Do Backlinks Still Matter for AI Citations? What 1.1 Billion Backlinks Say

Referring domains correlate with AI citations at 0.406 for Google AI Overviews and 0.377 for ChatGPT. That is across 486,003 and 255,413 domains respectively, significant at p < 0.001.

Those numbers come from DataForSEO’s Search Insider: 2026 Edition, which analysed around 1.14 billion backlink URLs against 137.9 million AI Overview answers and 27.5 million ChatGPT answers.

So the answer to “are links dead for AI search” is no. The answer to “will more links get me cited” is still no. Both of those can be true at once, and most of the argument happening about this collapses the two.

What the data actually shows

A correlation around 0.4 is moderate. It means link profile and citation volume move together often enough to be worth knowing about, and far too loosely to predict any individual outcome.

DataForSEO are careful about this throughout, and say plainly that the findings reflect associations rather than causes. Anyone quoting this report as proof that link building drives AI citations is quoting the opposite of what it says.

The finding I think matters most

Run the ranking analysis at domain level and at URL level, and the gap changes shape.

At domain level, sites ranking first had a median of 111 referring domains against 85 for positions four to ten. A 31% difference.

At URL level, first-place URLs had a median of 95 against 84 for positions four to ten. A 13% difference.

The same relationship, measured against the page instead of the site, is less than half as strong.

That tells you where the work pays. The authority that correlates with winning belongs to the domain, not to the individual page you have been pointing links at. Which matches what most practitioners have quietly concluded from their own work, and it is useful to see it separated out in a dataset this size.

It also reframes what a link campaign is for. Fifteen links to one commercial page is a weaker intervention than fifteen links to fifteen different pages on the same site, on this evidence.

Difficulty changes everything

The referring-domain profiles of sites actually cited in AI Overviews, split by how hard the keyword is:

A 5.6-fold difference between the easiest and hardest bands, and the spread widens as you go up. At KD 1 to 20 the typical cited domain sits in a narrow band between 236 and 447. At KD 81 to 100 the middle half runs from 518 all the way to 5,765.

For a specialist firm this is better news than it looks. The eye-watering numbers belong to the most competitive topics. If your buyers’ real questions sit at the easier end, which specialist questions usually do, the profile you need to be in the running is a couple of hundred referring domains rather than a couple of thousand.

Pick your questions properly and the link requirement drops by an order of magnitude. That is a strategy decision rather than a budget one.

Where this disagrees with everyone else

This is the part worth sitting with, because the published research does not agree with itself.

Ahrefs, across 75,000 brands, put referring domains at 0.295 and total backlinks at 0.218 for AI Overview visibility. Both lower than DataForSEO’s 0.406. In the same study, branded web mentions came in at 0.664, more than double any link metric.

SALT.agency, across 5,825 URLs, found backlink correlation with ChatGPT, Perplexity and Gemini visibility at 0.39 to 0.42, and with Google AI Overviews at only 0.25. Which is the reverse of DataForSEO’s ordering, where AI Overviews correlated more strongly than ChatGPT.

SE Ranking, across 129,000 domains, modelled it and found referring domains to be the strongest single predictor of ChatGPT citations, reporting that sites with 32,000 or more referring domains were 3.5 times more likely to be cited.

Semrush found nofollow links correlate almost as strongly as follow links, and that organic keyword coverage tracked AI visibility slightly more closely than total backlinks did.

Four studies, four different orderings. Most of those figures reach me through secondary write-ups rather than the primary reports, which is its own caution.

What survives the disagreement is a narrow band. Every study that has looked finds a positive, moderate association between link profile and AI citation, somewhere between 0.2 and 0.45, and every study that has measured brand mentions alongside links finds mentions correlating harder. Nobody has found links irrelevant. Nobody has found them decisive.

UGC links are the weaker half of a true story

Among domains with 10,000 or more UGC backlinks, 37.6% had 501 or more AI Overview citations. Among domains with none, 0.19% did. ChatGPT shows the same shape, 25.8% against 0.08%.

Those gaps are enormous and I would be cautious with them, because the base rates are tiny and domains with 10,000 UGC backlinks are not ordinary businesses. They are platforms.

The more useful comparison is the one in the same table. UGC backlinks correlate at 0.342 and 0.314. Referring domains overall correlate at 0.406 and 0.377. A varied profile tracks AI visibility more closely than a pile of forum links does.

Which is not an argument for ignoring Reddit, given it turns up in roughly a third of Google’s top ten results and is the second most-cited domain in both AI Overviews and ChatGPT. It is an argument for caring whether you are discussed there rather than whether you have links from there. Those are different jobs and only one of them can be bought, badly.

What I would actually do with this

Stop selling links as the AI citation lever. The honest version is that a reasonable link profile is the price of being considered a credible source, and after that something else decides. We set out the distinction in link building versus citation building.

Spread the links across the domain. The domain-level effect is more than twice the URL-level effect. Concentrating everything on one money page is optimising the weaker variable.

Choose easier questions. A 5.6x difference in required profile between easy and hard topics is the largest lever in this entire dataset, and it costs nothing but discipline.

Spend the remaining budget on being mentioned. Every study that measured mentions found them correlating harder than links. That means press, directories, review platforms, podcasts, forum presence, and being named by people who are not you.

Measure citations, not links. A link report proves activity. A citation report proves the thing you were trying to buy.

Our own approach to all of this is written up in the Synaptic Authority Engine, with pricing published rather than quoted case by case. If you want to know where your brand currently stands before deciding any of it, the AI Visibility Score is free.

The caveat

DataForSEO’s backlink dataset is global, which is unusual and welcome. The AI answer corpora behind it are United States English-language, so the citation side of every correlation here is American.

And none of it is causal. If you take one thing from this piece, take that correlations around 0.4 are interesting enough to plan around and nowhere near strong enough to promise anything on.