Reference rate, not traffic: the AI Overviews metric

Organic clicks on queries with an AI Overview fell 61% between June 2024 and September 2025, then partly recovered in 2026 — but remain 38% below queries without an AI summary. The one thing that grew steadily throughout: the value of being CITED in the answer. Cited brands get more than twice the clicks per impression of uncited brands on the same results page. That is why the metric we track before traffic is the reference rate: of the questions that matter to you, in how many AI answers you appear as a source.

Below are the numbers — all from studies with published methodology, with their dates and their limits. Not from blogs quoting each other.

What happened to the click:Four independent measurements

Seer Interactive (3,119 informational queries, 42 organisations, 25.1 million organic impressions, June 2024 → September 2025): organic CTR on queries with an AI Overview fell from 1.76% to 0.61% — a 61% drop. On paid, from 19.70% to 6.34% (−68%).

Ahrefs (300,000 keywords, aggregated Search Console data, December 2023 vs December 2025): the presence of an AI Overview correlates with a 58% lower CTR for the page ranking first. In their original April 2025 study the figure was 34.5% — so the effect deepened during 2025.

Pew Research Center (68,879 real searches, 900 US adults, March 2025): users who saw an AI summary clicked a traditional result in 8% of visits, versus 15% for those without one. They clicked links inside the summary in 1% of visits. And 26% ended their session right after the page with a summary, versus 16% without.

Cloudflare (June 2025): the "crawl-to-refer" ratio — how many pages an AI engine reads for every visitor it sends back — reached, in Anthropic's case, nearly 71,000 pages read per single referral. Models read enormously and send back little. Traffic is no longer the currency; being the source is.

Four teams, four methods, the same direction.

The 2026 recovery — and why it is not "the end of SEO"

In April 2026, Seer published its update on 53 brands and 5.47 million queries: organic CTR on AI Overview queries bounced back to 2.36% in February 2026, 85% above the December 2025 low of 1.3%. It still sits 38% below queries without a summary.

The honest reading: the collapse was real and it stopped; what remains is a search engine where the summary answers first and the click is the second option. It is not the end of organic search. It is the change of the metric.

Citation is the lever — with numbers

Seer again, in the same datasets: brands cited in the AI Overview had, in Q3 2025, an organic CTR 35% higher than uncited ones (0.70% vs 0.52%); by February 2026 the gap had grown to +120% clicks per impression versus uncited brands on the same page.

Ahrefs, on 75,000 brands (May 2025): the factor most correlated with appearing in AI Overviews is brand mentions on the web (Spearman coefficient 0.664), ahead of links, domain authority or traffic. Brands in the top quarter for mentions appear in AI Overviews 169 times on average; the next quarter, 14 times. And 26% of brands do not appear at all.

Correlation, not causation — Ahrefs says so, and so do we. But the direction is clear: machines cite what is mentioned, defined and structured, not what has the most links.

What the reference rate is and how to measure it honestly

Our working definition: of N questions your customers ask, in how many AI answers your site appears as a cited source — measured periodically, on the same questions, with a date on every verdict.

How to measure it without lying to yourself:

• Through the APIs that return the answer's sources, not through screenshots from apps. Models with integrated search can return, for each answer, the addresses they relied on — so "was I cited for X?" has a verifiable answer.
• With the honest label on the figure: what you measure through an API is the model, not exactly what a person sees in the app on their phone (personalisation, location, version). It is said next to the figure, not hidden.
• With the same questions, at the same interval — otherwise you compare apples with pears. The variability of AI answers is real; a single probe is not a measurement.
• Without promising growth: the levers that "increase citation" (from the llms.txt file to presence on forums) do not yet have solid primary evidence. You measure the effect; you do not sell the recipe.

What the reference rate does not give you: Google does not publish a citation metric in Search Console. Bing introduced a "citation share" in its console in 2026 — in preview, sampled, without an API. The rest you measure yourself, with discipline.

What to do to be cited, not just indexed

There is no trick — there is a page built to be read by a machine looking for the answer, not by an algorithm looking for keywords:

1. The answer in the first sentence, not after three paragraphs of introduction — the way the articles on this site are written.
2. Clear entities: who you are, where you are, what you do — the same data on the site, in the schema and on profiles. The full guide is in AEO & GEO, and the generative-engine part in the GEO guide.
3. Schema that tells the machine what is on the page: Article, FAQPage, Organization, Person, with a declared bibliography — the schema markup guide.
4. Text visible without JavaScript: an AI extractor does not run your scripts. What appears only after animations or after a click does not exist for it.
5. Sources cited in the article, with a year. Machines prefer pages that show their sources — and so should you.

And if you are wondering what to do about the traffic that no longer comes, the answer is in the article on monetising in the zero-click era.

What these numbers do NOT say

• They are US and global data. Nobody has published a study on Romania. The direction is the same; the local magnitude we measure ourselves, on the sites we work on — and we publish it when we have enough data, not before.
• They are informational queries. On transactional and local queries, AI Overviews appear less often and the effect is different.
• Correlation is not causation. Brand mentions correlate with citation; nobody has shown that buying mentions produces it.
• The numbers have dates. The 2025 studies said "collapse"; the April 2026 one says "partial recovery". Whoever cites only 2025 sells fear; whoever cites only 2026 sells comfort. Reality includes both.