Free scan Register
Menu

Changelog

Last updated 2 October 2026

Every change to the rubric, the checks, the AI models behind a measurement, and the tiers, with its date. A figure in a report can always be matched to the method that produced it.

2 October 2026

  • Rubric. Rubric 0.3.0: three new checks read the page where a business states what it sells and for how much, and score on AI Visibility. The first asks whether the prices sit in a list, a table or a price card rather than inside sentences; the second whether the pricing page declares its offers in its code; the third whether those declared names and prices match what the page shows. Two more are reported but not scored for now: whether the guide file for AI assistants (llms.txt) states the same prices and products as the pricing page, and whether a page's visible update date matches the one in its code. A site with no pricing page, or one that quotes on request, is not marked down. In a client's monthly report the previous month is recomputed on 0.3.0 from its stored scan, and a check that scan never ran is shown as new this month, never as a regression. No weight changed.

23 September 2026

  • Models. Moved the two jobs on our writing model to its newer version, which costs less per word: the pass that writes the Strategic Report's prose, and the internal judge that reads our own checkpoints. Nothing about the numbers changes: every figure in a report is built in code from the measured checks and the model only writes sentences over that list, and any page stating a figure the list does not carry is rewritten once and then withheld. The assistants we measure for a citation reading are asked through their own makers' models as before, with the ChatGPT column moved to the same newer version. Every model we use is reached through OpenRouter.

17 September 2026

  • Models. Pinned the model for the Strategic Report's narrative pass. The report's numbers are unchanged and still come only from the measured checks and studies: the model is handed a list of facts built in code and writes prose over it, and any page stating a figure that list does not carry is rewritten once and then withheld. No other job's model changed.

14 September 2026

  • Rubric. Rubric 0.2.0: the letter grade now reads off the open-finding count as well as the score. A score of 8 or more open findings (fails plus partial passes) on an axis drops its letter by one band, 12 or more by two; the number itself is unchanged, and the severity cap still governs where it binds. The weighted mean was insensitive to finding count: a site with nine open findings, a slow time to first byte and no confirmed identity could still read B, because no single category collapsed. The letter is the frame an owner reads, so it now follows both signals the report already prints. No weight changed.
  • Models. The checkpoint judge moved off our internal workhorse model. That model thinks before it answers, and on the first checkpoint it spent its whole budget thinking and returned nothing to read, twice in a row. The judge now runs on a model from a different maker with that thinking step switched off, so what comes back is the yes or no for each question and the passage that fails, and nothing else. It is also a different family from the models that write any text we send a client, which is what a judge should be. This corrects what we wrote here earlier today: the judge no longer runs on the same model as our other internal jobs. Everything else is unchanged. The workhorse model still runs our other internal jobs and judges accuracy twice at two settings, and the questions behind a citation measurement are still asked of the assistants being measured, each through its own maker's model: Perplexity, ChatGPT and Gemini.
  • Method. The Strategic Report now has a written standard: twelve criteria, each with the check that scores it, some by our code, some by a judge model, and only a major issue by a person. Before a change to the report line ships, a judge model reads the artifacts of that step against a fixed list of questions, answers yes or no, and quotes the passage that fails. The judge never sees the work of the model or person that built the artifact, and the list of questions is locked: a change to it is made by a person and recorded here. The judge runs on the same model as the rest of our internal jobs, at a setting that gives the same answer to the same input.

13 September 2026

  • Method. The Strategic Report is restructured for the decision-maker: an executive summary (the situation in one sentence, the one thing to do first, what it costs to leave it, quick wins, the next 90 days), your brief quoted back, one page per theme with its size of prize and its evidence, then the measured sections as annexes, the methodology and a glossary. Themes are chosen by fixed rules over the measurements, with thresholds printed in the methodology; a model words them from a digest of the theme's own evidence and every figure is checked back by code. A report whose themes could not be worded is held and reviewed by a person; it is never sent with anything in their place.

12 September 2026

  • Offer. The Strategic Report is now on sale, alone or in a bundle with the audit. It maps the demand in a market — 150 to 300 queries clustered by intent — then reports who ranks for each cluster, where an AI Overview or the local pack owns the answer instead, how often three AI assistants name the business and link to it, which of its own pages compete, three content briefs, the authority gap against the top three competitors, and a 90-day roadmap. Every number is measured by a named tool, and the report names the tool beside it. Delivered within 10 business days of payment, with a re-check 30 days later. The AI citation baseline, which the entries below call not yet sold, is part of this report and of AI-Ready.
  • Method. Changed how the AI citation baseline reaches its numbers, which is not yet sold. Whether an assistant named the business, and whether it linked to the business's website, are now counted by our code from the stored answer rather than judged by a model: the same stored answers give the same counts every time, with no model involved. Whether an assistant described the business correctly still needs judgement, so it is checked against a short list of facts our code has first confirmed against the business's own pages, and two judges must agree on it. Where they disagree, the answer is passed to a person rather than settled by us.
  • Models. Pinned the accuracy judges for the AI citation baseline. The questions are still asked of the assistants being measured, each through its own maker's model: Perplexity, ChatGPT and Gemini. Our workhorse model now judges accuracy twice, at two different settings, and both readings must agree; it no longer counts mentions or links, which our own code does. Any change to this list is recorded here.

11 September 2026

  • Site. Published a questions page with turnaround times, this changelog, and llms-full.txt, a plain-text version of our key pages for language models. Every item our tiers list is now tied to a code that is either built, delivered by a written procedure, or marked coming soon.
  • Models. Pinned the models for the AI citation baseline, which is not yet sold. The questions are asked of the assistants being measured, each through its own maker's model: Perplexity, ChatGPT and Gemini. Our workhorse model sorts the answers. Any change to this list is recorded here.

10 September 2026

  • Site. Launched aiqsight.com: the free scan, pricing, how it works, the rubric, a sample audit of a site we own, and a reference page for AI assistants. Opened the service to businesses in any country.
  • Rubric. Published the rubric as data at /rubric.json, generated from the same files the scorer reads.

9 September 2026

  • Rubric. Rubric 0.1.0 and check map 0.1.0: the scanner's checks mapped to two separate scores, Search SEO and AI Visibility, with a coverage floor below which a score is withheld, and a severity cap on each axis of its own.