How we measure
Written out in full, so any figure here can be checked against a run of your own.
The corpus
4,185 queries across 24 subject areas, built by expanding a list of named products and concepts through a fixed set of intent templates. It is generated by code rather than curated by hand, which means it is reproducible: the same generator gives the same list, so this week's measurement is comparable to last week's by construction rather than by luck.
Each query carries its intent — commercial, informational, comparison, navigational, support — assigned when it enters the corpus and never re-derived. That is what makes the intent splits on this site stable while the results underneath them move.
The measurement
Once a week, every query runs through one endpoint from the United States in English, with autocorrect on. A sample of 300 queries also runs from nine other countries. Nothing is re-fetched afterwards: the stored response is the record.
A block counts as present when the response contains its key. Google omits the key entirely when it renders nothing, so this is a presence test rather than an inference from an empty value — which matters, because the two disagree exactly where the interesting cases are.
The arithmetic
| Figure | How it is computed |
|---|---|
| AI Overview coverage | Queries whose response carried an aiOverview block, divided by queries measured |
| Cited without ranking | Citations whose domain does not appear in the organic results for that same query, divided by all citations |
| Weekly churn | One minus the Jaccard coefficient over page-one domains, this week against last |
| Country overlap | Jaccard coefficient over result domains, that country against the United States |
| Citation rate | Citations for a domain divided by the queries where it appears at all |
Jaccard rather than a raw count of shared domains, because Google does not always return ten results and a raw count would move when the page length did.
What we publish and what we do not
We publish measurements: positions, domains, result titles, which blocks appeared, and the list of sources an AI Overview credited. These are facts about the result page, and they are what every rank tracker has published since 2004.
We do not republish result snippets. Only their length is stored. A snippet is a fragment of somebody else's page in Google's rendering, and a collection of them is not ours to hand out.
What could be wrong
Three things, stated so you can weigh them:
The corpus is developer and marketing subject matter. Coverage figures here should not be read as coverage of Google overall — recipes and medical queries behave differently, and we do not measure them.
One measurement a week catches weekly change, not daily. A block that appears and disappears within a day reads as stable here.
Results are personalised and geolocated even when logged out. We hold location and language constant, which makes runs comparable to each other, not identical to what you see.
Run the method yourself
Free tier, no card. Everything above is ordinary calls to one endpoint on a schedule.