Measurement you can check, including where it fails.
Every study here runs on our own production corpus. Each one states the cohort and the window, publishes the numerator and the denominator, shows an interval, lists what it excluded and why, and names the claims the evidence does not support. Where a result is too small to carry a rate, we report that instead of a rate. The aggregate data ships with the study so anyone can run the same calculation against their own.
How eight AI engines answer the same questions
Perplexity cites on every answer and keeps 89% of its sources between two runs; Google AI Mode keeps 24%. The consumer ChatGPT cites on 65% of answers, the API on 24%. The best agreement between any two engines on the same prompt is one source in five, and six citations in ten fall outside the two hundred most-cited domains.
10,639 valid answers · 9 engine surfaces · 77,526 citations over 14,477 domains · 11 July to 2 September 2026
Per-engine aggregate (CSV)The noise floor of every engine
How much each engine disagrees with itself when asked the same question twice on the same day, the constants the product applies today beside a fresh recomputation, and the rule that decides what may be called a change.
7,902 repeat pairs · same prompt, same engine, same day · rule measured 27 August, recomputed 2 September 2026
Per-engine floors (CSV)A flip rate is not a stability metric
How often a brand mention changes between consecutive runs has a null model, and the null is not zero. Two engines in our corpus flip within 0.4 points of each other: one is genuinely steadier than chance, the other is indistinguishable from a coin. Correcting for the base rate inverts the engine ranking at both ends.
8,060 graded answers · 9 engines · 32 brand records · 11 July to 31 August 2026
Per-engine aggregate (CSV)Field notes
Shorter measurements and one-off checks. Same reporting rules, smaller scope.
How we report. The calculation and reporting contract behind these pages is published in full on our methodology page. Where a published number later changes, the study says what changed and when, rather than being rewritten in place.