Methodology
Seenfact reports what AI search surfaces answered to a fixed set of questions under fixed conditions. This page states exactly what that means and what it doesn't.
What we measure
We sample answers, not audiences. A rate such as "mentioned in 34% of answers" describes the answers we collected, not the share of real users who saw your brand.
Every metric shows its sample size and links back to the answers it was computed from.
Surfaces, not model APIs
We collect from the consumer surfaces buyers actually use — ChatGPT Search and Google AI Mode — not from the underlying model APIs. The same model answers differently, and cites different pages, behind an API than in a search product.
If a surface can't be collected on a given day, that day is recorded as missing for it. We never fill the gap with a different model or vendor.
How a sample is collected
Each question is sent verbatim — we never add your brand name to it — in the market and language you chose, without personal history.
The raw answer is stored before any analysis, with its time, surface, collection settings and the model the surface reported. Changing settings or vendors starts a new configuration version, and trends break there instead of silently mixing.
Missing samples are not "not mentioned"
A failed or missing sample is excluded from every rate and counted separately. Treating it as an answer without your brand would make collection problems look like lost visibility.
Questions that name your brand
A question such as "Is Acme any good?" or "Acme vs Foo" contains your brand, so its answers almost always mention you. These questions are excluded from visibility rates and used to check what AI says about you instead.
Questions like "Foo alternatives" don't name your brand and do count.
Rates and intervals
Brand mention rate is k/n: answers that mention your brand (k) over valid answers to questions that don't name it (n). Own-site citation rate counts answers that cite a page on your domain. Recommendation rate and average rank need the recommendation analysis stage and will appear once it ships.
Every rate carries a Wilson 95% interval, which stays honest at small sample sizes. Week-over-week changes are reported only when they are statistically significant.
p̂ = k / n, z = 1.96
center = (p̂ + z²/2n) / (1 + z²/n)
margin = z · √(p̂(1−p̂)/n + z²/4n²) / (1 + z²/n)
interval = [center − margin, center + margin]How diagnostic recommendations are drafted and reviewed
Mention rates, sample coverage, and confidence intervals in a one-off diagnostic report are always computed deterministically from frozen samples. A model never decides whether your brand was mentioned and never generates rates or counts.
Summaries and priority actions may be written manually by a reviewer, or drafted by a model from the order's frozen answers, cited pages, and captured brand website text, then verified and edited by a person before publication. Model drafts are never published automatically.
Every delivered diagnostic report records its drafting method in the immutable snapshot (manual, or drafted by a model from frozen evidence and reviewed by a person). Older reports published before this field existed do not have it and are shown as manual without backfilling.
Limits
Answers vary between runs, users and days; one day's sample is noisy, which is why we sample daily and read trends weekly.
Surfaces change models and layouts without notice. We record what the surface reported and annotate breaks, but we cannot see inside the systems we measure.