Everything in this section is a measurement we ran, published with its data, method and caveats. No opinion pieces, no "five tips" listicles: if an entry exists, there is a dataset behind it and a date on it. When a reading goes stale, the date says so before we do.
The Answer Volatility Reading
1,211 measured AI answers: 19% of same-question, same-engine repeats change their verdict; engines disagree with each other on three questions in ten. Recomputed monthly. CC BY 4.0, with sample data.
EV charger installation: who the AI names
108 measured answers, three engines, repeated runs. The live version runs on the home page.
Garden rooms: who the AI names
96 measured answers, three engines, repeated runs. The live version runs on the home page.
Loft conversions (London): who the AI names
102 measured answers, three engines, repeated runs. The live version runs on the home page.
The Instrument Reading
Before trusting our own volatility numbers, we tested the instrument: the same answers, classified repeatedly, to measure how much of the wobble is our classifier rather than the engines. Two-day protocol in progress; publishes with full sample sizes when it completes.
Updated July 2026 · new readings aren't announced anywhere special: they appear here, dated.