Brand Mentions vs Backlinks: What AI Studies Show

Before comparing anything, pin down the dependent variable, because the two most cited studies do not share one.

Ahrefs measured whether a brand gets mentioned inside an AI answer. Its researchers took 75,000 brands, pulled search metrics for each from Site Explorer, pulled AI mention counts from Brand Radar, and ran Spearman correlations between the two sets. The outcome being predicted is a brand name appearing in generated text.

Seer Interactive measured something close but not identical: whether a brand surfaced in answers generated by a specific model. Its team started from more than 300,000 finance and SaaS keywords, reduced roughly 600,000 People Also Ask questions down to 10,000 relevant ones, and ran those through the GPT-4o API, published in January 2025.

Neither study measured which URL got cited. That is a separate outcome with separate mechanics, and conflating the two is where a lot of downstream advice goes wrong. Getting your company name into a sentence and getting your page into the source list are different wins, and the same site can reliably earn one without the other.

So the honest framing of the whole debate is narrower than the headlines suggest. These are correlation studies about brand name visibility in generated answers, not experiments about citations, and not experiments at all.

Most coverage quoted two rows from this table. Here is the whole thing, from the Ahrefs analysis of AI Overview brand visibility factors published in May 2025.

Sample construction matters here, so note it: domains filtered to Domain Rating above 40, brands filtered to a highest-volume keyword of at least 800 monthly searches. About 74% of the brands in the set had at least one AI Overview mention and 26% had none.

Ahrefs later extended the same method across three surfaces, comparing ChatGPT, AI Mode and AI Overviews in a follow-up correlation study. The shape held, with one new entry at the top.

The authors are blunt about the limits of their own work: "correlation isn't causation. We've spotted patterns between search metrics and AI mentions, but that doesn't mean improving these metrics will automatically boost your AI visibility." They also warn that mass link building and programmatic content "doesn't seem to be the best course of action" on this evidence.

Seer's study of what drives brand mentions in AI answers approached the question from the ranking side and landed somewhere compatible.

Its strongest reported relationship was between first page search rankings and brand mentions in the model's answers, at roughly 0.65. Bing rankings tracked a little lower, in the 0.5 to 0.6 band. Backlinks were the surprise: Seer's team expected them to carry real weight and described their measured impact as weak or close to neutral. Content variety scored lower than the team anticipated too.

Read the two studies together and the useful signal is not "links are dead". It is that raw link counts predict AI brand visibility poorly, while things carrying your brand name predict it far better. Rankings, branded mentions, branded anchors and branded search all cluster above the generic link volume metrics.

The most repeated summary of the Ahrefs table runs some version of "brand mentions beat backlinks three to one". You can see how it was built: 0.664 divided by 0.218 is roughly three. That arithmetic is invalid, for a specific reason worth understanding.

Spearman correlation coefficients are not on a ratio scale. A coefficient of 0.664 does not represent three times as much of anything as 0.218 does. Correlations are bounded at 1, and the amount of variance a coefficient accounts for scales with its square, not with the value itself. Squaring both gives about 0.44 and about 0.05, which is a very different comparison from three to one, and still not a statement about importance or effect size.

Ahrefs itself put the whole table in perspective: "All the factors we studied revealed moderate to very weak correlations on the Spearman scale, but that doesn't necessarily mean they should be overlooked." The top factor is moderate. The bottom factors are weak. Nothing in the set is strong.

That distinction matters commercially, because the three-to-one framing has been used to justify shifting entire budgets out of link acquisition. The data supports "reweight and diversify". It does not support "stop".

Here is the finding that almost none of the coverage mentioned, and it sits two rows from the top of the same table everybody quoted.

Branded anchors scored 0.527, and a branded anchor is a backlink. It is a link pointing at your site whose anchor text contains your brand name. It is not a mention, not a citation, not a social post. It is link building, measured by an SEO tool, sitting comfortably above Domain Rating, referring domains and raw backlink count, and it sits above every other link metric on the list.

Line the link metrics up on their own and the pattern is clear:

So the split in the data is not mentions against links. It is branded signals against unbranded volume metrics. Branded web mentions (0.664) and branded anchors (0.527) sit at the top. Generic counts sit at the bottom. A link that names you behaves much more like a mention than like a link, at least as far as this correlation goes.

That reframing changes what you do on Monday morning. "Brand mentions beat backlinks" points a team toward unlinked PR placements. "Branded signals beat volume" points the same team toward earning fewer, better placements that actually say the company name, whether or not a link comes attached. The second instruction is supported by the table. The first one is not, quite.

There is a methodological reason to expect the link metrics in this dataset to underperform, and it is baked into the sample.

Both Ahrefs studies filtered to domains with Domain Rating above 40. Domain Rating is computed from the backlink profile. Filtering on it removes every site with a weak link profile from the sample before a single correlation is calculated.

This is textbook range restriction. When you cut the low end off a variable, you shrink its variance, and shrinking variance mechanically depresses any correlation that variable can produce. The remaining brands are all reasonably well linked, so link metrics have less room left to distinguish them.

Branded web mentions got no equivalent filter. That variable keeps its full spread across the sample, from brands nobody talks about to brands in the news weekly.

None of this makes the study wrong. Ahrefs states the filter openly, and studying established brands is a legitimate choice. It does mean the gap between the mention numbers and the link numbers is at least partly an artifact of who was allowed into the sample, and a study built on a wider Domain Rating range would very likely report stronger link correlations. Treat the 0.218 as "how much raw backlink count separates already well linked brands from each other", which is a much narrower claim than "backlinks barely matter".

Both studies predict brand name appearances. Citation behavior is measured separately, and it moves fast.

Ahrefs' analysis of AI Overview citations against the top 10 looked at 863,000 keyword SERPs and around 4 million AI Overview URLs. Roughly 37.9% of cited URLs also appeared in the first 10 organic results for the same query. About 31.2% came from positions 11 to 100, and about 31.0% came from outside the top 100 entirely. An earlier pass in July 2025 had put the top 10 overlap near 76%, so the share sourced directly from the visible results roughly halved.

Two practical consequences follow. First, a page well outside the first results page can still be cited, because fan out queries pull from SERPs you never see. Second, the metrics that predict brand mentions are not automatically the metrics that predict citation of a specific URL. Brand mentions look like a brand strength problem. Citations look much more like a retrieval problem, which is a question of whether the right passage on the right page is crawlable, extractable and quotable. We covered that mechanism in how AI answer engines retrieve chunks rather than pages, and the engine by engine differences in how AI assistants pick the sources they cite.

The row that resolves the argument for most teams is the fourth one. If your goal is brand visibility in generated answers, mention work and branded anchor work are the better bets on this evidence. If your goal is getting a particular page cited, retrieval and indexing work matter more than either, and links still do the discovery job nothing else does.

This audit takes about an hour and tells you which of the two problems you actually have. All figures below are illustrative examples, not SparkCliks data or study results.

Step 1: baseline your branded search demand. In Google Search Console, open Performance, then Search results. Set the date range to Last 28 days. Add a Query filter using Custom (regex) and enter your brand name plus its common misspellings, for example sparkcliks|spark cliks. Switch to the Queries tab and export. Record total branded impressions and clicks. Branded search volume was a mid table factor in both Ahrefs studies (0.352 to 0.466), so this is your cheapest available proxy for brand strength.

Step 2: baseline your citation surface. In the same Performance report, clear the query filter, go to the Pages tab, and export the top 100 pages by impressions. These are the pages that already have retrievable, indexed content. Anything not on this list is unlikely to be cited by an AI surface that reads the search index.

Step 3: measure assistant referrals. In Google Analytics 4, open Reports, then Acquisition, then Traffic acquisition. Change the dimension to Session source / medium and search for the assistant hosts one at a time: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com. Referral counts here are a floor, not a total, because AI Overviews traffic arrives as ordinary search traffic and many assistant clicks strip the referrer.

Step 4: compute the gap. Put the three numbers side by side. An example month might look like this:

Reverse the second and third numbers, with strong assistant referrals but almost no branded impressions, and you have the opposite problem: you get cited but nobody is looking for you by name.

Step 5: design the test honestly. Pick a 90 day baseline window and a 90 day change window. Choose 15 to 20 pages to work on and leave a matched control set of similar pages untouched, matched on impressions and topic. Read the change on the control set first. Whatever moved there is your noise floor, and only movement on the test set beyond that floor is worth interpreting. Expect this to be noisy: assistant referral volumes are small, seasonality is real, and models get updated without notice.

Run these against your own site before reallocating any budget.

1. Is your brand name spelled and formatted the same way across your site, your profiles and your press coverage? Inconsistent entity naming splits every signal in this article.
2. Do the placements you already earn actually say your brand name, or do they say "click here" and "this tool"? Branded anchors outscored every other link metric in the table above.
3. Do you know your current unlinked mention count at all? Most teams track referring domains and have no baseline for mentions.
4. Are your key pages crawlable by the assistant crawlers you care about, and readable as plain HTML without JavaScript execution?
5. Is your target outcome a brand mention or a URL citation? They need different work, and most plans quietly assume they are the same thing.
6. Have you set a control group, or are you about to read a seasonal bump as a win?

Say this part out loud before presenting any of it internally.

Every number in this article comes from a correlation study. Not one of them is an experiment, so none of them establishes that increasing a metric causes AI visibility to change. Both variables in every pair are plausibly downstream of one thing nobody isolated: how big and well known the brand already is. Large brands get mentioned more, get better anchors, get more branded searches and get into more AI answers, and a correlation table cannot separate those.

The studies are also snapshots of systems that change constantly. The AI Overview citation overlap moved from roughly 76% to roughly 38% between two measurements of the same metric. Treat any of these figures as current at the date of publication and check the source before quoting it.

One more thing, since this is a traffic company's blog and you deserve the plain version. Nothing SparkCliks sells produces brand mentions or AI citations. Our search click services deliver the clicks and settings you configure, visible in your own analytics, and that is the whole claim. Retrieval systems read crawled content, not visitor counts, and no assistant has access to your traffic numbers. Anyone selling AI citations as a deliverable is selling something they cannot control.