How we measure PR & Brand

Depth: wordings per question

How many wordings of each question the scan searches. Draw one is the question on your template; the rest are rewordings in your study's language, pooled under the question. Depth buys the sentiment read, which cannot be graded at all at one draw - the state of every scan run so far.

Why it stops at T2. Draws past three add half as much for twice the price.

* Our estimate, not yet validated by real runs.

Wordings, not repeats: two runs of one phrasing overlap about half to three fifths of the time, a meaning-preserving reword under a third, and wording carried 26.5% of the variance against the brand's own identity at 1.5%. Rewordings are written in your market by a cheap model held to a rubric (same meaning, no new constraints, places or brands). Any that come back as the original, a shuffle, a duplicate or an answer are thrown away; the draw then repeats the original, and a run whose rewordings all fell through is labelled repeats.

The mention rate barely moves with depth. Draws of one question pool as one cluster, so three count for about 1.44 independent reads: on the median scan of 52 questions the rate lands at about plus or minus 13 points at one draw and 11 at three. That correction was measured on repeats, and distinct wordings are less alike, so the real gain is larger; we publish the smaller figure until the wording correlation is measured on this audit, which is why every rung above one carries a star.

Two draws can only score 1.00 or 0.50, so two and three draws promise the same floor; your report shows what it actually measured, which may be Established. A fourth and fifth draw add 0.139 effective reads per question between them, against 0.298 for the first extra draw. The cost multiple is the multiple charged: writing the questions happens once, a few per cent of a scan, and is not discounted.

Breadth: questions per scan

How many questions the scan asks - on this audit, the axis carrying the weight. Each rung is a search depth on your template at the credits it is charged, and the count is the measured median, not a cap.

Why it stops at T3. At 250 questions a scan is still predicted to find sources at half its average rate.

* Our estimate, not yet validated by real runs.

Across 720 completed scans, sources found per question barely falls as the question set grows; the marginal rate between the ends is 1.21 new sources per extra question. The stars here are inferred from totals, because no scan recorded which question found which result. Each scan now records its own curve over the domains of admitted mentions and stamps its grade. Questions return in network order, so it is a rarefaction curve: how much we would have missed stopping earlier, not a planned sequence. Errored searches count as dry samples, nudging towards Saturated, and the report states how many errored.

The classifier: Jev

Since 5 October 2026 every scan is judged by Jev, a typed classifier that answers each judgment as a probability rather than writing text, at about a tenth of the cost per result of the language model that judged scans before. That changes what a scan can afford to read: every distinct result page is fetched instead of the top 30, and there is a fourth size, Giant, of 200 to 240 questions. If Jev is unreachable, a scan falls back to the earlier model rather than failing.

Pages are fetched one site at a time, as GeorgeBot, within each site's crawl budget and inside a five-minute limit. A scan of 1,006 distinct results across 270 sites fetched in 125 seconds, and one slow site set that time on its own. About a third of pages return readable text; the rest block crawlers or carry none and are judged from title and snippet as before. A page the time limit or a crawl budget skips is counted on the report.

Jev admits the sources the earlier model admitted: over 30 scans judged both ways, the median ratio of distinct source domains was 1.00. So the grades at 25 and 52 questions carry over unstarred from production, where 53% of 352 scans read Narrow at 25 questions and 51% of 105 at 52. Above 52 they are projections from the curves those scans recorded, using an incidence estimator that came within 15.5% when checked on held-out halves: at 100 questions 58% of scans are predicted Partial, at 200 92%. Each plus or minus figure is the worst-case 95% range on the mention rate at one wording.

Depth works as above: an extra wording is a new search, so a second doubles the searches, fetches and judging. No scan has yet run above one wording on either engine, so every rung past one stays an estimate. One proxy points the right way: judged in two halves, a document got the same sentiment from Jev 98% of the time and from the model 93%, so Jev's single read is the steadier one.

What the report claims

Fidelity: mixed

Why old trends are undefined

Until now every scan wrote fresh questions at a temperature chosen to vary them, and nothing recorded them, so no two scans of one brand asked the same things. A line between two of them compares two instruments. Our data shows it: ten pairs of scans of one template run within two days share only 55% of their sources, and two pairs run six hours apart share 56% and 61%.

Every scan now records its questions, the model that wrote them and the version of the admission rules, and can rerun the previous question set. That is off by default: fresh questions are what discovery needs. Freeze the set for a trend, leave it free for discovery. Older scans read as one draw with no question set recorded, and are not backfilled.

What we do not claim

• That this is what AI engines say about you. It is what search returns.
• That we found everything written about you. On this audit the breadth grade usually says more was still arriving.
• A confidence percentage on anything a model judged. A model confidently wrong three times running agrees perfectly.
• That two scans are comparable, unless one reran the other's questions. The report says which.
• That a draw is an independent look. Draws of one question are pooled as one cluster and the range is widened accordingly.
• That every draw was reworded. The report says how many fell back to the original.

Other audits