Original data on how the web performs in AI search
We point our own audit engine at large, disclosed samples of real sites and publish what we find, including the sample, the method, the limitations, and the numbers we decided not to publish.
AI knows your brand, not what you do
We audited 800 domains drawn from the Tranco ranking and asked Google AI a second, category question about a paired subsample. 84.6% are cited when asked about by name; 27.6% when asked about their own category.
Engineering notes
How the audit engine works, and where it has been wrong. These are not data studies: they are post-mortems on our own instrument, published because a measurement tool that never discusses its own defects is asking to be taken on trust.
A check that could never pass
For months, one of our GEO checks could not return a pass for any site on earth. Three independent bugs, each sufficient on its own, and every one of them hid behind a 200 OK. Had our 800-domain study run before we found them, its headline would have been a fact about our own code.
A sandbox DOI looks like a real one
Publishing our dataset, a rehearsal returned a DOI under a test prefix. Nothing about it looked wrong: same shape, same length, and the archive built, the upload succeeded and the page rendered a citation. The only check that separates it from a real one is resolving it.