Citations and mention share correlate at -0.229. That is a negative number, measured across 1,094 US categories in a Semrush AI Visibility Toolkit dataset analysed by Kevin Indig for Growth Memo. The domain cited most often in a category was also the brand mentioned most often only 20.8% of the time. Reverse the question and the asymmetry sharpens: the most-mentioned brand did pick up at least one citation 69.9% of the time.
Most AI-search reporting counts citations. It is the number vendors surface, the number that fits the link-shaped mental model everyone inherited from SEO, and the number a marketing lead can drop into a board slide without explaining. Across this dataset it moves against the thing that determines which brand a buyer actually hears named.
Scope bounds everything below. ChatGPT only. USA only. January to June 2026, monthly snapshots, five prompts per category. Semrush supplied the data and sponsors the post; Indig ran the analysis. One well-instrumented window into one assistant in one market.
Two scoreboards that disagree
A citation is a URL the assistant surfaces as a source. A mention is the brand named in the prose the buyer reads. The dataset spans more than 220,000 domains, more than 50,000 brands, more than 600,000 citations, and more than 220,000 URLs, which is enough volume for the divergence to be structural rather than noise.
Homepages accounted for 4% of citations. The rest went to deep pages, documentation, comparison articles, forum threads, and third-party coverage. Nothing there is under your homepage’s control, and the licensing arrangements that shape which domains get cited at all sit further outside it.
The practical consequence lands in reporting before it reaches strategy. A team can hold the strongest mention position in its category and see a mediocre citation count, then spend a quarter optimizing the citation count. The reverse also happens: a domain with heavy citation volume can be functionally invisible in the prose, because the assistant sources from it while recommending someone else. If your dashboard has one AI-search number on it, and that number is citations, it can move the wrong way while your position improves.
Ownership sticks once it exists
Indig classified every category, every month, against three thresholds. An owner holds the highest mention share, is named in at least 4 of 5 prompts, and leads the runner-up by at least 5 percentage points. An emerging leader is most-mentioned in at least 3 prompts but sits below that bar. A category is unsettled when no brand leads in 3 or more prompts.
Clear owners retained first place in 90.4% of month-over-month comparisons. That is the finding with the longest shelf life. Leaders did change in 1,950 of 5,470 comparisons, so the field is far from frozen, and the variable separating the two groups was margin. Categories where the leader switched had a median lead of 1.3 percentage points. Categories where the leader held had a median lead of 2.9.
Indig’s reading: “Narrow-lead categories are contestable; wide, sustained ownership is much harder to displace.”
Two points of separation is the difference between a position you defend and a position that changes hands next month. If you are measuring mention share at all, that threshold is the one to put on the chart.
The demand is in the unclaimed half
In June 2026, 15.2% of categories had a clear owner. 53.7% were open fields with several contenders and nobody in front.
Now split the 1,094 categories into two halves of 547 by AI-search volume. The top half holds 98% of estimated demand. It also has the lower owner rate: 11.3%, against 19.0% in the bottom half. Big categories are harder to own, which is unsurprising, and the effect is strong enough to flip the aggregate. Net result: 89.3% of estimated AI-search demand sits in categories with no clear owner.
That number cuts two ways depending on where you sit. If you already lead a small, well-defined category, the stickiness finding says defend the margin and widen it past two points. If you are looking at a large category with several contenders and no owner, the position is available in a way it has not been in organic search for fifteen years.
Legal has zero owners
Retail had the highest owner rate at 61.6%, 61 categories out of 99. Legal was the only industry in the dataset with no clear owners at all.
Legal is also where the downstream stakes are highest per query. In an earlier Growth Memo user study on high-stakes purchases in AI Mode, 74% of participants chose the item ranked first, the mean rank of the final choice was 1.35, and only 10% chose anything at rank 3 or lower. 64% clicked nothing at all before deciding. When the assistant is the last surface a buyer reads, the first name in the answer carries most of the decision.
An entire industry where high-stakes buyers are being routed by an assistant and no brand has established a durable position is an unusual thing to find in a dataset. It is also consistent with what we have argued about AI search as a governance surface rather than a channel: the firms best positioned to be named are the ones whose published expertise is verifiable, structured, and hard to fabricate.
What this data does not establish
Indig is careful here and the caution is worth copying. The analysis identifies traits associated with ownership. It does not show how ownership was created. Brands that own categories tend to have stronger branded demand and stronger organic performance, and none of that resolves the direction of causation.
He is equally direct about the monitoring implication: a month-over-month change in mention share is “a reason to investigate, not proof.” Five prompts per category is a thin sample by design. It supports a trend line and a triage queue. It does not support attributing a two-point move to last month’s content push.
That discipline is the difference between a measurement practice and a dashboard that generates false certainty. We catalogued the KPIs five companies actually track precisely because the reporting layer is where most AI-search programs quietly lose credibility.
Do this now
Pick your top three revenue categories. For each one, run five buyer-intent prompts through ChatGPT this week, record which brands are named in the prose, and calculate your mention share and your margin to the runner-up. Ignore citations for this exercise entirely.
You will land in one of three positions. Ahead by more than 5 points and named in four of five prompts: you own it, and the job is holding the margin. Ahead by under 2 points: contestable in either direction, and worth a defensive quarter. Not named in three or more prompts: unsettled, which is where 53.7% of categories and the overwhelming majority of demand currently sit.
Write the numbers down with the date. Repeat monthly. Treat a move as a question, never as an answer.
This analysis draws on Does topical authority matter in AI Search? (Growth Memo, July 2026).
Victorino Group helps organizations build AI-search measurement they can defend to a CFO, starting with the metric that actually moves. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation