I first learned of this from a post on X by Mr. Suzuki of the Japan SEO Association, and the paper itself was published on arXiv in August 2026 by Athena Chapekis and colleagues. It analyses one month of browser logs from 900 U.S. adults, drawn from the Ipsos KnowledgePanel, covering March 1 to March 31, 2025. The sample is 91,121 visits to Google search pages and 68,879 distinct search queries.
The paper raises one methodological question. When your site is cited in an AI Overview — the AI summary at the top of the search results — how many clicks does it actually get?
About 1% is the real click rate
Let me put the numbers in front of you.
After visiting a page that carried an AI Overview, users clicked on a source cited inside that AI Overview in about 1% of visits.
For comparison, here are the same metrics on search pages that did not carry an AI Overview.
| AI Overview present | AI Overview absent | |
|---|---|---|
| Clicked a link in the search results | 8% | 15% |
| Ended the browsing session | 26% | 16% |
Where an AI Overview is present, clicks on search results fall from 15% to 8% — roughly in half. Instead, sessions end earlier. The user gets what they need from the AI answer and closes the browser.
These correlations held even after controlling for differences between individual panelists and for the query attributes that make AI Overviews more likely to appear. The paper looks at query length, whether the query starts with a question word, and whether it contains both a noun and a verb. After adjusting for those, the result does not move. This is observational work, not a causal claim. But the correlation is robust.
What makes AI Overviews likely to appear
The paper identifies three attributes of the query.
- The query is longer.
- The query starts with a question word (what / how / why, etc.).
- The query contains both a noun and a verb.
Put simply, the more a user is searching for "give me the answer directly," the more likely an AI Overview is to appear at the top of the page. Conversely, when the user types a company name, a product name, or a short imperative query like a URL, the AI Overview is less likely to appear. That difference goes straight to KPI design, as we will see next.
"Citation as traffic acquisition" is structurally a loss
This is the core of the problem.
A lot of teams treat "Am I being cited by AI search?" on the same footing as legacy organic traffic. Citation counts, cited-query counts, traffic change on the cited page. These are being watched side by side with normal organic KPIs.
But given that the click rate is 1%, that lens does not work by construction. Say a cited page appears in an AI Overview 100 times. It probably gets one click from that. That one click converting is, at best, rare. Cost the citation acquisition against revenue — content production, optimisation, tooling — and most of it comes out as a loss.
This is not because AI is eating the site. The KPI design is wrong. AI answer formats taking a deeper position at the top of the search results is a real shift, and I wrote about that pattern in what LLMO means as an evolution of SEO. The problem is that the KPIs built on top of that shift still assume click volume as the unit of value.
How you count the number decides the verdict of your programme
This is where the counting method makes the difference.
If you evaluate citation acquisition as a traffic acquisition initiative, the ROI will always look like a loss. If instead you evaluate it as a seeding mechanism for awareness, a different picture appears. I track these two separately.
Why split them? In most cases, the effect of citations shows up through a medium-to-long-term lift in brand awareness. A user sees your company or product name once inside an AI answer. Days later, they think "that was the company I was looking for," and go and search for it directly. That lagged effect will not be captured by same-day click counts.
Three things you can do now
-
Record a baseline of branded search traffic. In Search Console or GA4, record traffic that comes from your company name, product name, and brand terms, every month. Watch the trend around the months where citations rise. Causality cannot be proven, but the direction of the correlation is readable. Find the months where citations rose and where branded search rose, and lay them side by side on the calendar.
-
Design pages likely to be cited in AI Overviews as "awareness pages." Citation acquisition is not for one click. It is for one moment in a user's memory. Answer in the first line. Show the evidence. Make the company or product name unmistakably readable. Accept that the click rate is fixed at 1%, and make "one person seeing this once" the KPI. This is the direction I've also written about in AI-generated content in SEO: clarity and freshness are the differentiation levers.
-
In KPI meetings, separate "citations" and "traffic" into two rows. The first row: citation count, cited-query count. The second row: clicks, traffic, conversions. Mix them and citations will always look like a loss, and the strategy that should actually be pursued will become invisible.
Conclusion: one click vs one impression
When your site is cited in an AI Overview, the real click rate on the cited source is around 1%. As long as you count citations as a traffic KPI, the ROI will always read as a loss. That is a KPI problem, not an AI problem.
But a citation is more than a click. It is a moment of awareness in which a user sees your company or product name, even once, inside an AI answer. That moment returns, a few days or a few weeks later, as branded search or as an enquiry.
A bad KPI breaks outcomes faster than a bad strategy. Start by counting citations correctly as moments of awareness. Then track what branded search and enquiries do in the weeks after. That alone moves the evaluation closer to what actually happened.
One impression, once, remembered. That is the axis of content strategy in the age of AI search. My own thinking on where AI writing has taken us, and where it has not, is in what changed in AI writing in 2025.
