Brand Mentions vs Backlinks for AI Visibility: A Claim Check on Ahrefs' 75,000-Brand Studies

"Mentions are worth three times what links are" has become a stock slide in AI search decks. It comes from a correlation table, and a correlation table does not say how much more important anything is.

Marketers weighing a shift of budget away from link building because of that line are the audience. We opened the Ahrefs studies behind it, read what each number measures and marked every claim as Holds, Partly holds, Unverified or Does not hold. We have no commercial link with Ahrefs. We did not check the Muck Rack report directly, and we say so where it matters.

The verdicts at a glance

Claim	Traced to	Verdict	Reason
Mentions outscore backlinks three to one	Ahrefs, May 2025	Partly holds	Dividing two correlations gives a ratio of fit, not of effect
Links have stopped counting in AI answers	Ahrefs, December 2025	Partly holds	Weak in the data, but weak is not zero, and the data is correlational
Unlinked brand names are ignored	Ahrefs, May 2025	Does not hold	The mention measure includes unlinked text
YouTube mentions are the strongest signal	Ahrefs, December 2025	Holds, narrowly	Highest value in the study, with a title-only definition
Most AI citations are earned media	Muck Rack, second-hand	Unverified	Read only through a summary

Claim 1: "Mentions outscore backlinks three to one"

Origin: Ahrefs, through Louise Linehan, released the first study on 26 May 2025. It covered 75,000 brands, limited itself to domains rated above 40 whose top keyword drew 800 or more searches a month, and measured brand mentions in Google AI Overviews through Ahrefs Brand Radar. Roughly 26% of the brands had no AI Overview mentions at all.

What the data says: On the Spearman scale, branded web mentions scored 0.664 against AI Overview mentions, while backlinks scored 0.218. The author is plain about the limit, stating that correlation does not establish causation, and that every factor sat between moderate and very weak.

Verdict: Partly holds. The division works out near three, so nobody miscalculated. But a correlation coefficient describes how closely two series move together, not how much one changes the other. Use it like this: write "mentions tracked AI visibility more closely than backlinks did" and keep the caveat right beside it.

Claim 2: "Links have stopped counting in AI answers"

Origin: A common reading of both studies.

What the data says: In the first study, referring domains reached 0.295 and Domain Rating 0.326, so link-based measures stayed positive. In the follow-up published in December 2025, Ahrefs extended the work to ChatGPT and AI Mode and described backlinks as very weak without giving exact values.

Verdict: Partly holds. For ChatGPT and AI Mode the evidence points to a thin relationship. It still does not show that links are useless, because a link and a mention often come from the same article. Use it like this: keep the link work that earns coverage, and stop counting links as the goal in themselves.

Claim 3: "Unlinked brand names are ignored"

Origin: Older SEO advice, repeated in AI discussions.

What the data says: Ahrefs defines branded web mentions as the brand name anywhere on the web, linked or unlinked, and that is the factor with the highest correlation in the first study. Branded anchors, the linked kind, came lower.

Verdict: Does not hold. For AI Overview visibility, the strongest factor in the study explicitly includes unlinked text. We do not know the mechanism, and the study cannot supply one. Use it like this: track unlinked mentions in PR reporting, as described in the PR machine audit.

Claim 4: "YouTube mentions are the strongest signal"

Origin: The December 2025 Ahrefs study of ChatGPT, AI Mode and AI Overviews, reviewed by Ryan Law.

What the data says: YouTube mentions scored about 0.737, ahead of branded web mentions, which ranged from 0.656 to 0.709 depending on the platform. The study counts a YouTube result only when the brand appears in the video title, and not when it appears in the spoken transcript.

Verdict: Holds, narrowly. It is the top value in that data set. A gap of a few hundredths between two correlations is small, and the study carries no replication. Use it like this: treat video titles as one more place to be named, not as a replacement plan.

Claim 5: "Most AI citations come from earned media"

Origin: Muck Rack's "What is AI Reading?" report. We reached it only through a ContentGrip summary, which reports 89% earned media across more than one million links, 95% from non-paid sources, and no definition of earned media or list of assistants.

Verdict: Unverified. Muck Rack sells media monitoring, the summary is second-hand, and later editions are said to report different shares. Use it like this: say "Muck Rack reported, as summarised by ContentGrip" and rely on the direction, not the figure.

What these claims share

Taken together they say that brands others name tend to be the brands assistants name. None says how much effort in produces how much visibility out. A fair summary: across big Ahrefs samples, widely discussed brands showed up more often in AI answers.

What a Spearman value tells you, and what it hides

Spearman's coefficient ranks brands on each factor and asks whether the two rankings agree. A score near one means the orders match closely. A score near zero means they barely relate. It says nothing about how many extra AI mentions one more link would produce, and it cannot separate a cause from a shared driver such as company size.

Two details in the follow-up help calibrate the numbers. The three platforms' brand mentions overlapped strongly, between 0.749 and 0.821 depending on the pair, so the same well-known names recur across ChatGPT, AI Mode and AI Overviews. That fits a story where fame drives everything. Ahrefs also cautions that its YouTube figure counts a brand only when the name sits in the video title, so a brand discussed at length in a transcript scores nothing.

Read these studies as a map of who is already visible. They are a weaker guide to how a smaller brand becomes visible, and that is the question most readers actually have.

One more caution on the first study: it measured Google AI Overviews only, from a sample of brands that already had a top keyword with real search demand. Newer or niche brands sit outside that frame, and so do products with little search volume. If that describes you, the studies are a hypothesis generator, and your own prompt tracking is the evidence.

Which claims to repeat, and which to retire

Repeat the narrow versions: mentions correlated more closely than links in these samples; unlinked text is part of the measure; YouTube titles scored highest in one study. Retire the broad versions: links are dead, mentions are three times better, one signal decides visibility. Each retired claim has the same flaw, which is turning a description of existing brands into an instruction for a different brand.

The cost nobody puts on the slide

Our read is that the studies mostly measure how big a brand is. Ahrefs itself notes that large brands such as Nike, Apple and Amazon appear across all three platforms, so the sample leans toward established names. A well-known company is mentioned everywhere and also gets recommended by assistants, with neither one causing the other. If you are small, a rank correlation across 75,000 mostly large brands tells you very little about what a month of outreach will do for you.

That is why we would not move budget on this evidence. Test instead. Put one half of a month's outreach behind a single original-data piece pitched to a few relevant writers, the other half behind ordinary link requests, and log mentions, links, branded searches and a fixed prompt list for each. A controlled test on small brands showing mention growth reliably leading citation growth would settle the matter; until then, either camp is arguing from another company's data.

A ten-minute routine for any statistic

1. Open the study. Ignore the headline that cites it.
2. Note what was measured: a correlation, a share, a rate or an effect.
3. Read the filters on the sample, for example a Domain Rating floor.
4. Check that the assistant studied is one your buyers use.
5. Rerun it on your own prompts using the spreadsheet playbook.