurement showing it can hand the recommendation to somebody else. Lily Ray analysed 100 B2B "best [category]" queries in Google AI Overviews at three checkpoints, 15 April, 15 May and 8 June 2026, with 80 of the 100 prompts triggering an AI Overview. Where a brand's own self-promotional listicle was cited, the brand was left out of the recommendation in 224 of 323 cases, 69%, and in 74 of the 100 prompts an AI Overview cited a self-promoter's listicle without recommending that brand, in "Why Calling Yourself the Best Could Be Helping Your Competitors Win in AI Search" (17 June 2026). Ray's own summary is that for many sites, self-promoting listicles might actually be more of a liability than an asset.

Read carefully, that is not an argument against comparison pages. It is an argument about what they are for. The page can win the citation, meaning coverage of the query, while the recommendation goes to a firm named inside it. So a comparison page earns its place as a genuinely useful comparison and as a route into the retrieval set, and not as a route to being the recommended name.

We publish two comparison pages of this shape ourselves, and SIGNALS appears in the table on both of them, so read our own entries the way you would read any vendor's account of itself. Both halves of that disclosure matter for the finding above: it applies to our pages as much as to anybody's, and it is the reason we would rather tell you what a comparison page can and cannot do than sell it as a shortcut.

The more useful read for winnability is which listicles the engine retrieved for your category. If the cited pages are third-party roundups, there is a route in: being considered for inclusion is ordinary outreach. If they are competitor-published pages that rank themselves first, the route is narrower and slower. Our guides on getting into the best-of lists AI engines quote and whether to write comparison pages yourself cover each path.

Which categories are winnable, and which are not?

Categories sort into four rough shapes, and the shape tells you which work would move it. The table below is the read we use, with the diagnostic for each and the honest prognosis.

What the answers look like What it means What moves it
Many firms named, citations mostly third-party pages Unsettled category, retrieval set still forming Page work plus inclusion in the roundups already cited. The winnable case
Few firms named, citations match the named firms Concentrated but page-driven Page work on the specific questions, starting with narrow qualified ones
Few firms named, their pages not cited Brand memory is doing the work Off-site presence and third-party coverage. Page work alone will not reach it
Answers vary wildly between runs, few citations at all Thin retrieval set, nobody has won it Publishing the page that answers the question. Cheapest opportunity on the list

The fourth row is the one buyers most often miss, because it looks like failure. A question that returns a vague answer with no citations is not evidence that the question does not matter; it is evidence that nothing good exists to answer it. Our page on whether AEO works if nobody searches your category yet covers that case in full.

The head of a category is almost always the third row. Broad terms are answered out of names the model already holds, which is why publishing one more explainer against them changes nothing. A qualified version of the same question, narrowed by industry, use case or constraint, often sits in the first or fourth row instead.

How do you run the winnability check yourself?

Run the winnability check in four steps, and write the results down so the next read has something to compare against. The whole thing takes an afternoon and needs no tooling beyond the engines themselves.

First, write the questions a buyer would actually type, including the follow-ups, and include the narrow qualified versions rather than only the category term. Second, run every question on at least two engines and run each one more than once, because the same question asked twice can name different firms. Third, for each answer record two lists separately: the firms named, and the domains cited. Fourth, classify each cited domain as your site, a competitor's site, a third-party roundup, a review site, a forum or a publication.

The classification in step four is where the answer comes from. A retrieval set dominated by third-party pages means the route in is getting into those pages. A set dominated by competitor sites means the engine is reading vendor material directly. That is the mode a serious buyer's deeper research uses, and depth on your own domain is what competes there. A set with few citations at all means nobody has answered the question yet.

One live example of reading the retrieval set rather than the answer, from our own category. Searching "best AI visibility tools 2026" on 17 November 2026 returned a first page made almost entirely of roundups published by firms that sell into the category, among them Conductor, NoGood, Clearscope and Trakkr, and the ones published by a platform tend to rank that platform first. For a new entrant that is a first-row category on the table above: the pages that get retrieved are lists, and lists can be joined. For an incumbent it is a warning that the query is already occupied by whoever publishes the most lists.

If you would rather have the baseline run for you, a free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude and reports the named firms and the cited sources separately, which is the distinction this whole page turns on.

What else do people ask about category competition in AI search?

Marketing leads sizing up category competition in AI search ask these five questions most often.

How many prompts do you need to judge whether a category is winnable?

Enough to cover the buying question and its follow-ups, run more than once, which in practice means a few dozen rather than a handful. One run of one prompt tells you almost nothing, because these systems sample rather than rank once and hold, so the same question asked twice in an hour can name different firms. Judge the set, not the answer.

Does a strong Google ranking mean the category is winnable in AI search?

No, and the two can point in opposite directions. Ranking first on a category term does not put you in the retrieval set an assistant builds for a buying question, because the engine rewrites the question into sub-queries and retrieves against those. A page can own the blue link and never appear in the answer.

Is a narrow category easier to win than a broad one?

Usually, and for a structural reason rather than a volume one. A broad category question is answered from pages with years of domain authority behind them, while a narrow, qualified question often has no good page at all, so the retrieval set is thin and a well-built page can enter it. The trade is that a narrow question is asked by fewer people.

What is the clearest sign that a category is not winnable with page work alone?

The engine names firms whose pages it does not cite. When the recommendation and the citation list disagree, the names are coming from somewhere other than the pages in front of it, and no amount of rewriting your own site reaches that. Off-site presence is the lever in that case, and it is a different kind of work with a different cost.

How often should you re-run the check?

Quarterly for a category read, and monthly if you have work in flight. Retrieval sets turn over fast enough that a six-month-old competitive picture describes a market that has moved, and the whole point of the exercise is to catch the moment a thin retrieval set gets occupied by somebody who noticed before you did.

Related guides

The Assessment

See who your category actually names, and what it cites.

A free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude, and reports the firms named and the sources cited as two separate lists.

Request a free visibility assessment →
SIGNALS · A BlackSig Systems company