How to Own a Topic in AI Search.

Owning a topic in an AI assistant is not winning one question.

It is being the brand named across the whole set of questions a buyer works through before they decide.

Semrush, working with Kevin Indig, tracked 1,094 US categories in ChatGPT every month from January to June 2026, across more than 50,000 brands and 220,000 domains. A brand counted as the owner of a category only if it held the highest share of mentions, appeared in at least four of that category’s five prompts, and led the runner-up by at least five percentage points.

By that bar, 15.2% of categories had an owner. In 53.7%, no brand even reached three of the five prompts.

For a content team, the useful part is not the scoreboard. It is that most topics are still open, and the route to taking one is coverage of a question set rather than one strong page.

What owning a topic actually means

Three conditions have to hold at once: the highest share of mentions, near-total prompt coverage, and a margin.

The coverage condition is the one worth sitting with. Being named in one prompt out of five can happen by accident. Being named in four is a pattern.

The margin carries just as much weight. Owners with a real gap held first place in 90.4% of month-on-month comparisons, and brands that lost the top spot had a median lead of 1.3 percentage points against 2.9 for those that kept it.

A narrow lead is not ownership. It is a good month.

Most categories still have no owner

The study split the 1,094 categories three ways:

• 15.2% had a clear owner.
• 31.2% had an emerging leader, named in at least three prompts but short of the margin.
• 53.7% were unsettled, with no brand reaching three of five.

The counter-intuitive part is where the demand sits. Among the top half of categories by estimated AI search demand, only 11.3% had an owner, against 19.0% in the bottom half. Across the whole set, 89.3% of estimated AI search demand sits in categories with no clear owner.

Bigger topics are less owned, not more. That is the opposite of how a search results page usually feels, and it means a content pillar you wrote off as too competitive may not be.

What this study can tell you, and what it cannot

Be precise about the boundary, because this is the sort of finding people over-read.

• One assistant. ChatGPT only, so it says nothing about Gemini, Perplexity or Google’s AI answers.
• One market. US categories.
• One window. Monthly snapshots, January to June 2026.
• Five prompts per category, which samples a question set rather than covering it.
• It measures whether a brand is named in the answer text. It does not measure sentiment, recommendation quality, trust, or whether anyone bought anything.

The authors were open about the gaps. Nearly half the cited pages were hard to classify, and the data cannot explain why leadership changes hands when it does.

Treat it as a careful read of one system over six months, not a rule about how machines think.

Search strength did not predict who owns a topic

This is the finding that should change a plan.

Comparing each category owner against its runner-up, the owner had higher branded search volume in 55.7% of comparisons, higher organic traffic in 48.4%, and a higher Authority Score in 52.5%. Only branded search volume reached statistical significance, and the edge was modest.

Two of those three numbers are close enough to a coin toss to be useless as a predictor.

There is a second split in the same data. Only about one in five of the most-cited domains in a category were also the most-mentioned brand. Being the source an answer draws on and being the brand it names are different outcomes, and the one that shapes a shortlist is the name.

Step 1: Pick a topic narrow enough to be the best answer on

The instinct is to claim the biggest category you plausibly belong in. The four-of-five bar makes that a bad trade.

Coverage is the constraint. On a broad topic, five representative questions span more ground than a small team can answer properly, so you finish adequate in three places and owning none.

Ask one question of any candidate topic: could we be the single most useful answer here within a year, with the people we actually have? If the honest answer is no, the topic is too big. Cut it until the answer is yes.

Narrow is not small. It is the size at which depth is affordable.

Step 2: Map the question set, not the keyword

A category in the study is not a keyword. It is a cluster of five representative prompts covering what a buyer asks in that subject area.

So the planning unit changes. It stops being a post or a page, and becomes the set of questions someone works through before choosing.

Write yours out in the language buyers actually use:

1. The definition question. What is this and do I need it?
2. The comparison question. What are my options and how do they differ?
3. The selection question. How do I choose, and what do good and bad look like?
4. The objection question. What goes wrong, what does it cost, why do people regret it?
5. The proof question. Who does this well and what did it get them?

Five is not sacred, it is the study’s sample size. What matters is planning against the whole set, and seeing which parts of it you have never answered.

Step 3: Find the categories nobody has settled

You can read your own space by hand in an afternoon.

• Write the five questions for each candidate topic, in buyer language.
• Run each one and record which brands are named in the answer text, in order. Ignore the link list for now.
• Repeat on a different day and from a clean session, because answers move.
• Count. One brand in four or five prompts with a visible gap means settled, and you are choosing a fight. Names that change every run, with nobody reaching three, means open.
• Date every log. The study took monthly snapshots because the picture shifts month to month, and yours will too.

Personalisation, memory and region all affect what you see, so treat your own reads as directional rather than exact. It is still more than most teams know about their own category, and it is the evidence an audit and strategy engagement starts from, before anyone commits a quarter of production.

Step 4: Cover four in five, then keep covering them

Partial coverage does not accumulate the way partial keyword coverage does.

A follow-up analysis of the same dataset found that appearing in only one of a category’s five prompts was associated with a fall in mention share, and that the penalty did not clear until a brand reached at least three of five. Thin presence across ten categories is worth less than solid presence in two.

The same analysis found that expansion works outward from strength. When brands turned up in a closely related category, 44% of those appearances included a brand mention, against 25% in a distant one.

So the sequence is: hold one cluster, step to the neighbouring cluster, then the one after that. Not five at once.

What to stop publishing to pay for the depth

Depth is bought with the capacity currently spent on breadth. Nobody is getting a bigger team for this.

The cuts that usually fund it:

• Trend reactions that map to no question in the set.
• The third format experiment on a theme you were never committed to.
• The platform kept alive out of guilt rather than plan.
• One-off posts commissioned because a stakeholder asked, not because a buyer did.

A quick way to size the problem: take your last sixty posts, tag each to a question in your set, and count the ones that map to nothing. That count is your budget for depth.

What to measure when the click never happens

An assistant naming you produces no session, no referrer and no line in a standard analytics report. If your measurement only counts clicks, this work will look like it did nothing.

Measure the mention instead. A monthly log does it: one row per prompt, with the date, the brands named in order, whether you appeared, and whether you were cited. Your ownership number is the count of prompts you appeared in, out of five.

Keep the citation column separate, because the two rarely agree.

Then pair it with demand signals you can see: branded search volume, direct traffic, and how people answer the “how did you hear about us” field. Branded search was the one metric that separated owners from runners-up at all, which makes it worth watching rather than proof of anything.

What good looks like

• Two or three topics chosen, not eight, with a written reason for each.
• A question set per topic, visible to whoever briefs the work.
• Every commissioned piece tied to a question, and anything tied to nothing declined.
• A monthly log of who gets named, kept even in the months nothing moves.
• Expansion decided by adjacency to what you already hold, not by what is trending.

How NBK thinks about topic ownership

This is a prioritisation and production problem wearing a technology costume. Coverage of a question set beats brilliance on a single page, and coverage is a capacity decision made months earlier.

So NBK treats pillars as an operating structure rather than a slide: a mapped question set per pillar, a repeatable format that answers each question, and a review that judges the theme rather than the post. It is the same argument as consistency on a topic beating volume across many, now with something measurable attached.

None of this promises a model will name you. What it does is stop a team spending a year as the fourth-best answer in five categories at once.

Next step

Most brands cannot answer a simple question about their own space: which topics are settled, which are open, and which one they could genuinely own.

If your output feels busy but spread across too many themes to be the best answer anywhere, that is a constraint in the system rather than a content problem, and it is exactly what an NBK audit is built to find.