How CompetLab measures AI visibility

How we collect

We collect AI-visibility data by putting real, buyer-style questions to the AI engines and reading what comes back — the same way a prospective customer would ask an assistant to recommend a tool in your category. Nothing is scraped from a search page; the measurement is the answer an engine actually gives.

Each project runs a set number of monitoring prompts — the questions a buyer might ask about your space — and that number is set per organization. For a check, CompetLab sends each prompt to five engines: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Perplexity, and Google AI Overviews. Every stored check records the engines and prompt count it actually asked, so that ask is a fact about the check rather than a fixed rule, and an older check may carry a smaller one. We then read every answer for the companies it names — your own domain, each tracked competitor, and every other brand that comes up.

Your monitoring prompts
   │  (buyer-style questions about your space)
   ▼
Ask each engine  →  ChatGPT · Claude · Gemini · Perplexity · Google AI Overviews
   │  every prompt × 5 engines, per check
   ▼
Read every answer for the companies it names
   → who is named (you + every brand that comes up)
   → how the chat engines describe them
   ▼
Roll up per brand
   → presence, pooled over up to 5 checks, with a 95% range
   → the market map (three zones) → AI Visibility Score (0–100)

As the diagram shows, the same cycle runs for your brand and every company the answers name, so the numbers are comparable across the whole field rather than only across the list you configured. From the parsed answers we compute each brand’s presence with its 95% range, place it on the market map, and compute the AI Visibility Score.

Every rate divides by the answers that came back, not by the queries sent. Those are usually the same number, but not always, and there are three outcomes rather than two. An engine that answered and named nobody is a measured absence — a real finding. An engine we could not read produced missing data. And an engine that was read but had nothing to show — today, a Google results page carrying no AI Overview — is a third thing again: measured, not a failure, and excluded from every count. Only the first is an absence; none of the three is ever reported as “not mentioned”. Never read a 0 as “we couldn’t measure.” An agent can pull exactly these fields through the MCP tools reference or the AI Visibility endpoints.

How we compute the AI Visibility Score

The AI Visibility Score is a 0–100 reading of where a brand lands when it is named. It counts only the top five positions in an answer, evenly spaced — first place the most, the fifth the least — and nothing below them. Each check’s raw figure is then averaged over that check plus up to four previous published ones, so a single unusual day cannot swing it.

It is pooled across every engine a check asked, and it is deliberately not a per-engine number — which is why a provider-filtered trend returns no score at all rather than a pooled figure wearing one engine’s name.

The score is not a statement about who is ahead. It reads position, not standing, and it can favour a brand named half as often as another. Any claim that you lead or trail belongs to presence — how often each brand is named — which is what the market map reports. A score of 0 for a brand the answers did name is a real measurement: it was named only below the top 5, or too seldom inside them for the average to register.

There is no average position anywhere in this dimension. CompetLab does not store, publish or display an average rank, by design: averaging ordinal positions across answers of different lengths from engines that phrase things differently moves far more from sampling than from anything a company did. Position survives only inside a single stored answer, which the API will hand you on request.

The evidence behind every figure is returned by the API — each brand’s presence with its interval, the answers counted, the per-engine breakdown, and the stored answers themselves — so you can audit any number in the AI Visibility endpoints rather than taking it on faith.

Refresh cadence

Refresh cadence depends on the dimension: the six monitored dimensions run continuously on a schedule, while the once-a-month Strategic Briefing carries the deeper read. AI Visibility is a monitored dimension, so your score is refreshed on a recurring cycle rather than only when you ask for it — no more often than every three days, which is also the floor for AI Sources.

Because every monitored run is kept, day-to-day movement is visible as a trend, not just a latest value — you can page through the check history or read a time series. The eight researched dimensions are refreshed as part of the Strategic Briefing, which regenerates every 30 days on Monitor, every two weeks on Process; see the strategic-briefing endpoint for the availability states and dating. For how each monitored dimension is scheduled, see Monitoring.

What we do not claim

AI answers are probabilistic, so we are deliberate about what these numbers do and do not prove. Being explicit about the limits is part of the methodology, not a disclaimer bolted on to it.

• Answers vary run to run. The same prompt can produce a different answer on a different day. A single check is a snapshot; the trend over time is the reliable read, which is exactly why we sample on a schedule and keep history.
• We measure a sample, not the whole universe. A check runs your configured prompts — a representative set of buyer questions — not every way a person could phrase one.
• We measure five AI engines. ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. We don’t claim to cover every model, assistant, or answer surface, and we don’t model how one logged-in user’s personalization or region might change a reply.
• Correlation, not causation. A score change lines up with what you and your competitors do, but it doesn’t prove why it moved. Treat it as a signal to investigate, not a verdict.
• We measure; we don’t control. CompetLab observes what the engines say. It doesn’t inject content, pay for placement, or steer a model.
• Conveniences aren’t recommendation levers. Things like an llms.txt file are forward-looking niceties for coding and RAG agents. We don’t claim they are proven factors in whether an AI engine names or recommends a brand, because they aren’t.

Authorship

This methodology is maintained by the CompetLab team that builds and operates the measurement pipeline described above. When the pipeline changes — the engines we query, how a check is composed, how the score is built — we update this page and the affected references together, so the docs don’t drift from what the product actually does.

If a figure here ever looks inconsistent with what the API returns, the API response is the authoritative source; tell us and we’ll reconcile the page. This page was last reviewed on 10 July 2026.

Related

Where these numbers live, and how to pull them yourself.