Citations Are the New Backlinks. The Data Finally Says So.

Last week I read every large study on AI search citations published since early 2025. Six of them, plus one controlled experiment. Going in, I expected marketing noise dressed up as research. Some of it was exactly that. But the pattern underneath held up better than I wanted it to, and it has changed how I plan link acquisition for clients.

So here is what the numbers say, and what I changed because of them.

The click economy broke first

None of this matters without the traffic collapse sitting behind it. In the US, 68.01% of Google searches ended without a click in the first four months of 2026, per SparkToro's clickstream analysis. Pew Research ran the cleanest version of this study, tracking 68,879 searches from 900 US adults. When an AI summary appeared, people clicked a traditional result 8% of the time. Without one, they clicked 15% of the time. Then comes the detail everyone skips over. Users clicked a link inside the AI summary in 1% of visits.

Meanwhile, Chartbeat data cited in the Reuters Institute's 2026 report showed Google traffic to publishers fell by a third globally in the year to November 2025. US publishers lost 38% of it.

Seer Interactive then split the surviving clicks into three tiers, across 5.47 million queries. With no AI Overview on the page, brands averaged a 3.3% click-through rate. When cited inside the AI Overview, that dropped to 2.1%. When present on the page but left out of the summary, it fell to 0.9%. So a citation is now worth roughly double the alternative.

So the thing we have spent a decade optimising for is shrinking under us. That is the backdrop for everything below.

Six studies, one uncomfortable pattern

Ahrefs studied 75,000 brands and ran correlations against AI Overview visibility. Branded web mentions scored 0.664. Backlinks scored 0.218. Roughly three to one, in favour of being talked about rather than linked to.

Then Muck Rack analysed more than 25 million links across ChatGPT, Claude and Gemini. Earned media accounted for 84% of all AI citations. Paid and advertorial content accounted for 0.3%. Not 3%. Zero point three.

Stacker and Scrunch ran the only proper experiment I found. They took 87 stories across 30 clients, held the content constant, and changed only the distribution. Stories pushed out to third-party outlets earned a 239% median lift in citations. Also, 97% of distributed stories earned at least one citation, against 82% for the same content sitting on a brand's own site.

Seer Interactive found something similar with review profiles. Brands with no Trustpilot presence had a 1% median citation rate. Brands with even a token profile, as few as one review, hit 53.5%.

The contradiction that turned out not to be one

Here is where I nearly binned the whole thing. SE Ranking studied 129,000 domains and found the opposite. Referring domains were the single strongest predictor of ChatGPT citations. Sites with more than 32,000 referring domains pulled 3.5 times the citations of sites under 200.

So which is it? Mentions or links?

Both, once you look at the methods. Ahrefs ran brand-level correlations against Google's AI Overviews and treated mentions as a separate signal. SE Ranking ran a machine learning model against ChatGPT citation counts and never tested mentions as an equivalent variable. Neither study kills the other. They measured different things on different platforms.

So here is the synthesis I landed on. Links are still how engines find and trust the third-party pages they cite. Mentions on those pages are what the model absorbs. Links are the delivery mechanism, and mentions are the payload. You need both, and the old habit of chasing one trophy link while ignoring twenty unlinked mentions now looks backwards.

What I changed

First, nofollow links and unlinked mentions became reportable wins. Previously I treated them as consolation prizes at the bottom of a client report. Now they sit in the same column as dofollow placements, because none of the citation data cares about the attribute.

Second, I moved outreach budget toward coverage breadth. Twenty solid trade placements beat two trophy links, since AI engines pull from a narrow set of third-party sources and repetition across them reads as consensus.

Third, I started fixing the pitch gap. Muck Rack found that the journalists PR teams pitch overlap only about 2% with the journalists AI cites. Two percent. That is the cheapest arbitrage in the discipline right now, and it takes an afternoon to fix. Pull the outlets and writers your target engines cite in your category, then pitch those people instead of the legacy media list.

Fourth, review platforms went on the standing checklist. Getting a client a basic Trustpilot profile is a two-hour job with an absurd return, given what Seer measured.

I am not the only one moving this way. BuzzStream's 2026 survey found 83.1% of digital PRs say the discipline matters more this year because of AI. Also, 75% of them have fielded a client question about AI citations in the last twelve months.

What I stopped doing

Then there is llms.txt, which is done. Ahrefs checked server logs across 137,000 domains and found 97% of llms.txt files received zero requests. Google has said plainly that it does not use the file for Search. So keep it for documentation sites feeding coding assistants, and drop it from client SEO scopes.

I also stopped promising click-based reporting on AI visibility. Semrush found 45% of marketing leaders cannot measure their brand's presence in AI answers, and only 9% have tools covering every relevant metric. Rather than fake precision, I now baseline citation share per platform and report movement against that baseline.

Where I stay sceptical

Nearly every study above was funded by a company selling AI visibility tools. Ahrefs, SE Ranking, Semrush, Muck Rack, Stacker and Seer all benefit from the story they are telling. Every correlation study also confounds mentions with plain brand strength, since big brands get mentioned and cited for the same underlying reason.

Then there is volatility. Reddit's share of ChatGPT citations reportedly collapsed from about 3.8% to 0.5% inside a single week last August. So any tactic tuned to today's citation pattern may decay fast.

The academic work runs more sober too. A July 2026 survey on arXiv documented that the famous 40% lift from the original GEO paper has been widely misquoted. It measured word count share among sources already retrieved, not a rise in citation rate.

So I treat these effect sizes as directional, never precise.

The part that survives

Strip out the vendor framing and one thing holds. Off-site presence now matters more than owned-content volume, and earned media is the channel that produces it. That was true before AI search arrived. It is just harder to ignore now that the traffic which papered over weak PR has gone.

If I had one line to hand a client, it would be this. Stop counting links, start counting mentions, and go get covered by the people the machines are already reading.