Methodology

An AI visibility score can be copied in a week, and that is not the problem. The problem is a score without meaning: a number that cannot be compared with the market and whose change cannot be told apart from chance. So we publish the method in full, while the data that gives the score its meaning (the national corpus, noise bands, client histories) remains the product of our own measurements.

Every claim on this page has a source in code or a dated measurement. The weights table is checked automatically against the scanner itself: if the code changes and this page does not, the test blocks the release.

The score measures one thing: whether the given page is ready to be read and quoted by an AI system. It consists of seven checks whose weights are listed in the table below. It is computed from the page HTML plus three extra requests (robots.txt, llms.txt, sitemap.xml) and one HTTP probe with an AI crawler User-Agent.

The score is deterministic: the same page on the same day yields the same number, and no language model takes part in it. Measured on 2026-08-31 with two consecutive runs: 92/100 and 92/100, all seven checks identical. The score changes only when the page or the server's behaviour towards crawlers changes.

When an extra request fails (timeout, DNS error, server 5xx), the check is marked as unmeasured. It earns no points, but reports and monitoring do not compare it with the past: a measurement that did not happen is not a zero.

The score does not measure whether AI mentions you. That is a separate layer (see “What AI answers”) with its own questions and its own numbers. A high score with zero mentions is a common result and is shown as such.

The score measures only the one page you entered; the rest of the site stays outside it. The audit scans more pages; the free scan says only what it saw.

The score says nothing about search rankings and promises none. We never promise rankings or mentions in AI answers: it is the one claim we cannot back with a measurement.

For each business type (e-shop, local services, B2B) we keep 20 questions in four languages, written the way a buyer asks rather than the way a marketer writes. They are sent through official APIs to four engines (ChatGPT with search, Perplexity, Gemini with search, Claude with search); Google AI Overviews are measured manually because there is no API.

A mention is recorded by three rules: your domain in the answer text or in the cited sources; your name with tolerance for Lithuanian, Latvian and Polish inflection; in doubtful cases a second opinion from a language model that can only reject a mention, never create one.

A refusal or an empty answer counts as an error rather than as “answered, no mention”. The denominator is shown openly in every report: “2 of 18 answered” rather than “2 of 20”.

Every finding in the report comes with a quotation from the answer and a date. The client sees what AI said about them, in its own words rather than ours.

AI answers to the same question on the same day differ. A single measurement is therefore a snapshot rather than a trend. We measure the noise band by repeating an identical run several times and recording how many percentage points the mention share swings in each engine.

Measured in 2026-09 for one client from three runs: ChatGPT 15, Perplexity 35, Claude 35 percentage points. This means most monthly changes across 20 questions cannot be separated from noise, and the monitoring report says so directly: every change is labelled “outside the noise band” or “within the noise band”. Until the band is measured for a client, the report says exactly that.

A score gains meaning only in comparison. We therefore scan a corpus of Lithuanian business websites (302 domains, first run 2026-07-13, second 2026-08-17) and from 2026-10 will show your score's percentile against it: “better than X % of N measured sites”, with the measurement date. A sector breakdown is shown only when the sector holds at least 30 sites.

The corpus is scanned respecting robots.txt, public HTML only, one page and three extra requests per domain. A site owner can ask by email to have a domain removed from the corpus; we do so within 5 working days.

Two external findings directly shape how we word findings. Quotations are given verbatim in the original language; each was verified on the date shown.

• 97% of llms.txt files are never read

  Ahrefs, llms.txt study, 137,210 domains, 2026-06 · verified 2026-09-13

  This is why the llms.txt finding says only “present” or “absent” and never claims that AI reads it.

• 75% of the pages LLMs cite were updated in the last year

  Seer Interactive, 7,683 pages, 47,097 citations, 2026-07 · verified 2026-09-13

  This is why the page's last-modified date is shown as evidence under the meta check, although it does not change the score.

Every change to the score formula is compared on the corpus before release: old and new scores for every domain, with the difference recorded here. A formula without a recorded difference is not released, because a changed denominator would write a false change into every client's monitoring curve.

1. 2026-09-14 · v1.0 First public edition. Formula unchanged since 2026-08-17 (agent-access probe); llms.txt finding reworded to “present / absent”; last-modified date added as evidence without points. Corpus comparison: 0 score differences.