Semrush ChatGPT Topical Focus Study (2026)

What the follow-up study asked

The July research established that most categories in Semrush's ChatGPT panel had no brand appearing consistently, and that domain-level metrics did not separate the brands that did from the ones just behind them. It left the obvious question open. If not those, then what?

The August study runs on the same panel — 1,094 US categories, five prompts each, January through June 2026 — and asks a narrower question: when a brand shows up in a category it has not established itself in, does the answer cite it, name it, or both, and does that change with how far the new category sits from the ones it already answers well?

Two constructs do the work. A brand's core expertise was any category where it had already appeared in at least three of the five prompts before the month being measured. Category closeness was scored by semantic similarity between a target category's prompt set and the prompt sets of the categories that brand already owned in that sense.

The resulting data set holds 283,215 domain-category citation observations and 76,493 observations of a brand being named in an answer, with 45,578 category-expansion appearances mapped across 1,458 brand entities. The original is published as the Semrush study on how topical authority spreads in ChatGPT.

Note the ratio in those two totals before going further. Citations outnumber named mentions by roughly 3.7 to one in this panel. The system links far more sources than it names brands, which is the background condition every finding below sits on.

Citations traveled. Named mentions did not.

Near a brand's established expertise, 74% of its appearances came with a citation, 44% with a naming in the answer text, and 34% with both. At the far end of the similarity range, those three figures fell to 50%, 25% and 9%.

Read the first and third columns against each other. Citation frequency dropped by about a third across the distance range. The share of appearances carrying both signals dropped by nearly three quarters.

Kevin Indig stated the asymmetry directly:

Being cited in many categories does not show a spread-thin effect in this sample. But brand recognition works differently. The categories where a brand consistently earns named mentions are the ones from which it can credibly expand.

A publisher can be linked as a source in a subject it has no standing in. It is much less likely to be named as a participant in that subject. Whatever produces a citation and whatever produces a mention are not responding to the same thing, and the gap between them widens with distance.

The conversion rate is the number to keep

The sharpest figure in the study is not either of those columns but the ratio between them. Of the citations occurring in the least related categories, 18% also carried a named brand. In the most related categories, 46% did.

That is a citation-to-mention conversion rate, and it is more useful than either raw count because it controls for how often a brand appears at all. A domain publishing heavily into a distant subject can raise its citation count indefinitely without moving the conversion figure, and the conversion figure is the one attached to the outcome the study treats as commercially meaningful.

It also reframes a result from the earlier study in this series, which found that only 21% of the most-cited domains in a category were also the most-mentioned brand, with a slightly negative correlation of -0.229. Read alone, that looks like citations and mentions are unrelated or perversely opposed. Read with the distance data, it looks like an average taken across a distribution that has structure in it: close to a brand's core the two signals converge, far from it they come apart, and pooling the range produces a number near zero that describes neither end.

No AI surface reports this ratio, and no vendor dashboard the author is aware of presents it as a metric. It can be assembled from any tool that separates citations from mentions at the category level.

Retrieval and generation are two different systems

The asymmetry has a mechanical explanation that the study does not offer and that follows from what the operators already document.

A citation is a retrieval outcome. A document matched a query, cleared the eligibility and quality conditions the platforms describe, and was linked. Retrieval scores documents. It holds no position on whether the publisher of that document belongs in the subject, which is exactly why a competent article can be surfaced in a category its author has no standing in.

A named mention is a generation outcome. The model produces a brand name because its representation associates that entity with that domain of expertise, assembled from the whole corpus it absorbed rather than from the documents retrieved for this one query.

Documents are portable between categories. Entity associations are not, because they were not built per query. That single distinction predicts the shape of the study's central chart without any appeal to a ranking factor, and it is testable: a brand whose citation share rises in a distant category while its mention share stays flat is the case the framing predicts.

It also means the two signals answer to different work. Citation share responds to publishing. Mention share responds to whatever put a brand into the corpus in the first place, which is largely other people's writing.

The depth threshold, and the penalty below it

Within a category rather than across categories, the study reports a threshold effect. Brands present in a single prompt out of the five saw their mention share fall rather than rise. That penalty did not disappear until the brand reached three of five prompts.

Stated as the study states it, partial presence in a category was associated with a worse outcome than the panel average, not merely a smaller gain.

If that holds up, it inverts a piece of standard practice. Publishing one article into a new subject to see whether it gains traction has been a low-cost, low-risk test in organic search for as long as organic search has existed, because a page that fails to rank simply sits there. A threshold effect of this shape would mean the test is not free.

Whether it holds up is the open question, and it is the subject of the section below. A single-prompt appearance is also what a small or unfamiliar brand looks like inside a category, which makes this the finding in the study most exposed to the direction problem.

Legal and healthcare did not follow the pattern

Sector changed the relationship. Financial services and real estate converted wider category coverage into stronger citation leverage. Legal and healthcare did not: their citation gains from breadth were weaker, and mention share trended negative even for brands covering all five prompts in a category.

Full coverage, and the mention trend still pointed down.

The shape of that will be familiar to anyone who worked on medical or legal sites through a decade of search quality updates. High-stakes regulated subjects have carried a raised bar in ranking systems for years, and something with a similar effect appears to operate here.

That is an inference and should be labeled as one. The study did not test author credentials, licensure signals, primary research or product specifics, and its authors said so. What the data supports is narrower: in these two sectors, coverage depth did not behave as a lever in the way it did elsewhere. What produces the ceiling is not established, and the obvious hypothesis — that credentialing and third-party validation carry the weight instead — has not been measured by anyone.

The direction problem this study cannot settle

Every relationship reported here is cross-sectional. There was no intervention, no control group and no randomization, and the authors consistently wrote associated with rather than any causal verb. That discipline is worth more than the findings and it is the first thing lost in summary.

The specific worry is reverse causation on the depth threshold. Thin coverage may suppress mention share. Thin coverage may equally be what an unfamiliar brand looks like inside a category. A company appearing in one of five prompts is plausibly a company few buyers could name unprompted, and its recognition would lag whatever it published. On this data the single-prompt penalty could be tracking brand size and reporting it as a coverage effect, and nothing in the design separates the two.

Sample scope is the second limit. The expansion analysis rests on 1,458 mapped brand entities, against the 50,000-plus brands behind the earlier study on the same panel. Narrow bases produce confident-looking curves.

The authors also list what the study does not account for: writing style, content quality, originality, overall brand authority and third-party mentions. Any of those could be the variable that moves both coverage and recognition, which is the ordinary shape of a confounder.

None of that reduces the study to noise. It places it where it belongs: a coherent, well-described pattern that justifies planning work and does not justify a causal claim, from a vendor that was careful to say as much.

Frequently asked questions