How Mark measures AI visibility

Transparent, reproducible and explainable – every number links back to the raw answers.

1. Question funnels

People rarely ask an assistant “is Brand X good?”. They start with a need and narrow down. Every project runs persona-based journeys – multi-turn conversations with these stages. They come from your own question templates and custom funnels, or are written by an AI that only sees your category and market (never your brand, product or website):

Awareness – a vague problem, no category or brand named.
Consideration – “what are the best options?”
Comparison – trade-offs, pricing, reputation for the persona.
Decision – “if you had to pick one…”
Brand probe (only in questions you write) – names your brand. Measures knowledge and sentiment; excluded from visibility rates. AI-written funnels never name your brand.

Each journey runs as a real conversation (previous answers are kept in context), on every selected engine.

2. Detection

Each answer is parsed for your brand name, aliases and domain; list position (rank); tracked and newly-discovered competitors; an estimated sentiment of the sentences that mention you; and the sources the engine cited. Only sources the provider returned from a real web search count as citations (OpenAI url_citation, Gemini grounding chunks). Links a model types into its answer are shown separately as unverified.

3. Scoring

engine score = 55% mention rate + 20% rank quality + 15% persistence + 10% sentiment
AI Visibility Score = mean(engine scores)

Persistence is the share of steps after your first appearance in which you’re still mentioned. Drop-off step is the step right after your last appearance in a conversation.

4. Estimated path to your brand

For each journey we compute the probability that your brand surfaces by step k as 1 − Π(1 − mention rate_i) and highlight the shortest, highest-probability chain of queries in the graph.

5. Real data only

There is no simulated, sample or invented data. An engine without a working key can’t be selected; a failed call is shown as a failure and excluded from the score. ChatGPT, Gemini, Claude and Perplexity are asked to search the web; whether a search actually ran (and which queries and sources it used) is recorded for every answer – some keys can’t search (e.g. a Gemini key without billing). Recommendations appear only when this run’s answers or the site audit triggered them, with that evidence attached.