Calibration corpus
Every threshold we use, published
These are the numbers that decide what we report and what we refuse to report. Each one is read directly from the constant that governs the behaviour, so this page cannot quietly drift away from the product. Where a number is a placeholder awaiting more observations, it says so.
It is common in this category to describe a private calibration corpus, or to state that a methodology is public without publishing one. Those claims cannot be checked, which is much of their appeal. This page exists so ours can be: every row names where it is enforced, so you can go and read the code that uses it.
When we refuse to report a change
Most of what a measurement tool gets wrong is reporting movement that is not there. These are the floors below which we say nothing moved.
Visibility score change
Provisional3 points
The composite is an integer built from seven pillars, several of them sampled, so it wobbles between consecutive scans without anything having changed. PROVISIONAL: this is the smallest move we are willing to call a change, not the measured standard deviation of no-change intervals. It will be replaced by that once enough consecutive-scan pairs exist.
packages/core/src/score-delta.ts
Source citation share change
Provisional0.5 share-points per period
Below this a source is reported as holding steady rather than as moving in either direction.
packages/core/src/source-trust-decay.ts
Scoring-engine version mismatch
ChosenComparison refused entirely
Two readings produced by different versions of our own scoring engine are not two measurements of the same thing. Their difference mixes the site's change with ours, so we report no delta rather than an ambiguous one. An unstamped reading is not treated as comparable to a stamped one.
packages/core/src/score-delta.ts
When we withhold a number entirely
A held number is not a missing feature. These are the points at which we judge the evidence too thin to publish anything.
Audit fix ranking
ChosenLeverage weight below 2 of 10
Fix weights are integers on a scale we authored, and the bottom step of a hand-authored scale carries no ordering information. A fix below the floor still appears in your audit, but with no number and no rank position: we would rather show the fix than publish a rank we cannot support. Printing a small number instead would read as a small effect, which is a different claim from an unrankable one.
packages/core/src/audit-triage.ts
Audit letter grade
ChosenHeld below 71% of components measured, and on any partial page fetch
A letter is the most quotable thing on a report and the least qualified. Nobody screenshots the caveat next to it, so where the evidence is thin we issue no letter at all rather than one that travels without its context. The grade covers on-page readiness only, never AI visibility overall, because a single free crawl cannot see the rest.
packages/core/src/audit-triage.ts
Composite visibility score
ChosenHeld below 5 of 7 pillars measured
A composite built from two measured pillars is an average of what happened to be available, not a measure of visibility. We show the coverage ratio instead, and never a partial average dressed as an overall score.
packages/scoring
Checklist-vs-citations correlation
ChosenHeld below 30 sites
The whole category asserts that on-site signals drive citations without publishing the correlation. Publishing ours from a handful of sites would be the same unbenchmarked claim with our name on it.
packages/core/src/proxy-vs-proof.ts
Source trust direction
ChosenHeld below 8 weeks of stored answers
A trend drawn across three weeks is a line between two points. We show the counts and withhold the arrow until the history supports one.
packages/core/src/source-trust-decay.ts
Single source trended
Chosen20 citations minimum
A share computed from a handful of citations moves several points on one lucky sample.
packages/core/src/source-trust-decay.ts
What counts as your brand being mentioned
Mention detection decides a number customers are judged on, so the rules are published rather than tuned quietly.
Names matched
ChosenBrand, asset label, hostname, plus owner-confirmed aliases
An engine writing a brand a different way still counts. Aliases are owner-confirmed, never inferred: an alias we applied unilaterally would be our claim about someone's brand rather than theirs.
packages/worker/src/citation-context.ts
Minimum alias length
Chosen3 characters
An alias raises a measured number, so a short or common one is a lever for inflating your own score. Common words are refused outright, with the exception that genuine brands which happen to be ordinary words are still accepted.
packages/core/src/brand-alias.ts
Maximum aliases
Chosen25
Not a plan limit, a correctness limit. Nobody legitimately has forty names, and a long list is where an over-matching entry hides.
packages/core/src/brand-alias.ts
Alias set changed
ChosenHistory comparison refused
Adding a name raises the mention rate with no engine having changed behaviour. Comparing across that change would attribute our own configuration to the market.
packages/core/src/brand-alias.ts
Thresholds that decide a verdict
Where a number becomes a word, the boundary is published.
On-site readiness called ready
Chosen70 of 100
Roughly where a heuristic checker would tell you the on-site work is done. Readiness deliberately excludes the citation result: including the outcome inside the proxy would correlate the number with itself.
packages/core/src/proxy-vs-proof.ts
Citation rate called cited
Chosen10% of sampled answers
Below this, engines are effectively not returning the site as a source.
packages/core/src/proxy-vs-proof.ts
Statistical significance on change
ChosenFisher's exact test, 95% interval
Applied to every before-and-after verdict, including Vendor Check. A change inside the interval is reported as no measured change rather than as a small win.
packages/core/src/measured-delta.ts
What Provisional means here
A provisional threshold is a deliberate placeholder: the smallest value we are willing to act on, chosen before we had enough observations to derive one. It is marked rather than hidden because a placeholder presented as a finding is the failure this whole document argues against. Each will be replaced by a measured value, and the version above will change when it is.