The citation is the conversion

There's a version of this argument that gets made badly, and it goes roughly: clicks are dying, so stop measuring clicks, and start measuring whether the machine says your name. It's a convenient argument if you happen to sell a tool that counts how often the machine says your name. The more useful question is narrower. Does a citation you never get a click from do anything measurable at all? There is now enough evidence to answer that, and the answer has a shape worth understanding before you rebuild a dashboard around it.

Start with the people who aren't selling anything

Pew Research Center ran a metered browsing study across 900 US adults for the whole of March 2025, capturing 68,879 unique Google searches, of which 12,593 produced an AI summary. It is the only dataset in this area with a nonpartisan, non-commercial publisher and a fully documented method.

With an AI summary8%

Clicked a search result

Clicked a link inside the summary

Ended the browsing session

18% of searches produced an AI summary during the study month.

Without one15%

Clicked a search result

No equivalent surface

Ended the browsing session

Nearly twice the click rate, on the same panel, same month.

Eight per cent against fifteen. And the link inside the summary — the citation itself, the thing everyone is now optimizing for — got clicked on one per cent of visits. Ahrefs reached a compatible number from Search Console data across 300,000 keywords, split evenly between those with an AI Overview and informational keywords without one. Using the non-AIO group as a counterfactual, they put the position-one CTR reduction at 58% as of December 2025, up from the 34.5% they measured in April of that year. The effect decays down the page: 50.8% at position two, 32.6% at position five, 19.4% at position ten.

What Google said, and what it conspicuously didn't

In August 2025 Google's Head of Search published a response. The load-bearing sentences: "total organic click volume from Google Search to websites has been relatively stable year-over-year," and "average click quality has increased and we're actually sending slightly more quality clicks to websites than a year ago." Third-party reports suggesting otherwise were characterized as based on flawed methodologies and isolated examples. The post contains no figures. No charts. "Quality click" is introduced as a term and never defined beyond a parenthetical about users not bouncing back quickly. Practitioners noticed. Authoritas' CEO pointed out that a company with the data would show the data. A former FT strategist observed the post claimed to share insights and contained none. Amsive's Lily Ray asked whether the hundreds of thousands of Search Console screenshots showing flat impressions against falling clicks counted as flawed methodology. Google is the only party that can settle this and it has chosen to settle it with an adjective.

Being cited beats not being cited. Probably.

Seer Interactive has the best client-side dataset on this — 53 brands, 5.47 million tracked queries, 2.43 billion organic impressions running from January 2025 through February 2026, joining Search Console and Google Ads data against AI Overview presence. Their finding is the strongest available case for citation-as-outcome. Being cited in the Overview delivered 120% more organic clicks per impression than not being cited on the same AIO-present query. An earlier cut of the same programme, covering 42 organizations and 3,119 search terms through September 2025, put it at 35% more organic clicks and 91% more paid clicks. The Q3 2025 numbers underneath that are worth seeing side by side. AI Overview present and not cited: 0.52% organic CTR. Present and cited: 0.70%. No AI Overview at all: 1.45%. And Seer, to its considerable credit, flags the confound itself rather than waiting for someone else to: it cannot prove citation causes the higher CTR, because it's equally possible that brands with stronger authority and higher baseline click rates are simply more likely to get cited in the first place. That caveat is the single most important sentence in the entire literature on this question, and it appears in the vendor's own write-up.

The part your analytics will never show you

Here's where it gets genuinely interesting, because the click studies all share a blind spot: they can only see people who clicked. If a citation does its work by lodging a name in someone's head who then goes looking three days later, no attribution model on earth records it. A June 2026 preprint attempts to measure exactly that, joining an opt-in clickstream panel to the same panellists' actual conversations with ChatGPT, Claude and Gemini. When an assistant recommended a brand to a user with no recent observed engagement with it, that user subsequently searched for the brand 4.3 percentage points more often, visited its site 2.4 points more often, and reached it at a retailer 1.0 points more often — measured against a within-user backward placebo. The authors note that standard web analytics attributes almost none of this to the assistant. The methodological contribution is the debunking, though. Pool every brand mention together and you get a larger, cleaner-looking funnel of 2.08, 2.37 and 0.52 points. The authors reject their own number. Pooling averages genuine recommendations with incidental name-drops — "you can use Samsung Pay," "your Netflix download" — where the downstream visits are existing customers doing what they were already doing. An event study shows activity on the named brand's site was already rising before the AI response. Incidental mentions move behaviour two to three times less than real recommendations. A vendor correcting its own headline figure downward is not vendor behaviour, which is why the study is worth taking seriously even though the vendor in question sells AI brand visibility software and declines to publish its sample size.

Now the ceiling

None of the above should be read as citation replacing the click, and the absolute numbers make that unarguable. Press Gazette's August 2026 analysis of the fifty biggest US news sites found combined monthly visits down 37% since July 2024, from a 4.9 billion peak to 3.1 billion, with organic search referrals down 42% over two years. CNN's search traffic fell 61%. And generative AI, across those same fifty sites, accounted for an average of 0.4% of total traffic. At the New York Times it was 0.2%. Only Reuters and The Hill cleared one per cent. SparkToro's clickstream analysis puts US searches ending without a click at 68.01% for the first four months of 2026, against 60.45% in 2024 — though the author is careful to note the historical points come from three different panels and the comparison is partly apples to oranges. Cloudflare's network data makes the exchange rate explicit. Between January and July 2025, OpenAI's crawlers made roughly 1,091 page requests for every single referral sent back. Anthropic's ratio improved dramatically over that period, from 286,930 to 38,066 crawls per referral, after Claude added web search with clickable citations. Google's sat at 5.4. Cloudflare flags its own blind spot: native app traffic arrives without a referrer header, so these ratios likely overstate the gap by an unknown amount.

What this actually changes

The strategic conclusion is real but it's a modest one, and it's worth stating without the drama. Citation is now the best outcome available on a query where an AI Overview appears, it correlates with meaningfully better click-through than absence, and it appears to produce off-platform brand-seeking behaviour that no analytics package will ever credit to it. Those three things justify measuring it. They do not justify abandoning click measurement, because AI referrals remain a rounding error in absolute volume. They do not justify treating citation share as a revenue proxy, because nobody has demonstrated the causal arrow. And they certainly don't justify a tool that reports your rank inside an AI answer as though such a thing were stable enough to have a position. The honest framing is that you have acquired a second, earlier goal that sits upstream of the click and cannot substitute for it. Presence in the answer is a precondition for being considered. The click, where it still exists, is what pays for the work.

Sources

Related reading

• What actually gets a page quoted in an AI Overview — the structural work that earns the citation in the first place.
• Not every prompt sends an AI to the web — which questions create a citation opportunity at all.
• Retrievable, extractable, citable, reusable — the mechanics that decide whether a machine can use what you wrote.
• Live answer-surface data — tracking citations across AI Overviews, PAA and the assistants.