That surprised me, because most of the other pairs in our test did switch.

What we ran

October 19, 2026. 14 pairs of Google searches from the US, both sides of each pair meaning the same thing. In most pairs, Google Ads volume marked one phrasing as common and the other as rare.

We read the titles of the top 10 results on each side.

The usual pattern

For the rare side of the 11 pairs with a clear volume gap, 56 of 83 top results were titled only in the common words.

"Customer management software for small companies" got 8 of 8 pages titled as best CRM for small business. "Egg cooking time" got 7 of 7 about how long to boil an egg.

The reverse didn't happen once. 0 of 90 results for the common phrasings used only the rare words.

So why did the brakes query hold?

Because pages titled "car makes noise when braking" already exist in numbers. Google seems to pull a query toward the popular words only when the searcher's own words don't have many pages behind them.

The British "how to fix a dripping tap" did something similar. It mostly got "tap" titles, several from UK sites, from a US search.

The 3% number

Across all 14 pairs, the two result lists shared 7 URLs out of 207.

So Google understood the meaning and still showed different pages for each wording. That's the bit I'd keep in mind when planning pages.

One page per meaning is still the default, since two pages on one site would mostly split the same searches. But a phrasing that already has lots of pages of its own, like the brakes one, can earn its own section.

Underneath: embeddings

We ran bge-small, bge-large and mxbai-large on a laptop. Each ranked five sentences for three queries: one real answer in other words, four decoys sharing the query's words.

The small model fell for a decoy all three times. Its top pick for the dripping tap query was a sentence about tapping a screen to fix a paint effect.

The large ones did better, two of three each. Still close calls. bge-large gave the faucet decoy 0.730 and the real answer 0.715, and mxbai-large lost the CRM query by 0.004.

ChatGPT and AI Overviews

ChatGPT, through a data provider's feed, searched the web for 9 of the 29 queries. When it searched, it rewrote rare wording into common wording: "customer management software for small companies" became a search for CRM software for small businesses.

For the other 20 it answered from memory and named both words, like "high blood pressure (hypertension)."

AI Overviews showed on 27 of 28 pair searches. Where both sides had one (13 pairs), they shared 27 of 179 sources, roughly 15%.

What I'd take from it

Title the page in the words most people use. Put the other phrasings in the body, not on a second page.

Exact strings are different. Model numbers and error codes need literal matching, and in an earlier test of ours on search operators, ChatGPT dropped quotes when it rewrote searches.

And write answers that still read fine when quoted on their own.

Limits

14 pairs, one day, one location, Google Ads volume as the common/rare signal, and a 3-query embedding check. Treat the ratios as rough.

If you'd rather track this than test it once: RankControl, an AI SEO and AEO platform (https://rctrl.com), has an AI visibility tracker. RankControl's AI visibility tracker tracks citations across 7 AI engines: ChatGPT, Perplexity, Claude, Gemini, Grok, Google AI Mode and Google AI Overviews. RankControl's free GEO Checker scores how likely AI is to quote a page, 0 to 100, with the five weakest passages.

Full method and all 14 pairs: RankControl's semantic search guide. The raw pairs are on GitHub.