Brand mentions beat backlinks for AI visibility (75k-brand study)

Across 75,000 brands, the off-site factor that lined up most closely with showing up in Google's AI Overviews was branded web mentions, at a Spearman correlation of 0.664. Backlinks came in at 0.218 — roughly a third as strong, and below every measure of how much the web talks about you by name. If your plan for getting into AI answers is a link-building budget, this is the study that says you aimed at the wrong end of the list.

We did not measure any of that. Ahrefs did. The study is An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied) by Louise Linehan and Xibeijia Guan, published 26 May 2025 — fourteen months before this post. We are putting it in the second paragraph rather than a footnote because we also sell an AI visibility tool, which means "mentions matter more than links" is a conclusion we have a commercial interest in being true. Ahrefs has the same problem in reverse: the numbers were produced by Brand Radar, a product they sell. Grade △▲ — large observational dataset, no control group, published by a company selling a related product.

So here is the rule for the rest of this post. Every figure attributed to Ahrefs is theirs. Every figure marked ◐ is our own first-party measurement on one brand, in one category, in July 2026. The two are never added together, and we say explicitly where ours cannot speak to theirs.

A branded web mention is any occurrence of your brand name in text on a page anywhere on the web, whether or not that page links to you. A backlink requires the link. That difference is the whole subject of this post.

Key takeaways

• Mentions outranked links, and the gap is wide. Branded web mentions 0.664, branded anchor text 0.527, branded search volume 0.392 — all three above the strongest link metric, Domain Rating at 0.326. Referring domains 0.295, backlinks 0.218 (Ahrefs, △▲).
• Paid barely registered. Ad traffic 0.216 and ad cost 0.215 sit within 0.003 of backlinks. Ranking factors that close together is reading noise; the article publishes no confidence intervals we could find.
• Every brand in the sample already had Domain Rating above 40. That filter removes the brands with weak link profiles before the link correlation is calculated. A weak coefficient measured on a sample with the low end cut off is not the same finding as "links don't matter."
• The shape is winner-take-all. By median, brands in the top quarter for web mentions recorded 169 AI Overview mentions; the quarter below 14; the second quarter 3; the bottom quarter 0. And 26% of brands had zero mentions in AI Overviews at all (Ahrefs, △▲).
• This is Google AI Overviews, and the study was published in May 2025. One surface, one vintage — and May 2025 is the publication date, not a collection window; we could not find a collection window stated anywhere in the article. In our own data the engines do not run the same mechanism: across 154 head-term citations, ChatGPT gave 66% to vendor product pages and 0% to video, while Gemini gave 38% to vendor listicles and 26% to video ◐.
• Correlation, and Ahrefs says so themselves. Their write-up states "correlation ≠ causation." Large brands get mentioned and get named in answers, and being important to the category is a plausible cause of both.

What Ahrefs actually measured

Spearman is the right choice for this question, and worth a sentence because it changes how you read the numbers. It asks whether the ranking of brands by web mentions matches their ranking by AI appearances, without assuming the relationship is a straight line. So 0.664 does not mean "66% of anything." It means the two orderings agree strongly but not perfectly.

The nine factors, in order

Strongest to weakest association with AI Overview presence, as plotted:

Two more factors appear in the same Ahrefs chart below these nine: URL Rating at 0.18 and number of site pages at 0.17. URL Rating is a link metric too, so the tidy line "backlinks came last of the link metrics" is not true — backlinks came last of the three link metrics in the ranking above, and URL Rating is lower still. We are pointing that out because it is the sort of sentence that survives being repeated after it stops being accurate.

It is two clusters, not nine factors

Reading this as nine independent levers is the most common mistake made with it. It is closer to two.

The top three all contain the word "branded." Branded web mentions, branded anchor text and branded search volume are three instruments pointed at one underlying quantity: how widely a brand name is known and repeated. They are not three findings that happen to agree. They are one finding measured three ways, which is why they cluster at the top together.

The link metrics are one cluster too. Domain Rating, referring domains, backlinks and URL Rating all describe the same link graph at different resolutions.

So the honest one-line summary is not "mentions beat backlinks by 0.446." It is: the awareness cluster ranked above the link cluster on this surface, and the paid cluster ranked below both. That is a weaker claim than the headline, and it is the one the data supports.

The bottom of the list is a tie

Backlinks 0.218, ad traffic 0.216, ad cost 0.215. Those three are 0.003 apart. Referring domains 0.295 and branded organic traffic 0.274 are 0.021 apart.

We could not find confidence intervals, standard errors, or per-factor sample sizes in the article. Without them, ordering coefficients that differ in the third decimal place is decoration. Treat the chart as three tiers — the branded cluster, the link cluster, and the paid tail — and do not build a strategy on any single position inside a tier.

The winner-take-all shape, which is the more useful chart

Ahrefs split brands into quartiles by web mentions and reported the median AI Overview mentions in each:

The step from 14 to 169 is about twelvefold. The step from 0 to 3 is the one most brands are actually standing on. And separately, 26% of the 75,000 brands had zero AI Overview mentions — on domains that all had Domain Rating above 40, which is to say a quarter of the sample was invisible on this surface despite being established enough to pass the filter.

We could not find a stated time window for the 169, so read it as the shape of a distribution rather than as a monthly rate.

Three ways this number misleads you

One: every brand in it already had Domain Rating above 40. This is the objection we think matters most and we have not seen it raised elsewhere. If you remove all the brands with weak link profiles before you compute the link correlation, the remaining variation in links is compressed — and a compressed range mechanically produces a smaller coefficient. Backlinks scoring 0.218 among brands that already have strong link profiles is not evidence that a brand with no links would do just as well. It is evidence that once you are above the DR 40 line, adding more links tracks AI presence weakly. Those are different sentences and only the second one is supported. To be clear about our own uncertainty: we have not quantified how much attenuation the filter causes, and neither, as far as we can find, has anyone else.

Two: it flatters big brands. Everything Ahrefs measured — mentions, branded search, Domain Rating — is downstream of already being big. A 75,000-brand correlation is a photograph of who is currently winning, not a set of instructions for how a small brand climbs. The lever it points at ("get mentioned more") is real; the coefficient is not your growth rate, and nothing in this dataset shows a brand moving from one quartile to another.

Three: it is one engine, one moment. AI Overviews is Google's surface, and the only date the moment is anchored to is the publication date, 26 May 2025 — no collection window is stated anywhere in the report that we could find. If your buyers ask ChatGPT or Perplexity, the retrieval mix — and therefore which signals win — can differ; our own engine split below is one demonstration that they do. The only way to know your position on the engines your customers actually use is to ask those engines, repeatedly, and count.

Where our own data touches this, and where it does not

First, the honest boundary: we have never measured mentions against citations. We cannot replicate Ahrefs, we do not have a 75,000-brand panel, and nothing below is a confirmation of 0.664. These are three places our measurements sit next to theirs, all ◐ — one brand, one category, July 2026, raw responses retained, and we sell a product in this space.

That last row is the one we would hand to anyone about to act on the 0.664. Branded signals and branded questions are different objects, but they fail in a similar way: both can produce a healthy-looking number that describes people who already know you.

The rank result is unpacked in Why your brand isn't in AI Overviews, and the question-type split in How to track your brand in AI answers.

What a growing brand can act on

None of this is a promise about your numbers. What follows is the mechanism each item works through, stated as such, plus what we can and cannot back.

• Get named in the comparison content of your category. "Best X for Y" posts, alternatives pages, category roundups. The mechanism: those pages are what an engine retrieves when someone asks for a shortlist, and a brand absent from the source list cannot be named in the answer. On the Gemini side of our data, vendor listicles were 38% of head-term citations ◐.
• Treat unlinked mentions as the goal, not the consolation prize. A review that names you without linking still puts your name in text a model can read. Ahrefs' top factor counts exactly this, and their fourth-ranked factor (Domain Rating, 0.326) does not.
• Feed branded search rather than measuring it. Branded search volume at 0.392 is a readout of demand, not a lever you pull directly. Everything that makes more people type your name — launches, talks, community presence, being useful in public — writes to the same variable.
• Do not stop building links on this evidence. The DR > 40 filter is the reason. This study cannot tell you what happens below that line, and both of the metrics it ranks fourth and fifth are link metrics sitting above backlinks.
• Measure the surfaces you actually sell on, over time. Because the distribution is winner-take-all, "we're probably fine" is an expensive guess. A single check is a screenshot; a series is a measurement.

If AI presence is driven by how much the web says your name, the number to watch is not your Domain Rating. But be exact about which number, because "mention share" is now used for two different things. Ahrefs' variable is web mentions: your name in text on other people's pages. What a tool like ours reports is answer mentions: your name inside the AI response. The first is the proposed cause and the second is the effect, so reporting the second as though it were the first folds 0.664 into a tautology.

And peekr does not measure web mentions. There is no web or news mention source wired into it — in our own off-page scorer, the two elements that would need one, linked mentions and unlinked mentions, are both marked unmeasurable for exactly that reason. So we cannot hand you the variable that scored 0.664. We can hand you the outcome side of it, and whether that is moving. For the input side you want Ahrefs' instrument or a media-monitoring tool, not ours.

What this post cannot tell you

• We did not run this study and we do not have the dataset. We read the published write-up. The underlying data does not appear to be public.
• We could not find confidence intervals, per-factor sample sizes, or a definition of what counts as one "AI Overview mention" in the article. Where we say a gap is within noise, that is our judgement about the missing intervals, not a computed result.
• We could not find a stated collection window anywhere in the article, including for the quartile figures. 26 May 2025 is the date it was published, not the period it covers.
• The range-restriction argument is reasoning, not a measurement. We have not quantified the attenuation the DR > 40 filter causes.
• Nothing here shows a brand moving. Every figure in the study is cross-sectional. No before-and-after, no control group, on either side.
• Our ◐ figures are one brand, one category, one week, and they were collected for different questions than this one.
• We cannot tell you that earning mentions will change your appearance rate, and neither can Ahrefs. That experiment has not been published by anyone we can find, including us.

Frequently asked questions

Do backlinks still matter for AI search?

This study cannot answer that, and the reason is its own sample. Backlinks correlated at 0.218 with AI Overview appearances, but every brand measured already had Domain Rating above 40 — the brands with weak link profiles were filtered out before the coefficient was calculated. What the data supports is narrower: among brands that already rank and already have links, adding more links tracks AI presence weakly, while being mentioned tracks it strongly.

What is a branded web mention?

Any appearance of your brand name in text on a page anywhere on the web, with or without a link back to you. Ahrefs measured this as a separate factor from backlinks and from branded anchor text, and it scored highest of the nine at 0.664.

Does this apply to ChatGPT, Gemini and Perplexity?

Unknown from this study, which measured Google AI Overviews only. Our own data says the engines differ enough that a single playbook is a bad bet: in one 154-citation set, ChatGPT cited vendor product pages 66% of the time and video 0%, while Gemini cited video 26% and vendor listicles 38% ◐. Treat 0.664 as a strong prior for Google's answer surface and as an untested guess everywhere else.

Is 0.664 a strong correlation?

For observational web data across 75,000 entities, yes — it is unusually high, and the distance from the next cluster (0.326 for Domain Rating) is the part worth noticing. But Spearman measures agreement between two rankings, so it says the order of brands by mentions closely resembles the order by AI appearances. It does not say what fraction of appearances mentions explain, and it does not establish direction.

How do I measure my share of AI answers?

Count every brand named in the answers, not only yours, and report each rate with its denominator, split by engine and by question type. The method and the four ways the number misleads you are in How to track your brand in AI answers. If you are comparing tools for it, the features and prices are in 8 AI visibility tools compared.

This is not the 0.664 variable, and the two are easy to mix up. Ahrefs counted your name on other people's web pages — the input. Share of AI answers counts your name inside the answer — the outcome. We measure the second one and not the first: peekr has no web or news mention feed, so it cannot tell you your web mention count or your correlation.

Should I write conversational long-tail pages to get mentioned more?

Those are two separate ideas that often get bundled. Engines rewrite a prompt into their own short queries before retrieving, so sentence-shaped pages built to match sentence-shaped prompts optimise a string your page never sees — the evidence is in Long-Tail Prompts vs Head Terms. Being mentioned on other people's pages is a different job from how you phrase your own.

Where our own numbers come from

Since this post is largely about someone else's instrument, it is fair to describe ours. peekr asks a fixed list of questions across ChatGPT and Gemini, stores each raw response whole with the citation list attached, counts every brand named rather than only yours, and reports the rate per engine and per question type with the denominator attached. Raw responses are retained, so when the parser changes the old answers are recounted instead of two series being spliced together.

That gives you your own answer-mention and appearance numbers on both engines. Two limits worth stating in the same breath as the pitch: re-measurement is triggered by hand today, so a series is something you build rather than something that accrues while you sleep; and this does not reproduce Ahrefs' study, does not count web mentions, and will not tell you your correlation coefficient.

Run the free check on one URL — it reads the submitted page for whether an engine can fetch and parse it, then asks one unbranded question about your category and shows the raw answer with its sources. One question, one sample: a demonstration of the mechanism, not a baseline.

Method

Source: Ahrefs, Louise Linehan and Xibeijia Guan, An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied), published 26 May 2025. Grade △▲ — observational, no control group, and the measurement was produced by a product the publisher sells. We read the published article directly on 11 August 2026 and checked every figure quoted here against it. We did not receive or inspect the underlying dataset.

Figures marked ◐ come from our own stored measurements on one brand in the AI video-editing category, July 2026: Google SERPs and AI Overviews through a SERP API across 13 queries with an overview appearing on 12, and engine responses through the OpenAI and Google APIs, producing 154 citations in the head-term set with retrieval enabled, and 336 stored answers in the question-type set with retrieval off. Those two sets are reported separately and never merged — different surfaces, not comparable numbers. One brand, one category, no control group, and we sell a product in this space.

Marks used above: ◐ our own first-party measurement, no peer review, no control group, vendor-affiliated. △▲ observational study, no control group, published by a company selling a related product.

If you have run a before-and-after on earned mentions with a control group, in either direction, I would like to read it.