Topical Authority in AI Search: Why Citations Travel Further Than Brand Mentions
Brands are often cited outside their established expertise, but they are rarely named there.
Brands that already own a category show a small but consistent advantage when they expand into other categories.
The topical authority effect is not the same in every industry.
A category is not won until the brand earns repeat mentions, not just citations.
Most content teams eventually face the same budget question. You can keep publishing on the few topics where your brand is already known, or you can go broad and try to show up in more of the questions your buyers ask. In traditional SEO, topical authority meant staying close to your core expertise because search engines rewarded depth. AI search adds a second question: when ChatGPT uses your page, does it also name your brand?
An earlier Growth Memo study found that some brands act as "category owners" in AI answers and hold that lead for months. This study asks what happens next. Once a brand has proven authority in one category, does that authority carry into nearby or distant categories, and if it does, does the brand show up as a recommendation or only as a source?
The short answer is that it depends on how far the topic is from what you already own.
What the data covers
The data comes from Semrush's AI Visibility Toolkit: US ChatGPT answers from January through June 2026, across 1,094 categories with 5 prompt variants each. Three tests were run.
Relatedness test. Compares a brand's new category appearances with the categories where it had already appeared in at least 3 of 5 prompts. This uses 45,578 expansion appearances by 1,458 mapped brands.
Future performance test. Tracks brand mentions and citations across 6 months.
Breadth versus depth test. Checks whether brands cited or named in many categories lose anything by covering topics broadly instead of deeply. This uses 283,215 citation observations and 76,493 named-brand observations, each paired with the next month's result.
Two outcomes are measured separately. A citation means ChatGPT linked to the brand's website as a source. A mention means ChatGPT named the brand in the answer. They are kept apart because ChatGPT produces far more citations than brand mentions.
All findings are associations after controls, so they show what tends to go together. They do not prove that publishing in a category, or expanding into one, caused the result.
Citations travel further than mentions
The study matched each of the roughly 1,000 categories against each brand's core expertise using semantic similarity, which is a measure of how close two pieces of text are in meaning.
When a brand appeared in a category far from its core expertise, about half of those appearances were citations the next month, about 25% were mentions, and only 9% were both cited and named. In the categories closest to the brand's expertise, 74% were cited, 44% were named, and 34% were both.
Citation and naming
Staying close to your expertise does not get you cited. It gets you named.
The share cited without attribution holds near 41% whatever the subject. Every point of gain comes from the named share, which runs from 9.0% at the edges of a brand’s expertise to 34.0% at its centre.
| Relationship to core expertise | Composition of all expansions | Citation only | Cited and named | Total cited | Of those cited, also named |
|---|---|---|---|---|---|
| Least related | 40.9% | 9.0% | 49.9% | 18.0% | |
| Less related | 42.4% | 16.9% | 59.3% | 28.4% | |
| More related | 40.7% | 25.9% | 66.6% | 38.9% | |
| Most related | 40.0% | 34.0% | 74.0% | 46.0% | |
| Spread | 0% 100% of expansions | 2.4 pts | 25.0 pts | 24.1 pts | 28.0 pts |
Bars run on a fixed 0 to 100% scale, so the grey segment holding its width down the column is the finding: the unattributed share does not move. Only two columns are independent measurements. Total cited is the sum of the two before it, and the final column is the named share divided by that total.
The citation-only share barely moves, from 40.9% to 40.0%. What changes is whether the brand also gets named. That means ChatGPT will use a well-built page as a source on almost any topic, as long as it meets the usual criteria for a citation, but it is much more likely to recommend the brand by name when the topic sits close to what the brand is already known for.
Depth matters more than breadth
A domain can show up once in 20 categories and own none of them. The study separates three measures:
Presence measures whether the brand appears in answers for the topic at all.
Depth measures how many of the 5 prompt variants the brand appears in.
Category performance measures the brand's share of the citations or mentions in those answers.
Showing up in many categories did not hurt citation share in this sample. The association with next-month citation share rose from +0.012 for brands that appeared in 1 of 5 prompts to +0.062 for brands that appeared in all 5. There was no sign of a "spread too thin" effect for citations.
Brand mentions behaved differently. Brands that appeared in only 1 of 5 prompts across their categories saw a negative association with mention share (-0.051). Brands that appeared in all 5 saw a slightly positive one. In essence, the penalty is for shallow presence. Being spread across many categories is not the problem when the brand shows up consistently in each one.
Two limits matter here. The data does not support a universal limit on how many categories a site can serve, and it also does not support the idea that a site can serve an unlimited number. The effects are also small, and other factors, such as writing style, overall brand authority, and third-party mentions, may matter more than topical authority.
The pattern differs by industry
Citation vs mention share
Citation share turns positive in every industry. Mention share splits by vertical.
At 3 and 5 of 5, citation share is positive across finance, real estate, legal and healthcare. Mention share is positive for finance and real estate at those levels, and negative for legal and healthcare at every level.
| Industry | Citation share | Mention share | ||||
|---|---|---|---|---|---|---|
| 1 of 5 | 3 of 5 | 5 of 5 | 1 of 5 | 3 of 5 | 5 of 5 | |
| Mention share positive at 3 and 5 of 5 | ||||||
| Finance | +0.054 | +0.105 | +0.139 | +0.069 | +0.113 | +0.041 |
| Real estate | +0.021 | +0.100 | +0.135 | −0.041 | +0.039 | +0.011 |
| Mention share negative at every level | ||||||
| Legal | +0.005 | +0.029 | +0.016 | −0.125 | −0.083 | −0.058 |
| Healthcare | −0.038 | +0.005 | +0.035 | −0.107 | −0.054 | −0.037 |
All 24 bars share one absolute scale: the centre line is zero and a full half-bar equals ±0.150. Gold extends right for positive values; charcoal extends left for negative.
A category counts once a brand appears in 1, 3, or 5 of the five prompt variants. Positive numbers mean broader coverage is associated with a higher share the next month after controls. Negative numbers mean a lower share.
Finance and real estate show the strongest repeat-citation pattern. In finance, the citation association rises from +0.054 at 1 prompt variant to +0.139 at all 5. Real estate rises from +0.021 to +0.135. In these industries, a domain that appears repeatedly across a category's prompts tends to hold more citation share the following month, so expanding into related categories looks like a workable way to grow source visibility. That does not mean every finance or real estate brand can win every adjacent category.
Legal and healthcare set a tighter limit on brand mentions. In legal, the mention association stays negative even at full 5-of-5 coverage (-0.058). Healthcare follows the same pattern (-0.037). More consistent coverage can improve a site's citation position in these industries, but it does not reliably lead to the brand being named. A likely explanation is that these are higher-stakes decisions, where being a credible source is not enough for ChatGPT to recommend the brand.
The practical result is that finance and real estate teams can use citations as an early success measure when they test related categories. Legal and healthcare teams need a higher bar, and a category is not won until the brand earns repeat mentions.
How to choose which categories to expand into
The study's author lays out a category-map method for deciding which categories are close enough to pursue. The steps below summarize it.
1. Find your anchor categories. Export the categories where your brand already appears in at least 3 of 5 prompt variants. These are your anchors, not every category where the brand appeared once. For each anchor, record the category name and its 5 prompts, citation depth and share, mention depth and share, and the customer, problem, product, and decision stage behind the prompts. Keep anchors where the brand is both cited and named separate from anchors where it is only cited, because the first group is stronger.
2. Build a target list from real customer questions. List the categories the business wants to win next, and write 5 prompt variants for each one that reflect how people would actually ask the question in ChatGPT. Add one row per target to the map, with the target category, closest anchor, shared customer problem, citation depth, mention depth, business value, and priority. Compare individual question clusters, not whole industries.
Expansion target
Connecticut credit-union CDs show one citation and no mentions. With high business value, it’s the one to test now.
The target shares its customer problem with an existing anchor, teacher credit-union CD rates. Depth is 1 of 5 for citations and 0 of 5 for mentions.
| Category fit | Current depth | Assessment | ||||
|---|---|---|---|---|---|---|
| Target category | Closest anchor | Shared customer problem | Citation | Mention | Business value | Priority |
| Credit-union CD rates in Connecticut | Teacher credit-union CD rates | Choosing a CD rate from a credit union | 1 of 5 | 0 of 5 | High | Test now |
Depth counts are shown as five dots: gold for each one counted, open for the rest; a count of zero shows grey dashes. Business value and priority are assessments, not measurements.
3. Score each target by how close it is to an anchor. Compare the target's title and prompts with every anchor and ask whether it covers the same customer, problem, product, decision, and type of evidence. A subject expert can do this by hand for a small set. For a larger set, use embeddings, which are numeric representations of text that let you measure similarity, and use the closest match as the target's score. Start with the closest categories that also matter to the business.
4. Give every target one success measure. A source-authority target succeeds when the domain is cited in at least 3 of 5 prompts and holds that coverage in the next snapshot. A brand-recommendation target succeeds only when the brand is named in at least 3 of 5 prompts and holds it in the next snapshot. Legal and healthcare teams should use the second measure.
5. Turn the map into a quarterly plan. Choose the 5 closest, highest-value targets. Build content, first-party evidence, and third-party proof around those questions, then review citations and mentions at the next snapshot. Keep expanding from categories that produce both signals, and pause categories that produce only occasional citations, especially when the brand is never named.
How to run the expansion
Decide what success looks like before publishing. Give each target category one job: to be cited as a source, to be named as a recommendation, or both. Most brands care more about mentions, unless the business is a review or publishing site. Build a simple scorecard for each category with its 5 prompts, the number of answers that cite the domain, the number that name the brand, and the citation and mention shares. That is the baseline.
The job decides the work. A source category needs the best answer and evidence. A recommendation category needs clear product proof, credentials, customer outcomes, reviews, and third-party validation. The study's example is a credit union. "How CD interest is taxed" is a source category, so the work is a correct, current explainer with primary sources and a review date. "Best credit union CD rates in Connecticut" is a recommendation category, so the work is current rate tables, plainly stated membership eligibility, branch and service proof, and listings on third-party rate sites.
Require repeat evidence. Run the same prompts at the next snapshot and record citations and mentions separately, from 0 of 5 to 5 of 5. Use 3 of 5 as the minimum for continued investment, then sort each category into one of four groups:
Scale: the domain is cited in at least 3 of 5 prompts in 2 consecutive snapshots. For a recommendation category, the brand is also named in at least 3 of 5.
Improve: the category produces 1 or 2 appearances but matters to the business. Strengthen the answer, the evidence, and the brand proof.
Pause: the category does not earn repeat citations after 2 snapshots. Move the effort elsewhere.
Reclassify: the domain earns repeat citations but few mentions. Keep it as a source category and do not report it as a recommendation win.
This keeps a team from treating one good answer as lasting visibility.
Expand in small clusters. Choose 3 to 5 categories around the same customer problem, product, or decision. Build one strong pillar page that answers the core question, with supporting pages or sections for the 5 prompt variants. Do not start a second cluster until the first has a result. At the next snapshot, expand to the next related category if repeat citations or mentions hold, deepen the existing category if the brand appears but not consistently, and stop if the brand is still absent.
Raise the bar in legal and healthcare. Treat citations as an early signal, not the finish line. In these markets, make the brand's proof visible on every page: name qualified authors and their credentials, show when the information was reviewed or updated, cite primary sources, guidelines, and research, explain the service or care model clearly, and add independent proof such as reviews, professional affiliations, case evidence, or reputable third-party coverage. Do not count a category as won until the brand is named in at least 3 of 5 prompts and holds that in the next snapshot.
What this means for your content program
If your goal is to be cited, you have room to publish beyond your core topics, as long as each page is strong enough to be used as a source. If your goal is to be recommended, which is the goal for most brands outside of publishers and review sites, the work stays close to home. Expand into categories that share a customer, a problem, or a product with the categories you already own, and go deep in each one before adding the next.
Over the next 6 to 18 months, this most likely matters more, because each new model release changes how many sources an answer uses and which ones get named. Brands that can show they own a set of related questions will most likely hold mentions better than brands with thin coverage across many.
How to start today
List the categories where your brand already appears in at least 3 of 5 prompt variants, and track citations and mentions separately. This is your baseline, and it shows which topics count as your core expertise in AI search.
Give each new target category one goal, either to be cited or to be named. The work is different: a citation goal needs the best answer and evidence, and a mention goal needs product proof, reviews, credentials, and third-party validation.
Expand in small clusters of 3 to 5 categories that share a customer problem with an existing core category. In the closest categories, 46% of cited expansions were also named, about 2.5 times the 18% in the least related ones.
Review each cluster at the next monthly snapshot before adding another. One good answer is not a trend, so look for coverage that holds across two snapshots.
In legal, healthcare, or other high-stakes categories, count a category as won only when the brand is named, not just cited.
Citations show that your content is useful, but mentions show that the market sees you as a choice, and the data suggests you earn those closest to the expertise you have already proven.
Methodology
Sample. Semrush US ChatGPT data from the AI Visibility Toolkit, January through June 2026: 1,094 categories and 5 prompt variants per category in every monthly snapshot. The breadth models use 283,215 citation observations and 76,493 named-brand observations from January through May, each with a next-month outcome. A domain that disappears from a category receives a share of zero. The relatedness analysis uses 45,578 expansion appearances by 1,458 mapped brands, with February through May as starting months, because each appearance needs prior history and a following-month outcome.
Measures. Citation share is a domain's share of all source citations in a category's answers. Mention share is its share of all named brand mentions. Depth is the number of a category's 5 prompt variants where a domain appears. Breadth is the number of categories where it appears in the current month. Demonstrated expertise means appearing in at least 3 prompt variants in a category before the target month. An expansion appearance is a brand showing up in a category where it had not previously reached that depth. Category relationship comes from multilingual semantic similarity between the target category's title and prompts and the brand's prior expert categories.
Controls. The breadth models control for current share, current depth, the other signal's share, Authority Score, organic traffic, branded search demand, and differences by category and month. The relatedness model controls for current citation and mention depth and share, breadth, Authority Score, organic traffic, branded search demand, same-industry status, and differences by category and month.
Limits. Citations are domains and named brands are text entities, and Semrush maps both to a brand where possible, so the relatedness results describe mapped brands only. The design measures associations, not a controlled test, so it cannot show that publishing or expansion caused the outcomes. It also does not produce a reliable maximum number of categories a site can serve.