Scorer changelog

13 August 2026 to 28 August 2026

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.

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.

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.

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.

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.

Until 27 July 2026

AVS weighted by sentiment, position and provenance.

From 26 August 2026 · scores move up

The answer-first check walked SIBLINGS of the h1, so any page builder that nests the heading in a div made a perfectly good lead paragraph unreachable and the check read an empty string. The social-proof checks never fetched /reviews/, /testimonials/, /press/ or /case-studies/, which are the pages that hold reviews and case studies. And the copyright proxy demanded a two-year span, so an accurate "2025-2026" failed while a stale "2025" passed. Measurement moves UP only, and only for sites that already had the signal.

From 25 August 2026 · scores move in both directions

Two readings of robots.txt were sharpened. A field our parser did not recognise, such as the Content-Signal line Cloudflare now adds by default, no longer ends the rule group it sits in, so a following Allow line is honoured rather than discarded. And a group is now closed by the next User-agent line, as the standard requires, so a rule written for one crawler can no longer apply to a crawler named earlier in the file. Measured by parsing the same 448 live robots.txt files with both versions: 94.0% of sites are unchanged, 21 moved up and 6 moved down, every move being 5.3 GEO points or less. Scores move in BOTH directions and some sites are newly reported as blocking, which is the correction working rather than a regression: a file that disallows everything and then whitelists named crawlers really does block the AI crawlers it leaves out, and the old grouping leaked those whitelist Allow lines to every agent in the file.

From 25 August 2026 · scores move up

A YouTube channel is now recognised when it is linked by its @handle, the format YouTube has issued since 2022, and when it is declared in Organization structured data rather than as a link. Measurement improved: worth 3 points of brand authority, and sites that had a channel all along are now credited for it, so scores move up only.

From 25 August 2026 · scores move in both directions

Claude-SearchBot, which indexes content for search, is now scored separately from ClaudeBot, which collects content that may contribute to training. The existing weight is split rather than added to, so no total moves. Measurement changed in BOTH directions: a site that allows search while declining training gains, and a site blocking the search crawler now sees the visibility gap it previously was not shown.

From 20 August 2026 · scores move up

The four trust-page signals now recognise contact, about, privacy and terms pages in languages whose words do not contain the English spelling, and the canonical-domain check is sharpened again. Measurement improved: worth up to 15 points of brand authority, and non-English sites gain most. Both changes move scores up only.

From 20 August 2026 · scores move up

The rendering signal now distinguishes a page that hides its content behind JavaScript from one whose crawler request was refused by a firewall or timed out. Measurement improved: a site that was never gated is no longer told that it was, and scores move up for every site whose audit had been blocked rather than gated.

From 19 August 2026 · scores move down

The llms.txt signal now assesses the file against the specification - a Markdown document with a heading and at least one link - rather than recording only that a file exists. Measurement changed and this one moves DOWN: a site whose file exists but does not meet the spec now scores lower than under geo-152. Adoption figures published before this generation were measured on presence and continue to say so.

From 19 August 2026 · scores move up

Article detection now distinguishes a published article from a grid of cards that happens to use the same markup. Measurement improved: homepages laid out as card grids are no longer asked for a publication date they were never going to carry, and gain the 6 points that question was costing them.

From 19 August 2026 · scores move up

Breadcrumb, Article and FAQPage signals in the SEO score are now judged across the same pages the AI-search score already used. Measurement improved: the two halves of a report now agree, and a site credited by one is no longer queried by the other.

From 19 August 2026 · scores move up

FAQ and speakable signals are now judged across every page the audit fetches rather than the homepage alone, and the audit reaches /faq, /faqs and /pricing alongside /about. Measurement improved: a site publishing questions away from its homepage is now credited for them, so scores move up only.

From 19 August 2026 · scores move up

The canonical-domain signal now compares the host the alternate address resolves to, so a site that correctly redirects one address to the other is recognised as consistent. Measurement improved: sites with a single correct canonical gain 3 points they were previously not credited with.

From 13 August 2026 · scores move up

The quotable-statistics signal now recognises money outside dollars, ordinary English sentences where the figure follows the object, verb-final languages, and a space as the thousands separator. It is worth 8 points. Measurement improved: scores read higher than geo-146 for any site previously unread, and non-anglophone sites gain most.

From 1 August 2026 · scores move up

The audit now reads authorship and publication dates from structured data, follows a site's own navigation to find its content, and no longer counts a soft 404 as a content page. Measurement improved: scores read higher than geo-145 for the same site, so figures must name the generation they were measured under.