Written 30 October 2026.
Before you paste a statistic into a deck, a post or a page, open the document where it was first published and ask it five questions. If you can’t find that document, don’t use the number.
What I measured
On 29 October 2026 I traced the 32 AI search statistics repeated most across 20 statistics roundups. 28 led back to a source document I could open. None of the 32 passed all five checks below, and the best managed 4 of 5. Study and CSV.
Three things the trace turned up:
- “83% zero-click when an AI Overview shows” is credited to Similarweb. Neither Similarweb page I opened contains it, and I found no source.
- Four figures have been replaced by a newer number from the same source, and roundups still print the old one. Ahrefs’ 34.5% drop in click-through (March 2025) is still repeated; Ahrefs’ own later figure is 58% (December 2025).
- Botify’s “up to 47%” of 120,000 searches has no date or market on the source page. All three roundups that carry it state it as a current share of Google searches.
The five questions
Before I use this statistic, the source document answers all five:
1. When: does it say when the figure was measured?
2. How many: does it give the sample size?
3. Where: does it name the country?
4. How: is the method published in full?
5. Can I check it: is the raw data published?
Then: does my sentence say what the source says?
Same metric, same base, same country, hedges kept.The last line matters as much as the five. In the study, 16 of the 28 traced figures were stated correctly by some roundups and changed by others: a different metric, a different base, the country dropped or a “may” removed.
How to write it once it passes
Put the answers to the first three questions in the sentence, so they travel with the number:
Pew Research Center, March 2025, 900 US adults: 8% of visits clicked a result when an AI summary showed, against 15% when it didn’t.That figure scores 4 of 5, as good as any in the sample. If yours scores lower, say what’s missing (“no sample size published”), or leave it out.
Ask an AI to find the source
I want to use this statistic: [paste it].
Find the document where the organisation that measured it first published it. A roundup or news story doesn’t count. Give me the link.
Then, from that document only, tell me: when it was measured, the sample size, the country, whether the full method is published, and whether the raw data is. Write “not stated” where it isn’t.Open the link yourself before you trust the answer.
What I will not claim
- That a statistic that fails a check is wrong. It means you can’t tell.
- That these questions catch every bad number. They test whether a figure can be checked, not whether it’s true.
- Anything beyond the 32 statistics and 20 roundups in the 29 October sample.
Related: put the date in the same sentence as the figure and five labels every AI visibility number needs.
This is one of the search hacks, each backed by a dated test.