This page is not meant to be a statistics dump but something usable: every figure carries its publisher, sample and date, and where studies contradict each other, that is said out loud. Numbers about AI search tend to wander: they get quoted without a source, lose their definition along the way, and six months later mean something quite different from what the original study measured.
One rule for the whole page: if you quote a number, quote the description of the measurement with it. Without that it isn’t data, only mood. aiseo42 applies the same rule to client reports — those state what was measured, when, and how.
How many Google searches end without a click?
| Figure | What exactly it measures | Source, sample, date |
|---|---|---|
| 68.01% | Share of US Google searches that sent no click anywhere (January–April 2026), on Similarweb’s desktop and mobile panel | SparkToro & Similarweb, June 2026 |
| 60.45% → 68.01% | The same metric from 2024 to 2026: a 7.56-point rise, the steepest two-year move since measurement began | SparkToro, June 2026 |
| 276 clicks / 1,000 searches | What now reaches the open web, down from 374 in 2024 | SparkToro, June 2026 |
| 22.4% | Stricter definition: desktop only, complete absence of a click (US) | Datos/SparkToro-based aggregate, 2026 |
The differences between these figures are not measurement error: the first looks at all devices and all click types (including clicks to Google’s own surfaces such as Maps and YouTube), the last only at complete click absence on desktop. Anyone quoting “zero-click is 68%” without the definition is selling a methodological choice as a fact.
SparkToro itself states the limitation: the historical comparison mixes three different data sources (2016 and 2019: Jumpshot, 2024: Datos, 2026: Similarweb), so these are not the same users or devices.
It also varies sharply by intent (SparkToro, June 2026): 31% for transactional queries, 46% for commercial investigation, but 74% for informational ones. That applies to your number too: the intent mix of your industry decides how much this affects you.
Takeaway: a substantial share of searches — a third or two thirds, depending on methodology — never reaches a website. The direction is the same in every study.
How often do Google AI Overviews appear?
The most detailed public time series is Semrush’s December 2025 report covering 10 million keywords:
- January 2025: 6.49% of keywords triggered an AI Overview.
- July 2025: 24.61% — the measured peak.
- November 2025: 15.69% — Google pulled back.
But this figure is sample-dependent, and that is exactly what makes it instructive. Nearly 60% of Semrush’s keyword set gets under 100 searches a month, so it leans towards the long tail. For the same period the commercially focused BrightEdge measured 48% across nine industries, Conductor measured 25.11% across 21.9 million queries in Q1 2026, and Google itself referred to roughly half of US queries.
Takeaway: not a monotonically rising curve, and there is no single “true” appearance rate. Anyone quoting a single peak value as a constant (the figure circulating most widely online is 42%, with no named measurement behind it) is selling a snapshot as a trend. aiseo42 removed that 42% figure from its own site because we could not find a verifiable measurement behind it — even though the number happens to be our brand name.
How much click-through is lost when an AI answer sits above?
| Figure | What it measures | Source, sample, date |
|---|---|---|
| 15% → 8% | Results-page click-through without and with an AI answer (≈47% relative drop). Only 1% of visits click a link inside the AI answer itself | Pew Research, 900 adults, 68,879 searches, March 2025; published July 2025 |
| −58% | Clicks to the top organic result when an AI Overview appears. Eight months earlier the same study measured −34.5% | Ahrefs, 300,000 keywords (150k with AIO, 150k without), Dec 2023 vs Dec 2025 |
| −61% | Organic click-through decline alongside AI Overviews; a cited source, however, received 35% more organic clicks than a non-cited one | Seer Interactive, 3,119 queries, 42 organisations, 25.1M impressions, June 2024 – September 2025 |
| −41% | Per the same study, click-through fell even on queries without an AI answer | Seer Interactive, September 2025 |
And one contradictory figure, which matters precisely because it contradicts: in the same report Semrush also measured what happens to the same keywords before and after AI Overviews appeared — there the zero-click rate fell from 33.75% to 31.53%. The explanation is methodological: asking “which queries get an AI answer” (typically informational, already click-poor queries) is a different question from “what happens to this same query before and after”.
Takeaway: AI answers do not affect every query equally, and the direction of the effect depends on how the study is built. That is why aiseo42 starts every engagement with a month zero measured on your own keywords — there is nothing to compare an industry average to.
How do B2B buyers research today?
| Figure | What it measures | Source, sample, date |
|---|---|---|
| 51% | Share of B2B software buyers who start vendor research in an AI chatbot, more often than with Google — up from 29% in 2025 | G2 “The Answer Economy”, 1,076 buyers, March 2026 |
| 94% | Share who used AI during their most recent purchase process | Forrester, “The State of Business Buying, 2026”, nearly 18,000 business buyers |
| 69% | Share who chose a different vendor than originally planned, on the AI chatbot’s guidance; a third bought from a vendor they had never heard of | G2, 1,076 buyers, March 2026 |
| 83% | Share who use AI tools in their work | Semrush, 622 valid responses from US B2B professionals, March–April 2026 |
The most important line in the G2 data is not the 51% but the 69%: the question is not whether buyers start with AI, but whether they end up saying a different name because of it.
Takeaway: your buyers ask AI about you whether or not you work on AI visibility. The only open question is what it answers — and that is exactly what the free aiseo42 check establishes.
Do buyers actually trust what AI tells them?
Trust in AI is where inflated numbers get quoted most often, so aiseo42 puts three measurements side by side, each with its sample:
- 75% — fully or mostly trust AI vendor recommendations (Semrush, 519 AI-using US B2B professionals, March–April 2026). Conflict of interest: Semrush sells AI visibility products, so this is a vendor study.
- 39% — trust in AI chatbots among 1,200 US business decision-makers (Reddit + SurveyMonkey, 2026).
- 94% — of those who used AI in the buying process, this share fact-checks at least some of the time what the AI told them (TrustRadius, 1,862 buyers and 444 vendors, January 2026; published July).
The same Semrush study also shows what happens after a recommendation: 71% visit the vendor’s website, 63% search for the company on Google, 46% compare it against alternatives. The AI mention does not close the decision — it starts a focused verification.
One figure deserves separate attention: per Semrush, only 7% of respondents notice a vendor in an AI answer because they recognise the name. What makes a vendor stand out is how precisely the description matches the buyer’s actual situation.
Takeaway: the AI answer is not the decision, it is the shortlist. Those who make it get verified; those never mentioned never get looked up. And because brand recognition barely counts, that shortlist is open to small and new players too — provided there is precise, citable text about them.
Is there a measurement of what makes a source get cited?
Yes, and it is the most important controlled experiment in the field. The KDD 2024 GEO study (Aggarwal et al., DOI: 10.1145/3637528.3671900) measured, on a benchmark of 10,000 real user queries, which textual interventions increase a source’s appearance in the generated answer.
It has two results, and both cut against established SEO reflexes:
- Adding statistics, quotations and citations raised visibility by up to 40% in the generated answer (KDD 2024, 10,000 queries). The effect varied by topic: statistics helped most in “Law & Government”, quotations in “People & Society” and “History”.
- Raising keyword density had effectively no effect, and in some settings made things worse.
This measurement is why this page is built the way it is — and why aiseo42 insists that every claim in a client’s text carries the number, the source and the date. That is not a style preference; it is the consequence of the field’s only controlled experiment.
Limitation: the KDD 2024 study is a benchmark experiment, not live traffic. The direction is reliable; the specific 40% figure does not generalise to every topic.
What do these numbers mean for your business?
Together they tell one story: discovery is moving from a list of links to a single synthesised answer. When that answer names a competitor instead of you, you don’t lose a ranking position — you lose the consideration entirely, and usually invisibly.
Three practical consequences:
- Ranking is necessary but no longer sufficient. You can be first on Google and still be absent from the AI answer above it.
- Citation is the new click. In a zero-click world, a mention or citation inside the AI answer is the impression that counts.
- The industry average is not your number. Every figure above is a direction, not a target. The only measurement that is about you is one run on your own keywords and your own market’s questions — before and after, the same way.
Closing that gap is exactly what generative engine optimization and answer engine optimization do. If you want to see where your business stands today on ChatGPT, Perplexity and Google AI, start with a free AI visibility check.
About this page’s methodology. Every figure above comes from a public, third-party measurement, with the publisher, the sample and the date named, and each carries a link to its source. Where two studies contradict each other, aiseo42 publishes both — AI search is too young a field to build on a single measurement. Where a study was produced by a company selling products in the field, that is flagged separately. This page contains no original data collection of our own; when we publish such data, it will be marked as ours. Last verified: 27 August 2026.