Scorer changelog
When we change how the score is calculated, your number can move without anything about your brand moving. Every one of those changes is listed here, with the dates each version was in force and the reason it changed.
Why this page exists
A visibility score is only comparable with itself. If the scoring method changes underneath a trend line, the line reports our arithmetic rather than your visibility, and nobody looking at it can tell the difference. The IAB framework asks a measurement programme to re-baseline when that happens and to report the periods before and after as separate series rather than stitching them together.
So each scan we store is tagged with the scorer version that produced it and with a measurement baseline. Change the scoring method, the engine panel, one of the underlying models or your question set, and a new baseline opens. Inside the product, a movement measured across a baseline change is not shown as a movement: the dashboard says the baseline reset and names the reason instead of printing a number that would be ours rather than yours. Both fields, the scorer version and the baseline, are also columns in the raw export, so you can group your own rows by them.
Scans collected before we started tagging carry no baseline. We treat an unknown baseline as "no change detected" rather than inventing a boundary, because a fabricated reset is as misleading as a missed one.
Versions
Three more ways of saying "I have never heard of them"
In force today13 August 2026 to today
The 2026-08-12 epoch stopped a knowledge disclaimer losing its veto to the remedy that followed it. Re-scoring against the labelled set showed six answers still counting as citations, every one of them an engine saying plainly that it did not know the brand. All three causes were vocabulary rather than logic. First, the disclaimer list held "could not find" and "cannot find" but not the contracted "couldn't find", so one spelling of a sentence was an absence and the other was a citation, decided by an apostrophe. Second, "Without public reviews or case studies on X, their effectiveness remains unclear" is an absence of evidence stated about the brand, and it was rescuing the mention because it is a clean clause that happens to name it. Third, a clause that puts a question back to the reader is part of the not-knowing response, not an answer: "do you have a website or company name associated with X?" is the engine admitting it does not know what X is. That rule only applies where the answer disclaims the brand somewhere, so an ordinary question in a real recommendation is untouched. This only ever REMOVES citations. Re-scoring all 5,977 stored cells, 11 stop counting and none start, moving the mention rate from 21.78% to 21.60%. Every one of the 11 was checked by hand and is an engine disclaiming knowledge of the brand.
2026-08-13.disclaimer-vocabulary
An engine saying it has never heard of you is no longer a citation
12 August 2026 to 13 August 2026
Three corrections to what counts as a mention, found by measuring the classifier against a hand-labelled set for the first time. Two of them REMOVE citations, so scores may step down, and that is the correction rather than a drop in visibility. First, a knowledge disclaimer no longer loses its veto to the remedy that follows it. "I have no information about X. To find out, search for X on Google" names the brand twice, and the second clause was clean, so the answer counted as a citation; the same held for "could you tell me more about X?". That is the stock response to a brand an engine does not know, and it was the single largest error in the classifier. Second, a name attributed to a different vendor is that vendor's product: "SearchScore by SearchUnify" and "SearchNode (formerly SearchScore)" were both counted as ours, which inflated our own visibility, and any brand whose name is a common compound was exposed to it. Third, a multi-word brand now matches its own singular or plural in the final word only, so "Millennium Hotels" is found in "Millennium Hotel and Conference Centre". That one ADDS citations that were always real. A single-word brand is untouched, so Boots can never match Boot.
2026-08-12.mention-integrity
Ranks must be corroborated, and sentiment carries its confidence
11 August 2026 to 12 August 2026
Three changes to how a citation is read. A rank is now stated only when the markdown reading and the classifier reading of the same answer agree on the same number; they agreed on 16% of the cells where either had an opinion, and the markdown reading alone was what customers saw. An uncorroborated rank is credited as a prose mention, never as a low rank. The markdown reading itself was also fixed: its list counter did not reset between lists, a prose mention above a list beat the rank below it, and a sentence after a list of rivals inherited the count, which reported a top-3 rank for a brand that was never in the list. Sentiment is now weighted by the confidence the classifier returned with it, which was previously stored and ignored, so a hedged verdict moves the score less than a certain one. Scores may step down slightly where ranks are no longer confirmed; that is the correction, not a change in visibility.
2026-08-11.corroborated-position
Transactional intent added
9 August 2026 to 11 August 2026
Adds a transactional intent, completing IAB's four intent types. It is decision-stage, being the closest of the four to a purchase, so 8 question-instances across the live sets move into the scored set. Prompt-type coverage goes from 3 of 4 to 4 of 4, which is the criteria-matrix row separating directional from decision-grade. Brand-named comparisons were NOT reclassified: that behaviour is deliberate and locked by tests/avs-brand-safety, and they are now marked by isBrandedComparison for separate reporting instead.
2026-08-09.intent-taxonomy
Provenance dropped from cell credit
27 July 2026 to 9 August 2026
Provenance dropped from cell credit: it measured engine retrieval capability (Perplexity 99% vs Gemini 18%) rather than the brand. Sentiment and position only.
2026-07-27.provenance-removed
The first AI Visibility Score
Until 27 July 2026
AVS weighted by sentiment, position and provenance.
2026-06-01.initial
What counts as a change
The version above moves whenever the scorer would produce a different result for identical input. A wording change, a new chart or a faster query does not move it. Anything that alters what a cell is worth does, and a golden corpus of stored answers is re-scored on every change so that "identical input, identical output" is checked rather than assumed.
Changes that are not ours are handled the same way. An AI engine swapping its underlying model, or joining or leaving the panel we query, also breaks comparability, so those open a new baseline for your account even though our code did not change. We record the model that actually answered on every observation for that reason.
Back to measurement standards, where we grade each plan against the IAB criteria and name what it falls short on, or on to the IAB disclosure fields, where the rest of the method is set out field by field and generated the same way this page is.
Current scorer version 2026-08-13.disclaimer-vocabulary. Framework reference: IAB, Measuring Visibility in the AI Era, August 2026.