How AgentPass measures AI visibility

AI Visibility answers one question: when people ask AI for the kind of thing your business offers, does AI mention or recommend you?

AgentPass keeps a list of customer searches for your business — questions such as “best waterproof jacket under $200 in Cedar Harbor, WA” — and asks each one on each supported AI surface, on a schedule. Every answer is stored as an observation: which businesses it named, whether it recommended yours, the order it listed them in when it used a numbered list, and the sources it cited.

Customer searches are written the way a customer would ask, without your business's name, so they measure whether AI finds you — not whether it knows you when asked directly. They contain only public information about your business, never customer data.

From those observations AgentPass reports how often you appear (AI Visibility), how often you're recommended, how you compare with the competitors you track, and which sources AI answers use. Every number is computed by fixed rules from stored observations.

Only valid observations count. A valid observation is a completed answer from an AI surface. When a provider returns an error, times out or rate-limits a check, that check is left out of every percentage — a provider problem never counts as “not visible”. The AgentPass simulation is never counted.

A mention counts when AgentPass matched your business with high or medium confidence: your full name, a name you added, or your website's domain. Weaker matches — a single word of your name, for example — are shown in the evidence but not counted.

If fewer than 80% of the checks on a surface completed in a period, that surface is marked Monitoring incomplete instead of showing a drop. Each point on the history chart covers the 7 days ending that day.

Mentioned

The answer names your business. Every confident mention counts toward AI Visibility — including a mention with concerns, such as “mixed reviews”.

Recommended

The answer suggests your business as an option: inside a list introduced as recommendations (“Here are some good options…”), or with a recommendation in the same sentence. A mention with concerns is never counted as a recommendation.

First recommended

Only when the answer itself was an explicit numbered list and your business was first. Positions (“2nd in an ordered list”) are only reported for numbered lists — otherwise AgentPass says Mentioned or Recommended, never a rank.

Not mentioned

The answer didn't name your business. That's one observation, not proof that AI doesn't know you — AgentPass could not verify a mention in that answer.

Every result is labelled with the exact AI surface that produced it. AgentPass never shortens an API surface to a consumer app's name — results from “OpenAI API + Web Search” are never shown as “ChatGPT”.

• OpenAI API + Web Search

  API evaluation

  OpenAI's API answering with its web search tool. Results may differ from the ChatGPT app.

• Anthropic API + Web Search

  API evaluation

  Anthropic's API answering with its web search tool. Results may differ from the Claude app.

• Gemini API + Google Search

  API evaluation

  Google's Gemini API answering with Google Search grounding. Results may differ from the Gemini app and Google Search.

• AgentPass simulation

  AgentPass simulation

  AgentPass checks whether your own website covers each search. This is not an AI platform result.

API evaluation results may differ from consumer AI applications such as the ChatGPT or Gemini apps. Consumer apps can use different models, settings, personalization and search features than an API request. AgentPass doesn't automate consumer AI apps.

AgentPass simulation is not an AI platform result. It checks whether your own website's public pages cover what each search asks about — a product type, “pickup”, your city. It's shown separately as “Information coverage — AgentPass simulation” and never counts toward AI Visibility or any other visibility number.

Which surfaces run for your business depends on your plan and on the surfaces available in your AgentPass workspace. A surface that isn't available reads “Not available yet”.

AI responses can vary between runs. That's why AgentPass monitors repeated observations over time rather than treating a single response as definitive.

• The same question can get a different answer minutes apart: models vary their wording, and web search results change.
• At every check, AgentPass asks each monitored search on each surface — on some plans more than once — and reports percentages across all of those answers.
• Searches that depend on the time, such as “open now” or “tonight”, are marked time-sensitive, and every answer shows when it was checked.
• Searches name the location explicitly (“in Cedar Harbor, WA”). An approximate location hint is also passed where the surface supports one.
• AI providers update their models and search indexes. Changes over time are shown with the date they happened — “Change occurred after …” — never as cause and effect.

Every visibility number carries a confidence level and the reason for it — for example “Based on 24 observations across 2 AI surfaces.”

High

At least 60 valid observations from at least 2 AI surfaces, with at least 80% of checks completed.

Medium

At least 20 valid observations, with at least 60% of checks completed.

Low

Anything less. Treat the number as an early signal.

For a single customer search or theme: High with 10 or more valid observations, Medium with 4 or more, otherwise Low. AgentPass only suggests an improvement for a search once it has at least 4 valid observations.

Competitors are businesses you choose to track. AgentPass also lists businesses that AI answers name often — “Frequently appearing businesses” — but never adds one as a competitor unless you do.

Competitive share of voice = your appearances ÷ (your appearances + the appearances of every competitor you track). Each business counts at most once per answer, and only confident matches count.

Based on publicly observable AI answers only. AgentPass has no access to competitors' private data.

When an AI answer cites web pages, AgentPass records each cited page and its website, and groups websites by kind: your website, competitor website, directory, review site, social media, forum or community, news or press, marketplace or delivery app, maps listing, other website.

Citation share, for each website = valid observations that cite that website at least once ÷ valid observations. One answer can cite several websites, so the shares don't add up to 100%.

Citations are observations. A citation shows what an answer referenced — not why it named one business instead of another. AgentPass never claims that a source raises or lowers your visibility.

AgentPass measures and improves how your business can be discovered. It doesn't manufacture demand. AgentPass doesn't:

• Write blog posts or pages with AI to try to sway what AI says.
• Rewrite your website automatically.
• Create backlinks, reviews, Reddit posts or social media posts — for you or about anyone else.
• Automate ChatGPT, Gemini, Claude or any other consumer AI app, or create accounts on them.
• Promise a position in AI answers.
• Generate fake customer demand or fake activity of any kind.
• Run or manage paid ad campaigns.
• Act as a search engine or a marketplace. Customers keep using the AI tools they already use; AgentPass measures and improves how your business shows up there.

• API evaluations aren't consumer apps. API evaluation results may differ from consumer AI applications.
• Searches are a sample. AgentPass monitors the customer searches you choose, not every question customers ask.
• Names are hard to match. A business with a very generic name (“Hot Chicken”) only counts on its full name near its location, or its website. A business with a similar name can occasionally be confused; weak matches are shown but not counted.
• Absence isn't proof. When an answer doesn't name you, AgentPass could not verify a mention — it doesn't mean AI doesn't know your business.
• Sources depend on the surface. Some surfaces cite few pages, or return redirect links instead of the page's address; AgentPass records what the surface provides.
• Small samples move more. A few answers can swing a percentage; that's what confidence levels are for.
• Stored answers expire. Full answer text is kept for 180 days. Counts, mentions and sources are kept so your history stays complete.

You don't need any of this to use AI Visibility.