89.3% of AI Search Demand Has No Owner Yet — But the Brands That Do Own a Category Keep It Nine Times Out of Ten

Open ChatGPT and ask it the way a customer would: “which wedding photographers in Taipei would you recommend?” — or swap in your own industry. Look at which names come back.

Most people’s first reaction is “why those?” There’s usually a competitor or two you’ve never heard of on the list, and you’re not on it. More common still: ask the same thing a different way — “how do I choose a wedding photographer in Taipei,” “what should wedding photography cost” — and a different set of names comes back.

That “the list changes every time I ask” effect is the subject of this piece. It has a specific name: the category has no owner. And a large-scale tracking study published in July 2026 puts that state at 89.3% of all AI search demand.

In AI search, the unit of competition is the category, not the keyword

One conceptual shift first, or the numbers below won’t land.

In SEO, the unit of competition is the keyword. “Taipei wedding photographer” is one target, “wedding photographer recommendations” is another; you rank for them separately, measure them separately, and climb one term at a time.

In AI answers that unit stops working. Nobody types keywords into an assistant — they type a whole sentence, and they ask about the same decision five different ways. What are my options. Is A better than B. What else is there besides A. Which type suits my situation. What should I budget. To an SEO tool those are five different keywords. To the person asking, it’s one job: picking a photographer.

So what actually gets contested in AI search is the whole category: whichever angle a customer comes in from, does the assistant reliably say your name? Showing up in one of those five answers doesn’t mean you own the category — it means you got remembered once.

That distinction sounds abstract, but it determines how the research below measures anything at all. They didn’t measure rankings. They measured whether a brand can show up across five different ways of asking.

Where the data comes from

From January to June 2026, Semrush tracked ChatGPT’s answers across 1,094 US categories, month by month. Each category got five fixed buyer-style prompts, re-asked every month. The sample covers 50,000+ brands, 220,000+ domains and 600,000+ citations. Kevin Indig got access to the dataset and published his analysis on 20 July.

Two design choices matter. Monthly re-testing means the study sees not only who’s on top now, but whether the person on top changes — which is where the single most important number comes from. Five prompts per category cross-checks each brand from five angles, so “got lucky in one answer” doesn’t get mistaken for “owns this space.”

The bar for “owning” a category is higher than you’d guess

Their definition of a category owner requires three conditions at once:

1. The highest share of mentions (share of mentions = of all the brand names appearing across the answers, what fraction are yours)
2. Appearing in at least four of the five prompts
3. A lead of five percentage points or more over the runner-up

Miss one and you don’t count. Show up in three of five prompts? Not an owner. Show up in four but lead by only two points? Also not an owner.

The bar is strict because what they’re trying to capture is being reliably recalled, not having once been mentioned. It’s also what gives the 53.7% below its weight.

By that standard, the 1,094 categories split three ways:

Read that last row slowly. It doesn’t mean no brands get mentioned in those categories. It means that across five consecutive prompts, no single brand shows up dependably. The little experiment you ran at the top — different phrasing, different names — is most likely this row.

The bigger the category, the emptier the seat

Here’s the most counterintuitive part of the study.

They ranked all 1,094 categories by AI search volume and split them into two halves of 547. The top half absorbs 98% of demand in the sample (unsurprising — demand concentrates in big categories). Then they checked ownership rates in each half:

• Big-category half: 11.3% have an owner
• Small-category half: 19% have an owner

The bigger the category, the lower the chance anyone owns it.

That’s the reverse of the SEO experience, where head terms are the most fiercely contested and get locked up by large brands first, leaving long-tail terms as the openings for smaller companies. In AI answers the direction flips.

The plausible explanation is the width of the question surface. The five prompts under a big category can span five genuinely different buyer situations — someone asking “what are my options” needs very different content from someone asking “which type suits my situation.” No brand’s content covers all five well enough, so each prompt has its own winner and nobody assembles four. In a narrow category the five prompts sit close together, and one company that explains itself clearly takes the whole set.

The 89.3% headline comes from weighting ownership by demand: the big half holds 98% of demand and is almost entirely unowned, so in aggregate, 89.3% of AI search demand sits in categories with no clear owner.

What that means for you: the big category you’ve been writing off as “too competitive, not my league” is precisely the one most likely to still be empty.

But the door swings one way: 90.4%

Up to here this reads as good news, and good news gets read as “big opportunity, no rush.”

The second number closes that reading off: brands that already own a category held first place in 90.4% of month-over-month comparisons.

Six months, re-tested every month, and owners almost never change. Nine out of ten chances to unseat someone went nowhere.

Put the two numbers together and the shape becomes clear: the seats are mostly empty, and whoever sits down is very hard to remove. Empty and hard-to-take aren’t contradictory — they describe two stages of the same thing. Nobody’s claimed it yet; once claimed, it sets.

Which is why time runs one way here. Skip it this year and the category might still be open next year, or it might not. And if it isn’t, that’s not something a bigger budget buys back — the next section explains why money does so little in this particular game.

Three percentage points is the survival line

The distance between holding a category and losing it is narrower than it sounds.

The team compared two groups: categories where the top brand changed, and categories where it held. Then they looked at the median lead in each group.

• Categories where the leader changed: median lead of 1.3 percentage points
• Categories where the leader held: median lead of 2.9 percentage points

The whole difference sits in that one-to-three-point band. Three points is roughly the line: above it you’re reasonably seated; below it you haven’t won, you just haven’t been pushed off yet.

This has a direct practical consequence. If you measure AI visibility as “do I appear in the answer,” then a one-point lead and a five-point lead look identical to you — you appear in both. Then one month you drop off the list, you go looking for a cause, and you find nothing, because you never measured the thing that decides it: how many points separate you from the runner-up.

The bad news: your SEO numbers don’t predict who wins the category

The most uncomfortable part of the study asks a very practical question: can existing SEO metrics predict who ends up owning a category?

The method was to pair every owner with its runner-up and check whether the owner was actually stronger on each metric:

(Authority Score is Semrush’s site-strength score computed from backlink and related signals — comparable to the DR figure most people know.)

All three sit near 50%. Organic traffic actually lands below it — in this sample, the side with more organic traffic was marginally less likely to own the category.

Indig’s own reading: brand strength may help a domain get into the answer set at all, but who takes the category is mostly decided by topic-specific factors these metrics never captured.

This section matters because it dismantles a comfortable assumption: “our SEO is decent, so we’re probably fine in AI too.” The data says those two things are statistically almost unrelated. Ranking first on Google and getting named by ChatGPT run on two different mechanisms.

It also explains why spending more does so little here. Money buys traffic and links — and the table above already showed what those correlate with.

To get a first read on your own gap, run our ten-minute self-test, then come back for the five-prompt version below.

The most-cited site is the most-mentioned brand only 21% of the time

One more, for anyone investing seriously in content.

Within a given category, the most-cited domain was also the most-mentioned brand only 21% of the time. The correlation is −0.229 — the minus sign means a weak inverse relationship (correlation runs from −1 to +1, where 0 means unrelated; −0.229 sits close to 0, so: nearly unrelated, and leaning slightly negative).

In plain terms: being used as source material and being named as a recommendation are two different outcomes. And doing more of the first may very slightly work against the second — you become “the site that supplies the facts” rather than “the brand worth recommending.” If you’ve published mountains of industry data that AI quotes constantly, yet you’re absent when someone asks “who should I hire,” this is the shape of that problem.

Being crawled, being cited, being written into the body of the answer, and being recommended as an option are four separate gates. We took apart the mechanics of each in our breakdown of how five AI platforms pick sources.

Practically, this changes your measurement sheet: make brand mentions the primary metric and treat citation counts as secondary. Plenty of AI visibility tools default to the opposite, because citations are easy to capture (there’s a URL, you can count it) while mentions require semantic reading (did that sentence actually refer to you, and in what tone). Easy to measure isn’t the same as worth measuring.

What this data cannot tell you

Worth stating plainly, so you don’t lean on it for decisions it won’t hold.

US categories, ChatGPT only. Six months and 1,094 categories sounds vast, but it’s the US market, English prompts, a dense field of mature brands. And ChatGPT is one engine among several, each with its own logic for picking sources.

It only records whether a brand appears. It can’t tell you whether the mention was flattering, how strongly you were recommended, or whether any of it converted. Appearing on a list described as “cheap but risky” counts the same as being enthusiastically recommended.

Close to half the cited pages failed classification, so questions like “what kind of content gets cited most” are out of this dataset’s reach.

It is correlational from end to end. Nothing in it demonstrates that doing X wins you a category. Anyone citing this study to guarantee that their method will make you the owner is selling something.

For markets outside the US, discount it further. Chinese-language markets are tracked and contested far less densely than this sample. Directionally the seats are emptier still — most Chinese-language categories probably have nobody near that 15.2% bar — but you can’t lift 11.3% or 89.3% and call it your local reality. To know whether your category has an owner today, you have to go ask.

A self-test you can run tonight

The good news is that the study’s method isn’t complicated, and you can run a scaled-down version yourself.

Step one: write down your categories. Not keywords — the way a customer would classify you. Say you run an interior design studio specialising in old-apartment renovations; your categories are things like “old apartment renovation design” and “pre-owned home remodelling,” not “interior design,” which is too broad and isn’t how your customers look for you. Three to five is plenty.

Step two: ask five prompts per category. The five map to five buyer situations. Don’t skip any:

1. Definition: “What does old-apartment renovation design involve?” (See how the assistant frames the category — and which names it reaches for as examples)
2. Comparison: “Is A or B better for renovating an old apartment?” (Put yourself and your biggest competitor in the prompt)
3. Alternatives: “Besides A, who else does old-apartment renovation in Taipei?” (See whether you make the “other options” list)
4. Use case: “A thirty-year-old apartment needs full rewiring — what kind of design team should I hire?” (See who gets recommended in a concrete situation)
5. Buying question: “How are renovation design fees calculated, and how do I avoid getting burned?” (See who gets treated as the trustworthy reference)

Step three: record three things — how many of the five prompts you appeared in, which names appeared in each, and how far you sit behind the most-mentioned brand. You don’t need decimal precision, but you need a feel for it: is it “they’re in every answer and I’m in one,” or “we’re both in all five and they’re ahead by one mention”?

Step four: read the result. Four or more appearances with a clear lead means you’re in that 15.2% and your job now is defending it. Two or three appearances, or lots of appearances but neck-and-neck with the runner-up, is the 1.3-point median from the study — contested, and liable to flip. Fewer than two out of five means that category currently doesn’t have you in it at all.

Step five: find the gaps. Pull out every prompt where a competitor appears and you don’t — that’s your content gap. Absences on the comparison and alternatives prompts usually trace back to a page that simply doesn’t exist: nobody has written up how you differ from X, so the assistant can only talk about X. How to structure those two page types is covered in our piece on comparison-intent keywords.

One evening of work gets you a picture of your position more honest than any tool dashboard.

Running it once is easy. Running it every month is the job.

Then you hit the real problem.

The value isn’t in one run, it’s in the trend. A single snapshot tells you today’s list but not whether you’re being squeezed out or working your way in — and the three-point survival line only becomes visible across consecutive months. To mean anything, it needs the same categories and the same five prompts, re-run every month.

It also can’t stop at ChatGPT. The study tested one engine; your customers are spread across Gemini, Perplexity, Claude and every AI search surface in between, each with its own source logic. The same category can have a different occupant on each one.

And the heaviest part: once you find the gaps, somebody has to close them. What’s missing is rarely an article or two — it’s binding your brand name to the category term consistently, in enough places: your own site says it clearly, and third-party discussion, reviews and roundup pieces point at the same thing. That’s an ongoing job, not an afternoon project.

Twenty categories × five prompts × every month × several engines, plus the content work behind it. Laid out like that, it stops looking like something anyone finishes after hours.

And right now the return on that work is probably the highest it will ever be: the seats are still open, and once you’re in one, there’s a 90.4% chance nobody takes it from you. No one can give you a closing date, which is exactly what makes it uncomfortable — no notice arrives; you just ask ChatGPT one day and find the seat taken.

It could have been yours.

That’s the work we take on. See how GEO managed service works →