One-line takeaway
AI Overviews did not take a share of traffic. They turned clicking itself into minority behavior: searches that touch no outbound link went from 54% to 73%.
Three datasets, three levels of evidence
Much of the discussion about AI Overviews and traffic treats weak and strong evidence alike. Three bodies are worth keeping separate.
A randomized controlled trial, the strongest for causality. Saharsh Agarwal and Ananya Sen grouped US desktop Chrome users between January and February 2026: one group saw AI Overviews, one had them hidden, and behavior was compared across the two. Random assignment means other conditions are approximately equal, so the difference is attributable to the overviews themselves.
A large keyword panel. DataForSEO covers 20 million keywords. It is good at describing what happened, not at proving why.
A difference-in-differences study of Wikipedia traffic in arXiv paper 2602.18455, using relative changes across language editions as a natural experiment.
The three point the same way, but their scopes differ: the trial measures user behavior, the panel measures result-page composition, the paper measures the actual traffic of one very large site.
The core numbers
| Indicator | Control (overviews hidden) | Treatment (overviews shown) | Change |
|---|---|---|---|
| Outbound clicks per search | 0.62 | 0.37 | −39.8% |
| Searches with no outbound click | 54% | 73% | +19 points |
The same trial found ad clicks were largely unaffected. The authors estimate that if overviews keep expanding, sites stand to lose roughly 18.5% of organic traffic overall.
Note that 39.8% and 18.5% differ: the first is the click drop on searches that show overviews, the second folds in coverage to estimate the effect on a site's total. They are not the same measure.
Coverage: from edge feature to default
| Date | AI Overview coverage |
|---|---|
| September 2024 | 5.93% |
| August 2026 | 61.43% |
Other result types were squeezed over the same period:
| Result type | Sept 2024 | Aug 2026 |
|---|---|---|
| Featured snippet | 8.96% | 0.44% |
| Forums block | 21.13% | 7.42% |
| Video block | 9.67% | 19.98% |
Featured snippets falling from nearly 9% to under half a point is the most telling signal. They were the original "answer extracted and shown separately" format, and AI Overviews now occupy that role.
Click-through: rank one is no longer safe
| Indicator | Sept 2025 | Aug 2026 | Change |
|---|---|---|---|
| Desktop top organic CTR | 17.83% | 7.43% | −58% |
| Mobile top organic CTR | — | — | −17% |
The desktop drop is far larger than mobile, consistent with the trial's desktop Chrome setting: desktop is where overviews display most fully and where click loss is clearest.
For a content site the implication is direct: ranking first no longer means capturing a fifth of clicks. The 17.83% era assumed users had to click a link to get an answer.
Intent mix is shifting
| Search intent | Sept 2024 coverage | Aug 2026 coverage |
|---|---|---|
| Informational | 48.42% | 70.44% |
| Commercial | 26.30% | 51.35% |
| Navigational | 17.04% | 34.03% |
| Transactional | 10.82% | 28.64% |
By vertical, health reaches 69.68% and finance 68.80%, while real estate sits at 22.27%.
Commercial queries crossing 50% matters more than informational. Comparison, selection and review pages — the ones closest to conversion — are also the ones now most heavily summarized.
Cited is not visited
One easily missed figure: as of March 2026, AI assistants accounted for just 0.3% of total website traffic.
That sits in sharp contrast to "the AI Overview cited me." A citation is brand exposure, not a visit. The overview hands over the answer, and readers often do not click through.
Wikipedia shows the scale effect: the arXiv paper estimates AI Overviews reduced English Wikipedia's search traffic by 5.45% relative to German and 4.82% relative to French, roughly 100.27 million visits per month. Wikipedia is among the most-cited sites on the web; ordinary content sites are unlikely to fare better.
So what should GEO be
Given the data above, the available moves are few but concrete.
Write pages as extractable answers. An overview needs a passage that stands on its own: a clear conclusion up front, with specific figures, dates and definitions. Vague openings and rhetorical throat-clearing get skipped.
Prefer structured facts to narrative. Tables, numbers with units and dates, explicit subjects and actions are easier to cite than long paragraphs. This is not for crawlers but so a summarizing model can extract safely — misreading one figure costs more than omitting a sentence.
Track citations, not only rankings. When CTR falls across the board, rank monitoring misleads. Useful substitutes include branded search volume, direct-traffic share, AI-assistant referrals (identified separately in logs), and how often your content is mentioned in answers.
On November 12, 2026, Google updated its search documentation with a section on identifying traffic from Google Search, covering robots meta, data-nosnippet, X-Robots-Tag and generative AI controls. It also added guidance on evaluating SEO advice and tools to the "Do you need SEO?" page. Both updates point the same way: make "which traffic came from search" and "which content may be used for generation" configurable.
Separately, on May 20, 2026, Google clarified in its documentation updates that AEO and GEO do not constitute new ranking systems. Marketing claims about "making AI like you" therefore lack official support.
Three realistic judgments
Informational traffic keeps shrinking, but not to zero. 73% of searches touch no link; 27% still do. That 27% skews toward queries needing depth, comparison and judgment — exactly what a content site can still hold.
Commercial pages need new entry design. A selection page offering only "which is better" gets absorbed by the overview. One offering verifiable comparison tables, failure conditions and scenario differences only loses part of its value, and the rest still requires a click.
Do not buy "guaranteed GEO ranking" services. No official documentation acknowledges a GEO-specific ranking mechanism. What can be verified is the old set: content quality, structure, crawlability and link signals.
One more practical step: build a separate channel for AI referrals. Split by referrer and user agent in your logs to see which pages get cited and which passage gets used. Without that split you see "organic traffic is down" without knowing which query class fell.
Boundaries
- The click figures, no-click share, unaffected ad clicks and 18.5% estimate come from Agarwal and Sen's study and secondary coverage; the full paper was not reviewed here, and sample size and confidence intervals are not disclosed;
- DataForSEO figures come from a 20-million-keyword panel and are descriptive, not causal;
- The Wikipedia decline comes from arXiv preprint 2602.18455, not yet through formal peer review;
- The 0.3% figure is a March 2026 industry statistic for AI-assistant share of total site traffic, not identical to AI Overview referrals;
- Google documentation details come from the official Chinese-language update log;
- This article contains no method for manipulating AI Overviews, fabricating citations or evading crawling.
Sources
- RS Web Solutions: randomized trial finds AI Overviews cut outbound clicks by about 40%
- Ecosistema Startup: AI Overviews cut clicks by 40%, what sites should do
- Pithwire: Google's AI answers cut search traffic by about 40%
- arXiv: the effect of AI Overviews on Wikipedia traffic (2602.18455)
- Google Search documentation updates (Chinese)
- Pacific Tech: Google's ~90% search share and AI Overview impact
