Do brand mentions matter more than backlinks for AI visibility?

The short answer

The evidence does not support a clean "mentions beat backlinks" rule.

It supports something narrower and less useful as a slogan: among domains with an established web presence, being widely talked about is associated with appearing in AI answers, more strongly than the number of links pointing at the site. That association changes sharply by platform and by sample, and every study reporting it says correlation is not causation, including the vendors who sell tools for it.

Four signals, not one

The practical mistake is treating four different observations as one metric called "AI visibility".

An answer can cite a page without naming its publisher. It can recommend a business while citing a directory. Measure them separately or the numbers mean nothing.

What the 75,000-brand study found

Ahrefs published this on 12 December 2025, covering 75,000 brands across ChatGPT, AI Mode and AI Overviews. Branded web mentions correlated with brand visibility as follows:

These are correlation coefficients. They are commonly rewritten as "mentions are three times more important than backlinks", which is wrong twice over. Dividing one coefficient by another does not produce an effectiveness multiple, and the outcome measured was how often a brand appeared, not whether its site was cited, visited or chosen.

Ahrefs states the caveat itself:

The usual disclaimer applies: correlation isn't causation.

The sample limitation most summaries omit

Two thresholds decided who was in the study:

• Domain Rating above 40.
• A highest-volume keyword with at least 800 monthly searches.

Those are thresholds on web presence, not on company size. Domain Rating measures backlink profile strength and the keyword threshold measures search demand for a term the domain ranks for. A two-person consultancy with an established domain can clear both, and a large company with a young domain may not.

What the sample does exclude is domains without an established link profile or measurable search demand, which correlates loosely with newness rather than headcount. The generalisability concern stands. The version we published before, which described the excluded group as small consultancies and boutique agencies, inferred company size from a web-presence metric and was wrong to.

The honest use of this study is to identify an association among domains with established web presence. It is not a formula for a five-person agency, and the same caution applies to the citation-share statistics used to sell digital PR.

Mentions and citations come apart

Semrush analysed 3,981 domain appearances from 115 prompts across four engines and 14 countries, published 9 June 2026:

• 61.7% were citations without a brand mention in the answer.
• 13.2% were both cited and mentioned.
• 25.1% were mentions without a citation.

The prompt count is small and the headline pools engines and countries, so treat the exact percentages as indicative. The structural finding is the useful part: most of the time, being a source and being named are different events. A page can support an answer and receive no visible credit for it.

That has a direct consequence for reporting. If a client asks whether AI visibility improved, "we were cited more" and "we were named more" are answers to different questions, and only one of them puts the business in front of a buyer.

What backlinks still do

Google says links help it discover pages and act as relevance signals, and that its generative features are rooted in core Search systems. That is enough to reject "backlinks are dead" without needing a study.

A relevant editorial link can still help crawlers discover a page, send referral visits, provide anchor and surrounding context, support ordinary search visibility, and place a brand mention inside independently published evidence. One good link often delivers all five at once.

Raw link volume is a poor objective. That is a different claim from links being irrelevant, and the two get conflated constantly.

Start with the problem, not the metric

The right mix usually includes both links and mentions, because a strong independent source tends to supply both at once. Google separately warns against chasing inauthentic mentions, which rules out the obvious shortcut of manufacturing forum conversations or buying list placements.

Track it with a fixed prompt set recording mentions, recommendations and citations as separate fields. If they share a column, the data cannot answer the question.

Limits

• Every study cited here is observational. None randomised anything, so none can establish causation.
• Ahrefs and Semrush both sell tools in this category. In each case the caveats we rely on are their own, which cuts against their products rather than for them.
• The 75,000-brand sample requires DR above 40 and a top keyword above 800 monthly searches. Those describe web presence, not company size, and we previously mischaracterised them as a proxy for it.
• The Semrush prompt set is 115 prompts pooled across engines and countries. Treat the split as indicative, not precise.
• Figures are dated December 2025 and June 2026. These systems change, and a correlation is a snapshot.
• We have not run a controlled test of mention-building against a holdout, and we have not found anyone who has.

The objective is not to win an argument between two metrics. It is to make the business easy to discover, easy to verify, and worth recommending.