Why Google Clicks Are Disappearing: AI Overviews, Zero-Click, and CTR
Pew, Ahrefs, and academic research measure fewer clicks when AI Overviews appear. What the figures actually show and which SEO metrics matter now.
There used to be a golden rule: rank #1 on Google, pop the champagne, enjoy the traffic. These days, a top spot alone gets you crickets. During a Google visit with an AI summary, 8 percent of the users studied clicked a traditional result. Without an AI summary, the figure was 15 percent. The Pew Research Center observed the browsing behavior of 900 US adults in March 2025 for this analysis.1
These figures do not establish a global click-through rate or a forecast for every website. They do show a specific risk: if Google completes some of the search work directly on the results page, a strong organic position may deliver fewer visits.
Zero-click is not a single event. Not clicking a result, running another search, leaving Google, and ending a browsing session are different things. Pew measures them separately. In the US sample, the browsing session ended after 26 percent of Google visits with an AI summary, compared with 16 percent without one. A link within the AI summary was clicked in 1 percent of visits.1
From a user's perspective, this is often just convenient. If you want to know how warm it is in London right now, you don't want to click through to weather.com and accept a cookie banner. You want the number. Now.
Four studies, four different metrics
The percentages that are often cited at present cannot be combined into a single loss rate.
| Source | What was measured | Result in the respective dataset |
|---|---|---|
| Pew | Click and session behavior among 900 US adults, March 2025 | 8% clicked traditional results with an AI summary, 15% without one |
| Ahrefs | Position-one CTR across 300,000 desktop keywords, data through December 2025 | 1.57% for AIO keywords compared with a modeled 3.73% without an AIO |
| Khosravi/Yoganarasimhan | Daily traffic to English-language Wikipedia articles | About a 15% decline under AIO exposure in the setup studied |
| Similarweb | Referral conversion on transactional websites | Historical provider benchmark of 7% for generative-AI referrals and 5% for Google referrals |
Ahrefs compares 150,000 keywords with an AI Overview against 150,000 informational keywords without one. For position one, observed CTR in the AIO sample was 1.57 percent; without an AIO, Ahrefs modeled 3.73 percent. The resulting decline of about 58 percent is a correlation in this desktop dataset, not a general causal rate.2
The preprint by Khosravi and Yoganarasimhan uses a difference-in-differences design across 161,382 matched pairs of Wikipedia articles in different languages. For English-language articles, the authors estimate about 15 percent less daily traffic under AI Overview exposure.3 This is early evidence of a substitution effect for Wikipedia. It is not a CTR forecast for shops, banks, or local providers.
Similarweb’s 7-to-5-percent comparison applies to transactional websites and referral sessions. The statistics page linked in the old article no longer directly substantiates this specific figure; a later trend summary still cites it as a historical provider observation.4 It is useful for your own planning only as a hypothesis: fewer sessions may have a different composition, but they do not necessarily convert better.
Why users are not simply “switching to AI”
Let's be honest: as a user, you often find that convenient. Traditional search isn't going away, though. When it comes to health, money or a major purchase, and this is just my hunch, most of us still want to see a source we trust ourselves.
In a small qualitative study by the Nielsen Norman Group, nine people from North America and the United Kingdom used both traditional search and generative systems. AI relieved them of some comparison and sorting work. At the same time, no one relied on it completely.5
The cases described in the study report are vivid, but they remain individual examples. One participant started out shopping for soccer goal nets for his son: searching, clicking, jotting measurements on a sticky note, comparing prices. Tedious. Later, faced with a burst pipe, he tried Gemini instead and got an answer that fit exactly. No searching, no clicking, just a solution. “It feels like I saved time,” the study report quotes him as saying. Another participant wanted to bookmark Perplexity. Both are individual cases from a sample of nine people. This points to coexistence, not a linear switch for all users.
The Nielsen Norman Group’s Google AI Mode test was exploratory as well: seven people took part, and four had not previously noticed or used AI Mode.6 The test identifies specific usability problems, not general rejection of the product.
Content needs to give people a reason to click
A factual question that can be answered directly often needs no additional page. A click becomes more likely when the source contains something that cannot be reduced to a short answer. “5 Tips for Better Email Marketing” doesn't pull the way “How We Burned 10,000 Euros on a Single Email, and What It Taught Us” does. Specific cases with real numbers deliver what a short answer cannot:
- original, verifiable data or a reproducible test
- a specific case with its method, limitations, and result
- a tool or calculator that users can operate themselves, still underrated, still effective
- primary sources and documents for an important decision
- a next step that goes beyond the summary
First-hand experience is not a decorative first-person signal. It has to appear in verifiable details. A generic story about a supposedly expensive mistake is no better than a generic list of advice.
For general AI visibility, crawlability, and source selection, see the GEO guide. The article on Content Governance explains how regulated organizations can manage claims, evidence, and approvals. This article remains focused on clicks, traffic, and measurement.
Take control of what shows up under your brand
When someone searches your company name, they often see plenty already, without clicking. It's worth deciding for yourself what's there.
Clear FAQs, pricing pages and comparisons on your own site give Google a clean source to draw from. Without them, a summary can end up pulling from a less reliable source instead, a forum post, say, though not every answer draws on forums.
Writing your own “us versus the competition” piece also lets you set the criteria for that comparison, rather than leaving it to someone else.
Spread out, don't just stockpile
Keeping content locked away on your own blog gets riskier the more search work happens right on the results page. Two kinds of channels can round out your own site:
- owned channels like LinkedIn, YouTube and a newsletter, with full control and long-term brand value
- earned attention on Reddit, Quora or niche communities, with outsized impact when a post fits the place
How well any single channel performs against your own site depends heavily on topic, platform and timing. Blanket comparisons between formats aren't reliable enough to lean on.
Technical foundations remain necessary, but they are not an antidote
For generative search features, Google continues to recommend traditional SEO fundamentals: helpful content, crawlability, internal discoverability, and visible text.7 Google does not name schema, llms.txt, or prescribed chunking as additional requirements for AI Overviews. What these building blocks actually do, and where their limits lie, is covered in Schema, llms.txt and Structured Outputs.
A fast, well-linked page is useful for users and search. It does not prevent click loss if Google already provides the answer on the results page. Likewise, allowing an AI crawler does not guarantee a citation. Technical hygiene keeps the page in contention; it does not automatically create demand.
What reporting should include now
Rankings and clicks remain important, but they are no longer enough as the only diagnostic measures.
- Web search: impressions, position, CTR, and landing page in Search Console.
- Generative Google search: use the separate generative-AI report if it has been enabled for the property. Google has been rolling it out to some websites since June 2026.8
- Qualified visits: measure engagement and conversion by landing page and channel, not as an average across all referrals.
- Business impact: track leads, branded searches, and sales contacts separately. A later direct visit cannot be reliably attributed to an AI click.
- Study figures: never adopt them as targets for your own website. Your own time series is the relevant comparison.
Manual prompt sets can provide an additional snapshot of whether a brand is mentioned or a page is cited. They measure a particular model, date, language, region, and prompt, however. That is a different measurement from CTR.
What follows from the figures
AI Overviews can reduce clicks and traffic. The extent depends on the query, content, country, device, period, and website. Pew measures visits, Ahrefs measures keyword CTR, Similarweb measures referral conversion, and the Wikipedia preprint measures daily article traffic. Each source answers a different question.
For SEO, this means measuring click loss properly first. The content then needs a defensible reason for someone to open the source anyway. Another GEO promise is no substitute for either.
Your website will probably see fewer visitors. With a clear reason to click, they will hopefully be the right ones.