Only 3% of Google's results overlapped between two searches that mean the exact same thing.
That's the part of a test we ran in October that stuck with us. We took 14 pairs of queries, same meaning, different wording, and pulled the top 10 results for both sides of each pair. Across all 14 pairs, the two lists shared just 7 URLs total, out of 207 pages looked at. Google clearly understood the two searches meant the same thing. It still showed almost entirely different pages for them.
The other half of the test went the way you'd expect, mostly. For 11 of the 14 pairs, one phrasing was clearly more common than the other, going by Google Ads search volume. We checked whether the rarer phrasing's results used the searcher's own words or the popular ones. Two in three used only the popular ones (56 of 83), and for one pair it was total: all 10 results for "inexpensive airfare to london" were titled "cheap flights to london," not one page using the words someone actually typed. The pull went only one way. Common phrasings never got swapped into the rarer wording, not once across 90 results.
There was a clean exception. "Car makes noise when braking" kept every one of its 8 results in its own wording, no swap toward "squeaky brakes" at all. Enough pages already exist with that exact phrase, so Google had no reason to redirect it toward the more popular term. That's the practical rule buried in the data: Google reaches for the crowd's wording when your phrasing is thin on content of its own, and leaves it alone once it has pages to stand on.
Put together, that argues against two habits we see a lot. One is writing a separate page for every synonym of a topic, since two pages on one site would mostly just split the same searches. The other is assuming your team's internal phrasing is safe to put in a title, when the data says the words most people actually type are the ones that absorb the rare searches too.
We also ran three free embedding models (the tech underneath most semantic search) against five candidate sentences per query. The smallest model picked a decoy that shared the query's words every time. The larger two got it right two times out of three, and on one query the decoy scored 0.730 against the real answer's 0.715. RankControl's free GEO Checker scores how likely AI is to quote a page, 0 to 100, with the five weakest passages flagged.
RankControl is an AI SEO and AEO platform. The full method, all 14 pairs and what ChatGPT and AI Overviews did with the same searches are in the write-up, and the raw pairs are on GitHub.