Writing · Series: How AI Search Works · Part 13 of 16
The AI Search Statistics Everyone Repeats, and Where They Actually Came From.
I traced the numbers this industry quotes back to their sources. One rests on a survey of 299 people. One I published incorrectly myself. Two hold up completely. Here's the audit, and the method for running it yourself.
Published5 September 2026
Read time13 minutes
Filed underSeries · Evidence · Part 13 of 16
Every deck about AI search runs on about six numbers. Almost nobody who repeats them has read the source, because the source is usually two clicks further than anyone goes.
I went and read them. What follows is what I found: which numbers are load-bearing, which are scenario modelling dressed up as forecasting, and which are vendor marketing that has been laundered into fact through repetition. One of the failures is mine, and it's the one I'd most like you to read.
"Search volume will drop 25% by 2026."
This is the most-quoted statistic in the category. It has justified budgets, repositioned agencies, and opened more conference talks than anything else on this list.
Gartner published it on 19 February 2024, attributed to analyst Alan Antin. The evidence underneath it is a survey of 299 consumers, conducted in August 2023, in which 79% of respondents said they expected to use AI-enhanced search within the next year.
Read that again, because there are two separate problems in it. The sample is 299 people. And the question measured what people expected to do, which is a notoriously poor predictor of what they actually do. Gartner itself clarified in interviews that the 25% figure was scenario modelling rather than a forecast, a caveat that has been comprehensively stripped away in the retelling.
The window has now closed. It's 2026, and search volume did not fall 25%. The prediction has never been withdrawn, and it is still being quoted in the present tense.
299 people said what they expected to do. An industry rebuilt its pitch around it.
"Only 14% of AI Mode citations rank in the top 10." (I got this one wrong.)
Start with the correction, because I'm the one who needs to make it.
Across several articles on this site I attributed the 14% figure to Google, in at least one place calling it "Google's own data" from I/O 2025. That was wrong. Google has never published a figure for the overlap between AI Mode citations and organic rankings. I've corrected every instance and left dated notes where the error was most explicit.
The number is real, and it comes from SE Ranking. Their study analysed 9,734 keywords that triggered AI Mode, with data collected on 20 June 2025, published that August. The finding: average exact URL overlap with the organic top 10 was 14%.
The word doing the work there is exact. Match at domain level instead, and the same study reports 21.9%. Other vendors, using different matching rules and sampling windows, report materially different figures. So the number isn't wrong, but "14%" without "exact URL match, one vendor, June 2025" is a much stronger claim than the evidence supports.
What survives is the direction, not the decimal. A large share of what AI Mode cites doesn't rank in the top 10, and rankings are a weak predictor of AI visibility. That conclusion is robust across every study I've seen. The precise number is not.
"AI traffic converts 23x better."
You've seen this one with various multipliers attached: 23x, 27x, 4.4x, 14.2%. They circulate without provenance, and they're almost always sourced to a vendor measuring its own customers, using last-click attribution, on a self-selected sample.
There is a defensible version of this claim, and it's smaller and much more useful.
0.5% / 3x
AI referral traffic is tiny and converts well. Orbit Media analysed 97 GA4 accounts across B2B lead generation sites, covering 29 million visits between 1 July 2025 and 30 June 2026. AI sources accounted for just 0.5% of all traffic. That traffic converted to high-intent leads at 1.91%, against 0.71% for Google organic: roughly 3x. For ChatGPT specifically, 2.1% versus 0.5%.
Both halves of that matter, and the people quoting the multiplier usually drop the first one. AI referral traffic is currently a rounding error in volume terms. It also converts substantially better than organic search on high-intent actions.
What makes this study worth citing when the others aren't: Orbit Media sells no GEO tooling, so nobody's commercial interest is served by the answer. The methodology is stated, including that key events were manually categorised into high and low intent, with misconfigured events excluded. And the sample is large enough and boring enough to be credible.
Use 3x. If someone quotes you 23x, ask who ran it and on whose clients.
Two numbers that hold up completely.
This isn't a piece arguing that all research is bad. Some of it is excellent, and two studies from this year are worth more than everything else on this list combined.
AI recommendations are wildly non-deterministic. SparkToro had 600 volunteers run 12 prompts between 60 and 100 times each, 2,961 runs in total, across ChatGPT, Claude and Google's AI. There's under a 1 in 100 chance that any two responses return the same list of brands. For the same list in the same order, it's closer to 1 in 1,000. Separately, 142 human-written prompts expressing identical intent had a semantic similarity of 0.081.
That finding should reshape how you report AI visibility. Any single-run measurement is noise. Any month-on-month movement you report without knowing your own variance is probably you reporting variance. And since people phrase the same question very differently, your prompt set samples the phrasing space rather than covering it.
ChatGPT no longer runs on Bing's index. A Search Engine Land teardown analysed 1,200 ChatGPT answers, 88,000 search results and 26,900 distinct pages, and found that only 1.5% of the URLs ChatGPT retrieves appear in Bing's top 20 for the same queries. It retrieves through OpenAI's own hub instead.
Both of these are specific, methodologically transparent, and state their sample sizes up front. That's the pattern. Good research tells you how it was made.
Four questions to ask any AI search statistic.
You don't need to run a full audit on everything you read. You need four questions, and about ninety seconds.
1
Who benefits if this number is true?
A tool vendor publishing research showing that the thing their tool measures is critically important is not neutral. That doesn't make it wrong, but it lowers the evidential weight considerably. The most useful studies in this space come from people with nothing to sell you.
2
What is the sample, and when was it taken?
A number without an n and a date is not a finding, it's a claim. This field moves fast enough that a figure sampled 14 months ago may describe a product that no longer exists. If you can't find the sample size in two clicks, that's usually the answer.
3
Is it measuring behaviour or intent?
The Gartner number failed here. Asking people what they expect to do next year produces survey data about optimism, not a forecast of behaviour. Watch for this in anything derived from a consumer panel rather than from logs, clickstream, or a controlled experiment.
4
Does the source say what the repeater says it says?
This is where I came unstuck. Statistics acquire attribution as they travel, and "a study found" quietly becomes "Google confirmed" after enough retellings. Click through to the primary source before you put a number in a deck with your name on it.
The uncomfortable conclusion is that the confident numbers are mostly the weak ones. The Gartner figure is quoted with certainty and rests on 299 people. The Orbit Media figure is quoted rarely and rests on 29 million sessions. Confidence and evidence are close to inversely correlated in this field right now.
That's an opportunity if you're the one being careful. Clients notice when you say "one vendor study, June 2025, exact URL matching" instead of a bare percentage. It sounds like less certainty, and it buys you considerably more trust.
Frequently asked
Did search volume actually drop 25% by 2026?
No. Gartner published that prediction in February 2024, based partly on a survey of 299 consumers about what they expected to do, and described the figure as scenario modelling rather than a forecast. The window has closed and the drop did not happen. The prediction has not been withdrawn.
Where does the 14% AI Mode statistic come from?
SE Ranking, from a study of 9,734 AI Mode keywords with data collected on 20 June 2025 and published that August. It measures exact URL overlap with the organic top 10. Matching at domain level in the same study gives 21.9%. It is not a Google figure, despite being widely attributed to Google.
How much better does AI traffic actually convert?
About 3x on high-intent conversions, per Orbit Media's analysis of 97 B2B GA4 accounts and 29 million visits. AI sources converted at 1.91% against 0.71% for Google organic. The much larger multipliers in circulation generally come from vendors measuring their own clients with last-click attribution.
How much of my traffic should I expect from AI search?
In B2B, currently around 0.5% of total sessions based on the largest independent dataset available. That's small enough that you should size the opportunity honestly rather than reallocating budget on the strength of a projection. It converts well, which is the reason to care.
Why do AI visibility numbers change so much between tools?
Partly methodology, partly genuine non-determinism. SparkToro found under a 1 in 100 chance that two runs of the same prompt return the same list of brands. If a tool reports a single figure without a variability range, it is presenting a sample of a distribution as though it were a measurement.
How do I check a statistic quickly?
Ask who benefits if it's true, what the sample size and date are, whether it measures behaviour or stated intent, and whether the original source actually says what the person quoting it claims. If the sample size isn't findable within two clicks, treat the number as an assertion.
Next in the series
One recommendation gets repeated more than any statistic: publish an llms.txt file. Article 14 looks at whether anything actually reads it. Read Article 14 →