Credible Roots / AI Search Statistics
Traced September 2026
AI search statistics, traced to the primary source
There are dozens of “57 AI SEO statistics” posts and most of them cite each other. This one lists only figures we followed back to the organisation that gathered the data, records the method and sample behind each, and names separately the widely repeated numbers we could not trace at all.
The best-sourced figure in AI search is SparkToro’s: 68.01% of US Google searches ended without a click between January and April 2026, measured on Similarweb desktop and mobile clickstream panel data, up from 60.45% in 2024. Expressed as traffic, every 1,000 US Google searches now send about 276 clicks to the open web, against 374 two years earlier — roughly a quarter of the open web’s clicks per search lost in two years. Two cautions travel with it that are almost never repeated. First, the long-run trend line stitches together three different measurement panels — Jumpshot for the 2016 and 2019 figures, which has since shut down, Datos for 2024 and Similarweb for 2026 — so year-on-year comparisons cross incompatible methodologies rather than tracking one consistent instrument. Second, the most-quoted AI Overview prevalence figures disagree by roughly threefold depending on the keyword set sampled, with Semrush’s analysis of more than 200,000 keywords putting coverage near 16% by November after a July 2025 peak around 25%, while figures near 48% circulate widely. Where a number matters to a decision, the sample it came from matters more than the number.
- 68.01% zero-click, US, Jan–Apr 2026, SparkToro on Similarweb panel data.
- 276 clicks per 1,000 searches now reach the open web, from 374 in 2024.
- The trend line crosses three panels. Jumpshot, Datos, Similarweb. Rarely disclosed.
- AI Overview prevalence estimates differ ~3x by keyword set. Check the sample.
On this page
Why another statistics page
Search for AI SEO statistics and you will find roundups promising 50, 60 or 100 data points. Follow the citations and a large share lead to another roundup, which leads to a third, which links to a page that no longer exists. The number survives; the evidence does not.
We have written about this pattern before in a narrower case: the “7-11-4 rule” is cited across the marketing industry as Google research, and Google has never published it. It seemed worth applying the same test to the statistics now being used to sell AI search services, including by us.
So the rule for this page is simple. A figure appears in the traced section only if we followed it to the organisation that collected the data and can state the method, the sample and the period. Everything else is in the untraced section, which is not an accusation — it means we did not find the primary publication, not that the number is wrong.
The figures that survive a trace
| Finding | Figure | Source, method and period |
|---|---|---|
| US Google searches ending without a click | 68.01% | SparkToro, using Similarweb desktop and mobile clickstream panel data, US, January–April 2026 |
| The same measure two years earlier | 60.45% | SparkToro with Datos panel data, 2024. Different panel from the 2026 figure |
| Clicks reaching the open web per 1,000 US Google searches | 276, from 374 | SparkToro, same studies. About a 26% reduction in two years |
| AI Overview coverage, peak and settled | ~25% July 2025, ~16% November | Semrush, analysis of more than 200,000 keywords |
| US searches concluding on the results page | 58.5% US, 59.7% EU | Semrush zero-click study, 2025, clickstream analysis |
| AI citations drawn from earned media rather than owned content | about 84% | Muck Rack, May 2026 |
| Brands cited via third parties versus their own domain | 6.5x | AirOps, October 2025 |
| ChatGPT citation presence in professional services prompts | under 4% | Similarweb, May 2026 |
| AI referral traffic as a share of total web traffic | about 1% | Search Engine Land, across 3.3bn sessions |
| Largest published AI search visibility prompt study we found | 24,067 prompts, 5 platforms | Minuttia, an agency in this field, published on its own site |
The last row is included deliberately. Minuttia is a competitor of ours, and it has published a larger original prompt study than we have. A statistics page that quietly omitted the best research in its own category because a rival produced it would fail its own test.
Where the numbers genuinely disagree
Two live disagreements matter more than any single statistic, because they determine whether a figure you are quoting means what you think it means.
AI Overview prevalence differs about threefold
Semrush’s keyword-set analysis puts AI Overview coverage at roughly 16% by November, after a July 2025 peak near 25%. Figures around 48% of tracked queries circulate widely elsewhere. Both can be internally valid: prevalence depends entirely on which keywords are tracked, and a set weighted toward informational, health or how-to queries will show far higher coverage than a set weighted toward navigational or transactional ones.
The practical consequence: “AI Overviews appear on X% of searches” is close to meaningless without the keyword set. It is not a property of Google. It is a property of the sample.
Zero-click rates measure different things
SparkToro’s 68.01% counts US Google searches that end without a click to the open web, which includes clicks back to Google’s own properties. Semrush’s 58.5% counts searches concluding on the results page. These are different definitions producing different numbers from overlapping realities, and quoting one alongside the other as a trend is an error.
The methodology problem nobody mentions
The most widely reproduced chart in this field shows zero-click searches rising over roughly a decade. It is assembled from three different measurement panels: Jumpshot for the 2016 and 2019 figures, a company that has since shut down; Datos for 2024, now owned by Semrush; and Similarweb for 2026.
Different panels recruit different users, instrument different devices and define a “search” differently. Stitching them produces a line that is directionally useful and numerically unreliable, in the same way a chart combining three countries’ unemployment definitions would be.
None of this makes the underlying conclusion wrong. Zero-click behaviour is rising, several independent instruments agree on the direction, and the 2024-to-2026 comparison is the most defensible segment because it is the most recent and best documented. But anyone presenting the decade-long line as a single consistent measurement is presenting something the data does not support, and that includes several agencies selling against it.
Figures we could not trace
These circulate widely across AI SEO roundups. In our check we did not locate a primary publication stating the method and sample behind them. That means untraced, not false. If you know the original source for any of these, tell us and we will add it and log the change.
- “AI Overviews reduce clicks by 58% for the top-ranking page.” Widely repeated; primary study not located.
- “Organic CTR dropped 61% where an AI Overview is present.” Appears in multiple roundups citing each other.
- “CTR bottomed at 1.3% in December 2025 and rose to 2.4% by February 2026.” Precise enough to imply a dataset; we did not find it.
- “Brands cited in AI Overviews earn 120% more organic clicks per impression.” Frequently quoted in GEO sales material.
- “AI Overviews appear on 48% of tracked queries.” Plausible for some keyword sets, but the set is rarely named — see the disagreement above.
- “60% of US adults read AI summaries at the top of results.” Survey-shaped; no instrument or sample found.
We are not suggesting these were invented. Several are probably real findings that lost their citation somewhere in the chain. But a number you cannot trace is a number you should not put in a client deck, and that standard applies to us first.
How to check a statistic in two minutes
- Click the citation, then click its citation. Most chains break at the second hop. If hop two is another roundup, keep going or drop the number.
- Look for a sample size and a date range. Real research states both. A figure with neither is a claim wearing a percentage sign.
- Ask what instrument produced it. Clickstream panel, survey, keyword tracker and log analysis measure different things and are not interchangeable.
- Check whether the definition matches your use. “Zero-click” means at least two different things in current research.
- Note who benefits. Much of this data is published by companies selling the remedy, ours included. That does not make it wrong; it makes the method worth reading.
Questions people actually ask
What percentage of Google searches end without a click?
68.01% of US Google searches ended without a click between January and April 2026, according to SparkToro using Similarweb desktop and mobile clickstream panel data. That is up from 60.45% in 2024, though the earlier figure came from a different panel. Semrush’s 2025 study, using a different definition — searches concluding on the results page — reported 58.5% in the US and 59.7% in the EU.
How much traffic do AI search engines actually send?
Around 1% of total web traffic, measured by Search Engine Land across 3.3 billion sessions. This is the figure most often left out of AI search sales material, and it is the reason we tell clients to buy answer engine work for position rather than for sessions this quarter.
What share of searches show an AI Overview?
It depends entirely on the keyword set, which is why published figures differ by roughly threefold. Semrush’s analysis of more than 200,000 keywords put coverage near 16% by November after a July 2025 peak around 25%. Figures near 48% circulate widely but usually without naming the sample. Treat any such percentage as a property of the keyword set rather than of Google.
Is the zero-click trend line reliable?
Directionally yes, numerically less so. The long-run chart stitches three measurement panels — Jumpshot, which has shut down, for 2016 and 2019; Datos for 2024; Similarweb for 2026 — which recruit different users and define a search differently. The 2024 to 2026 segment is the most defensible part. The decade-long line should not be presented as one consistent instrument.
Where do AI citations actually come from?
Predominantly from earned third-party coverage rather than a brand’s own website. Muck Rack put this at about 84% in May 2026, and AirOps reported brands being roughly 6.5 times more likely to be cited through third parties than through their own domain in October 2025. This is the single most commercially inconvenient finding for anyone selling on-site optimization alone.
Why publish statistics that undercut your own service?
Because the alternative is selling on figures we could not defend if a client checked them. The 1% referral traffic number and the earned-media citation share both argue for buying less on-site work than an agency would like to sell. We would rather state them and lose a badly-fitting engagement than have a client discover them in month four.
Related reading
Sources
- SparkToro. In 2026, less than one third of Google searches still send a click. 68.01% zero-click, US, January–April 2026, Similarweb clickstream panel.
- Search Engine Land. Google zero-click searches reach 68% in early 2026. Reporting on the SparkToro study.
- Semrush. Zero-click search study. 58.5% US and 59.7% EU; AI Overview coverage across 200,000+ keywords.
- Minuttia. AI search visibility data study, 24,067 prompts across five platforms.
- Credible Roots. The citation research, each figure attributed to its publisher. Muck Rack, AirOps, Similarweb and Search Engine Land figures.
How this was made: each figure was followed back toward the organisation that gathered the data, and recorded with its method, sample and period where those were stated. Figures we could not trace to a primary publication are listed separately as untraced rather than omitted or repeated as fact. Some sources were reached through reporting on the original study rather than the study page itself, which is noted where it applies. Corrections will be published in the corrections log with the original wording intact. Drafted with AI assistance, reviewed and approved by the named author before publication. Our editorial standards.
We will run these numbers on your category
Industry averages are the wrong instrument for a decision about your business. The diagnostic measures your prompt set, your citation share and who gets named instead of you, and reports every figure with its sample size.
Book a 30 minute call