Do backlinks work on AI search? Two studies, two answers

Two studies, two answers, and nobody has closed the gap

Semrush took 1,000 randomly selected domains and measured them against ChatGPT, ChatGPT with Search, Gemini, Google AI Overviews and Perplexity. In the study published on 16 October 2025, site authority against AI mentions came out at 0.65 on Pearson and 0.57 on Spearman. Strong, by the standards of a field where most published correlations sit under 0.4. Their conclusion: "AI visibility comes from sustained authority, not one-off link wins."

Ahrefs took 75,000 brands, ran millions of AI responses through Brand Radar, and published on 12 December 2025. Domain Rating against brand mentions came out at 0.266 in ChatGPT, 0.285 in AI Mode and 0.326 in AI Overviews. Their conclusion, word for word: "Once again, we found very weak correlations between link metrics ('number of backlinks' and 'URL rating') and brand mentions across all AI systems."

Same question, same eight-week window, opposite verdicts. Both companies sell backlink tools, so neither has an obvious incentive to talk the answer down. Neither published raw data. If you have read a confident post about whether links move AI visibility, it was built on one of these two and almost certainly did not mention the other.

We are not going to pretend to adjudicate it, because the honest position is that it has not been adjudicated. Two different companies measured two different proprietary metrics against two different proprietary AI visibility datasets and got answers a factor of two apart. Nobody has run the experiment that would settle it: build links to a page, hold everything else still, and measure whether citations move. Until somebody does, everything below is correlation, including the part we find most convincing.

The nofollow number is the one that matters

Semrush split the link types, and this is the finding that got the least attention and deserved the most.

Follow links against AI mentions: 0.334 on Pearson, 0.504 on Spearman. Nofollow links: 0.340 and 0.509. Nofollow came out fractionally ahead.

A nofollow link is a link whose publisher has explicitly told search engines not to count it. Passing no ranking credit is the entire point of the attribute. If link equity were the mechanism moving AI visibility, the follow column would pull away from the nofollow column. It does not. It loses to it by a rounding error.

Semrush read that as models treating both link types alike when they assess authority. There is a simpler reading available. Nothing is flowing through the link at all. Both columns are measuring the same underlying thing, which is how many times a page somewhere on the web talked about you and pointed at you while it did. The href is incidental to what is being counted. It is just the easiest part to count.

Ahrefs arrives at the same place from the other direction. In their data the strongest correlate of AI visibility was not a link metric at all, it was YouTube mentions, and branded web mentions landed in the 0.66 to 0.71 range against 0.27 to 0.33 for Domain Rating. Their own summary of it: "brands with strong YouTube presence, widespread mentions, and branded anchors are the ones appearing in AI responses."

Read the two studies as one body of evidence and most of the disagreement dissolves. Mentions correlate. Links correlate to the degree that links are mentions. The purest link-graph metrics, the ones stripped of everything except who points at whom, correlate weakest of all. That is consistent with what we described in how to get cited by ChatGPT, where the unit of retrieval is a passage a model can quote, not a domain it has ranked.

Authority does not pay out in a straight line

There is a second detail in the Semrush numbers that most write-ups skipped, and it has direct consequences for how link building gets reported.

Pearson measures a straight-line relationship. Spearman only asks whether the ranking order holds. For links, Pearson sits around 0.33 while Spearman sits around 0.50. When Spearman runs that far ahead of Pearson, the relationship is not linear. It steps. Semrush put it plainly: "Big visibility gains only appear once a site reaches the higher authority tiers."

Which makes the standard monthly link report close to meaningless as a progress measure. Five links this month, seven next month, a chart going up and to the right, and none of it corresponds to anything visible until the profile crosses a tier boundary. We think that is why link building has the churn problem it has, though we are inferring the mechanism rather than measuring it: the campaign gets cancelled in month four for showing a flat outcome curve, at close to the point where the spend was about to matter.

If you are buying links, the number to hold your supplier to is not how many arrived. It is whether the referring-domain profile has moved band, and whether your mention volume moved with it.

The bridge from rankings to answers is real, and it is narrowing fast

There is one route by which links reach AI answers that nobody disputes, which is the ordinary one. Links help rankings, rankings feed the AI layer. The question is how much traffic that bridge still carries.

Ahrefs measured it twice. In July 2025, roughly 76% of pages cited in AI Overviews also ranked in the top 10. By the follow-up published 2 March 2026, across 863,000 keyword SERPs and 4 million AI Overview URLs, that figure was 38%. Of the rest, 31.2% ranked somewhere between position 11 and 100, and 31.0% did not rank in the top 100 for the query at all.

Ahrefs attribute the collapse to query fan-out, where the original question gets split into several sub-queries and citations are drawn from those source SERPs instead of the one the user typed. That is a mechanism, not a guess, and it matches what Google has said about how the feature works.

Two readings of the same number, and both are true. Ranking still buys you a real shot at citation, so the case for links is not dead. And that shot halved in eight months, so anyone modelling AI visibility as a downstream benefit of rankings is modelling a bridge that is being dismantled while they stand on it. Nearly a third of cited pages are winning citations from outside the top 100, which is a route that no amount of link authority explains.

What Google says, which is less than people quote

Google publishes its own guide to optimizing for generative AI search, last updated 10 July 2026. Links, backlinks, domain authority and off-site signals do not appear in it. Not downplayed, not qualified. Absent.

What it does say is that "the best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems." That is an argument from inheritance. It tells you the AI features sit on top of the ranking systems that links feed. It does not tell you that links are a factor in what gets cited, and Google has not said so anywhere we can find.

The same document is unusually direct about shortcuts. It says you do not need llms.txt files, you do not need to chunk your content, you do not need to write differently for AI, and structured data is not required. And it names the obvious exploit that falls straight out of everything above: "Seeking inauthentic 'mentions' across the web isn't as helpful as it might seem."

That last line closes the loop on the cynical version of this article. Yes, mentions correlate more strongly than links. No, you cannot go and manufacture ten thousand of them. Google names the tactic specifically, which means they are looking for it, and nobody outside the company knows how well they find it.

The only causal test anyone has run was about content

Every study above is observational. There is one piece of work in this field that ran a genuine experiment, and it is worth knowing what it found even though it is about the other half of the problem.

A team from Princeton, IIT Delhi and elsewhere built GEO-Bench, 10,000 queries with the top five search results for each, and tested content interventions against generative engines. The paper was presented at KDD 2024. Adding quotations from credible sources improved visibility by up to 40%. Adding statistics improved it by around 30%. Citing sources improved it by around 30%. Keyword stuffing performed worse than doing nothing. On Perplexity, in live conditions, they measured improvements up to 37%.

Two things about that. It is from 2024 and predates every engine version currently shipping, so treat the percentages as a direction rather than a forecast. And every intervention that worked was the same intervention wearing three hats: put verifiable, quotable, attributable material on the page. None of them involved a link pointing at you.

What this changes about buying link building

The uncomfortable version first. If you are buying link building specifically to move your visibility in AI answers, you are buying an instrument that the published evidence supports weakly, indirectly, and only above a threshold nobody has located. Anyone selling you links on an AI visibility promise is ahead of the data.

The case for buying it anyway is real, and it rests on different ground. Rankings still convert into citations 38% of the time in AI Overviews, which is a worse bridge than it was and still the best-documented one there is. And the work that earns a link is the same work that earns a mention: a real pitch to a real editor about something worth writing up. You are paying for the outreach motion. The link is the receipt, and the mention is the part that correlates.

Four things we would change about how the work gets specified.

Prospect by citation, not by Domain Rating. Run a fixed prompt set for your category, record which domains the engines actually cite back, and target those specifically. A link from a site that gets quoted in the answers your buyers see is worth more than a link from a higher-DR site that never appears. Almost nobody prospects this way, which is exactly why it is available.

Count mentions, not just links. If unlinked brand mentions correlate more strongly than referring domains, then a campaign that produced eleven mentions and four links reported four. Track both, or you are grading the campaign on its weaker output. That is the baseline we set in measuring AI visibility.

Make sure the entity is unambiguous before you spend on volume. A model that cannot reliably tell your company apart from a similarly named one cannot attribute a mention to you at all, and every mention you earn until you fix it is being credited to a fog. That is the argument in entity SEO, and it is cheap work compared with an outreach retainer.

Give the target page something worth quoting. The one causal study in the field says statistics, quotations and cited sources move citation rates by roughly a third. A page with nothing extractable on it is a page an engine has no reason to lift from, however many links point at it.

That is how our Link Building Subscription is scoped, and it is why the qualification bar is set on topical relevance and real traffic rather than on a Domain Rating threshold. It also constrains volume, honestly. Outreach that earns a mention worth having does not run at forty sends a day.

A live example on a domain you do not own

A LinkedIn-specific study published in August 2026 found the same pattern holds inside the platform itself. Among 645 LinkedIn feed posts it tracked, those with at least one backlink ranked in Google's top 20 31.2% of the time. Posts with none managed 6.7%.

The mechanism is simple: a link gives Google (and, to a lesser extent, Perplexity) a path to find and assess a URL that would otherwise just sit on the platform. We are testing that mechanism on our own LinkedIn piece here: answer engine optimisation agencies are being outranked by $6 LinkedIn posts. One inbound link from one site is a small test, not a repeat of the study at scale, treat this as an experiment we're running in the open, not a guarantee.

What we cannot tell you

The two correlation studies are vendor research. Both companies sell tools that measure the metrics they found to be predictive, both used proprietary AI visibility datasets that cannot be independently reconstructed, and neither released raw data. That does not make them wrong, and both disclose their methods more openly than most of what circulates on this topic. It does mean no independent party has replicated either.

Ahrefs state the caveat themselves, and we will let them say it: "The usual disclaimer applies: 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."

We also cannot tell you why the two authority figures diverge as far as they do. Authority Score and Domain Rating are different measurements built by different companies to different recipes, and the composition of each is the obvious place to look, but we could not read Semrush's current documentation on how Authority Score is calculated while writing this. So we are leaving the gap where it is instead of filling it with a plausible story.

Finally, correlation studies at this scale describe the population, not your site. A 0.5 correlation across 1,000 domains tells you the tendency. It does not tell you that your next twenty links will do anything. Measure your own before and after, on a frozen prompt set, or you are running on somebody else's average.

Sources

The 0.65 and 0.57 authority figures, the follow and nofollow split, and the threshold reading come from Semrush's Do Backlinks Still Matter in AI Search? by Margarita Loktionova, published 16 October 2025, covering 1,000 randomly selected domains across ChatGPT, ChatGPT with Search, Gemini, Google AI Overviews and Perplexity. The Domain Rating correlations, the branded mention comparison and the correlation-is-not-causation quote come from Ahrefs' Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews by Louise Linehan and Xibeijia Guan, published 12 December 2025, covering 75,000 brands. The 38% and the position breakdown come from Ahrefs' update on AI Overview citations and the top 10, published 2 March 2026, across 863,000 keyword SERPs and 4 million AI Overview URLs, which revises their own July 2025 figure of roughly 76%. Google's position, and the quotes on inauthentic mentions and on SEO best practices, come from Optimizing for generative AI search in Google Search Central documentation, last updated 10 July 2026. The content intervention results come from GEO: Generative Engine Optimization by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD 2024, using a 10,000 query benchmark.

Common questions