Research

How machines decide who you are

We publish what we measure. Both papers below exist because we needed the answer to build the product, and both changed what we built. Figures are dated, methods are described, and the limitations are printed next to the results rather than left out.

Paper01

What makes a trusted source

AI citationsInternational registers15 Aug 2026

Which sources ChatGPT, Gemini and Perplexity actually read when asked about a person, and why the answer has so little to do with how well known a website is. Includes the international register comparison, from Companies House to the SEC.

An open wiki ranking sixth on Google was cited by none of the three models. A public register was cited twice. Volume of coverage is not the mechanism.

Read the paper

Paper02

What actually resolves a name

Precision at 10Common names15 Aug 2026

If a search for your name returns three different people, something has to separate you from the others. We measured the four obvious candidates: job title, company, city and industry. Three of them made the results worse.

Adding your job title dropped precision from 55 to 45 percent. Adding your employer dropped it to 40. It is the field every form asks for.

Read the paper

How we run these

Every figure carries a date. Search results and model answers change weekly. A number without a date is a claim, not a measurement.
We publish what did not work. Both papers contain results that were inconvenient for us, including one that made us redesign our own sign-up form.
Limitations sit next to the findings. Sample sizes here are small. We say so in the paper rather than in a footnote nobody reads.
Absence is never assumed. When a check cannot establish something, we report that it could not be established, which is a different result from proving it is not there.
Clients are not named without a reason. Where a subject is a real person, they are described rather than identified unless they have specifically agreed otherwise.