How we measure AI visibility — steps, time and limits.

The steps

1. 1. Entity inventory

   2–3 days

   We map how the brand is described today: own site, registries, directories, Wikidata and third-party mentions. Conflicting descriptions are the most common reason a model reproduces the wrong facts.

   Output: List of conflicting descriptions and missing entity links.

2. 2. Prompt panel

   1–2 days

   We fix a panel of the questions a buyer actually asks — per persona, topic and language. The panel is frozen; it does not change between runs, otherwise quarters cannot be compared.

   Output: Frozen prompt list with expected entities and expected URLs.

3. 3. Baseline measurement

   1 week of continuous runs

   The panel runs against every engine with several replicates per question, because answers vary between runs. We log citation, correct entity recall, correct URL and competitor mentions per answer.

   Output: Baseline per engine: citation rate, entity recall, misrepresentation.

4. 4. Retrieval analysis

   2–3 days

   The bot log shows which pages AI crawlers actually fetch. Pages fetched but never cited have an answer problem; pages never fetched have an indexing or linking problem. Those are two different fixes.

   Output: Prioritised list: answer problem vs. retrieval problem, per URL.

5. 5. Page-level work

   2–6 weeks

   Direct answer on top, key facts in table form, verifiable sources, stated limitations and a named author. Plus the technical base: truthful lastmod dates, IndexNow push and a documented bot policy.

   Output: Rebuilt pages following the page model, with before/after dates logged.

6. 6. Quarterly follow-up

   ongoing

   The same panel runs every quarter. We report crawl, retrieval, citation, clicks and conversion as separate metrics — never as one merged “GEO score”, because that hides where the chain breaks.

   Output: Quarterly report with raw data and change versus the previous quarter.

What is required of you

• —Access to the site's code or CMS so the page model can be changed.
• —Server or CDN logs so bot retrieval can be observed.
• —A decision maker who can approve factual claims and act as reviewer.
• —Tolerance for effects measured in quarters, not weeks.

What the method does not cover

• —Nobody can guarantee citation in a language model. Answers vary between runs, accounts, regions and model versions.
• —The method measures the frozen prompt panel — not the whole market. The panel can be extended, but then the comparison restarts.
• —Training data updates with a lag. New content appears first in the retrieval log and last in the model's memory.
• —We do not measure answers behind login or inside closed enterprise assistants.
• —The method does not replace classic technical SEO — it assumes it.

Frequently asked questions

How long before AI visibility changes?

Retrieval patterns in the bot log normally shift within a few weeks of rebuilding the page model. Citation rate in answers moves more slowly and is evaluated per quarter, because the models' underlying data updates with a lag.

Why several replicates per question?

The same question can produce different answers at different times. With several replicates per question and engine we measure how often a brand is cited, instead of drawing conclusions from a single answer.

How is AI visibility different from ordinary SEO?

Ordinary SEO optimises for a ranked list of links. AI visibility optimises for being retrieved and cited inside a generated answer — where the page must stand alone, with a direct answer, sources and stated limitations.

Can I run the measurement myself?

Yes. The prompt panel, run schedule and parser version are published with every quarterly report under CC BY 4.0, so the method can be replicated without going through Norrhavet.