**TL;DR** — Across 21 client sites in April 2026 we measured AI Overview citation click-through: once Google's AI Overview cites your URL, how often does a user actually click through to your page? Our cohort median was 4.1% — far below traditional SERP CTR at equivalent visibility, and far below what most clients expect. Three properties of the citation moved CTR measurably — vertical position inside the answer card (above-the-fold citations clicked 3.6× more often than below-the-fold ones), the visible link label the composer chose (page-title labels beat domain-name labels by 2.2×), and the "answer completeness gap" between the AI Overview prose and the question (the more the prose answered the question on its own, the lower the CTR). Citation is not traffic; it is permission to be visible, and the conversion from visibility to visit is where the next tier of operational work now sits.
Why CTR-after-citation became the question this quarter
Through 2025, every team we worked with was trying to get cited. Through April 2026, every team we worked with started asking a different question: we are cited, and the traffic still is not arriving — why. The dashboards we built in 2025 stopped at "citation rate" and treated citation as if it were the deliverable. It is not. The deliverable is whatever the citation was supposed to produce — usually a visit, sometimes a brand-search lift two weeks later, occasionally a model-training exposure that pays back over months. Citation is a leading indicator of all three, but the conversion rate from "cited" to "visited" varies by an order of magnitude across our sample, and a citation that does not produce its downstream effect is operationally indistinguishable from no citation at all.
There is a second motivation: GSC and the AI-search-side analytics we wire into GA4 disagree by 30–60% on attributed visits, which makes the citation-to-visit gap impossible to read from any single tool. We needed a controlled measurement to know which gaps were instrumentation noise and which were real user behaviour. The answer turned out to be mostly the latter — the CTR after citation is genuinely lower than the SERP CTR most teams expect, and the variance inside it is operationally addressable.
How we ran the measurement
21 client sites — 8 SaaS, 6 publisher, 4 DTC, 3 B2B services — across April 2026. We instrumented every AI Overview citation that pointed at any client URL with a unique `?ref=aio` parameter (stripped before canonical normalisation, kept in GA4) so we could attribute referrer-less visits accurately. We then captured the citation event itself via Playwright twice daily across a 60-query basket per client, recording the citation's vertical position in the answer card, the visible link label, the answer prose itself, and the cited paragraph or section on the source page. CTR equals visits-with-`?ref=aio` divided by citation impressions, where a citation impression counts each unique (query, capture) pair the URL appeared in.
We dropped two cohorts. Pages cited only on queries with low absolute volume (fewer than 50 SERP impressions per week per GSC) were too sparse — the CTR ratio was unstable. We also dropped any citation that appeared in an "expanded" AI Overview state only, where the user had to click "show more" to see it; those have fundamentally different mechanics and pollute the headline number. The reported numbers below are for visible-on-load citations on commercial-intent queries with at least 50 weekly impressions, which is the population that produces the bulk of traffic for our clients.
The shape of the CTR distribution
Median citation CTR was 4.1%, with P25 at 1.6% and P75 at 8.3%. About 22% of cited (query, URL) pairs produced zero attributed visits across the entire month — they were cited daily and never clicked. The distribution is heavy-left: a small fraction of citations (around 14%) produced more than 15% CTR, and another small fraction produced essentially no traffic at all. Compared to position-1 organic SERP CTR (still in the high 20s on commercial queries in 2026), citation CTR is roughly 5–6× lower at equivalent visibility. The shape is also very different — SERP CTR is a tight distribution; citation CTR is bimodal, which means the operational story is about pushing your citations from the low cluster to the high cluster rather than nudging an average.
A small surprise: the same URL cited on different queries produced wildly different CTR — the median ratio between a URL's highest-CTR and lowest-CTR citation was 4.7×. So CTR is not a property of the page; it is a property of the (query, citation context) pair. A page that converts brilliantly on one query can be ignored on another even when the cited paragraph is identical, because the visible context the composer wraps around it changes the user's decision to click.
Driver one: vertical position inside the answer card
The single strongest predictor of CTR was where the citation appeared inside the answer card. Citations rendered in the first 60% of the visible answer (above the natural eye-line break, before the user scrolls) clicked at 6.4% median; citations rendered below that break clicked at 1.8%. That is a 3.6× difference for what is, structurally, the same citation. The composer chooses citation position based on which paragraph of which source it pulled from — first-paragraph citations on a source page tend to be rendered higher in the answer card, late-paragraph citations tend to be rendered lower. Practically: the paragraph the composer extracts from determines half of the visible-prominence delta, and most teams are not editing with this in mind.
We now treat the first 150 words of a commercial-intent page as load-bearing answer real estate. If the page's first 150 words read like a generic introduction ("In this article, we will explore..."), the composer either skips them or extracts later — either way the citation position drops. If the first 150 words contain a tight, claim-and-evidence answer to the most likely query the page targets, the composer extracts there, and the resulting citation lands above the fold of the answer card. The CTR delta from this single edit — rewriting the lead into a direct answer — was about +1.9 percentage points of CTR in our before/after sample of 84 pages.
Driver two: the visible link label the composer chooses
The composer renders each citation with a visible label — sometimes the page's `<title>`, sometimes its OpenGraph `og:title`, sometimes the domain name, and occasionally a generated label paraphrased from the cited paragraph itself. Citations rendered with a page-title or `og:title` label clicked at 5.4% median; citations rendered with just a domain-name label clicked at 2.4%; citations rendered with a generated label sat between at 3.9%. The label is the only thing the user reads before deciding whether the source is worth a click. A domain-name label tells them almost nothing about what the page actually says; a clear title tells them whether the page goes deeper than the AI Overview already has.
Which label the composer chooses appears to follow a confidence cascade: it prefers the page `<title>` when the title is short, descriptive, and distinct from the page's H1 in a way that suggests editorial care; it falls back to `og:title` when the `<title>` is generic or keyword-stuffed; it falls back to the domain name when neither title looks usable. Pages with `<title>` tags that read like the page itself — "How citation decay works in AI Overviews | Ranko" rather than "Ranko — Hong Kong SEO Studio" — were 2.1× more likely to be rendered with a title label. The work is title-tag hygiene, and the CTR payoff has materially increased in 2026.
Driver three: the answer-completeness gap
The third driver is the one most teams overlook because it cuts against intuition: the more completely the AI Overview answers the question on its own, the less likely a user is to click through. We measured "answer completeness" by hand-rating each captured AI Overview prose against the implicit question on a 1–5 scale, and CTR fell almost linearly with rated completeness — 7.8% on "partial" answers (rated 1–2), 4.2% on "mostly complete" answers (rated 3–4), 1.4% on "fully complete" answers (rated 5). When the AI Overview has already given the user what they came for, the citation is read as a footnote, not as an invitation.
This is an uncomfortable observation because it means optimising for citation rate and optimising for CTR partially trade off. The page that gets cited is the one the composer can extract a clean answer from; the cleaner the extraction, the more complete the AI Overview answer, the lower the downstream CTR. There is no clean way out of this, but there is a partial one: shape your content so that the cited paragraph answers the immediate question but raises one specific follow-up that the on-page version answers and the AI Overview cannot. Citations earned on pages with a strong follow-up hook clicked 2.7× more often than citations earned on pages whose cited paragraph closed the loop completely. The composer takes the answer; you reserve the follow-up for the on-page reader.
What changed in our content checklist
Four additions. We rewrite the first 150 words of every commercial-intent page to be a tight claim-and-evidence answer to the primary query, with the page's actual lead moved to the second paragraph. We audit and tighten `<title>` and `og:title` on every page receiving an AI Overview citation — short, descriptive, distinct from the H1, no brand-suffix padding. We test every cited paragraph against the implicit question and rewrite it where the AI Overview can fully resolve the question from the paragraph alone, building in at least one on-page-only follow-up. And we tag every AI Overview citation with `?ref=aio` so the GA4 attribution is unambiguous and the CTR can actually be measured.
We dropped one habit. Through 2025 we coached writers to "make the answer obvious in the first paragraph" — good advice for traditional SERP snippets, partially counter-productive for AI Overviews. The new instruction is "make the answer obvious in the first paragraph and reserve the depth for the second." The two-paragraph structure is the operating unit; the first feeds the composer, the second pays the reader for clicking through.
- 01Measure CTR per (query, URL) citation, not just citation rate. Median 4.1%, P25 1.6%, P75 8.3%, and ~22% of cited pairs produced zero visits across a full month in our cohort.
- 02Rewrite the first 150 words of every commercial-intent page into a tight claim-and-evidence answer. The composer extracts from the lead; above-the-fold citations click 3.6× more often than below-the-fold ones.
- 03Audit `<title>` and `og:title` on cited pages. Citations rendered with a title label click 2.2× more often than citations rendered with a domain-name label.
- 04Build a follow-up hook into every cited paragraph. Citations whose paragraph fully closed the question clicked at 1.4%; citations with an on-page-only follow-up clicked at 3.8%.
Where this argument breaks
For sites cited on fewer than about 25 (query, URL) pairs per month, the CTR sample is too noisy to act on at the per-citation level and the audit becomes a qualitative review rather than a measurement. For news publishers, the answer-completeness driver inverts — AI Overview users on breaking-news queries click through to confirm currency, not to learn the answer, and the completeness penalty disappears. In Chinese-language search the link-label mechanics are different — 文心 and 通义 render citations with their own conventions that do not map cleanly onto the title/OG hierarchy, and the audit needs separate basket runs. Outside those carve-outs, citation CTR is the next operational metric most teams should add to their AI-search dashboard, and it is the metric where the bulk of the citation-to-revenue gap actually lives.