AI visibility
How often and how favourably AI systems mention a brand when they answer questions. What it measures, how it is tracked, and what it misses.
AI visibility is how often and how favourably AI systems mention a brand when they answer questions. It is measured by sending a fixed prompt set to chat models on a schedule and counting mentions, position and sentiment.
How the measurement works
Every product in this category does roughly the same thing.
- You choose a set of prompts. "What is the best hosted Postgres?" "Cheapest transactional email provider?"
- The tool sends them to several assistants on a schedule, usually daily.
- It parses each reply for brand names, where they appear, how they are described, and which sources were cited.
- It draws a chart.
Some tools add real user session data rather than only synthetic prompts, which is a better signal and costs more.
The metrics
| Metric | Definition |
|---|---|
| Visibility | Percentage of tracked prompts where the brand appears |
| Position | Where in the answer the brand appears, first mention being strongest |
| Share of voice | Your mentions as a proportion of all brand mentions |
| Sentiment | Whether the description is positive, neutral or negative |
| Citations | Which URLs the model linked or drew on |
The thing to understand about the score
The number depends entirely on the prompt set you chose. Change the prompts and the score changes.
That makes it useful in two ways and useless in a third.
Useful: tracking your own number over time with a fixed prompt set. Comparing yourself with named competitors inside the same set.
Not useful: comparing your score with a number from another company, another tool, or another prompt set. There is no absolute scale.
What it misses for a developer tool
The measurement assumes a person asked a question and will read the answer.
A large and growing share of developer tool adoption now happens differently. A developer opens a repository in a coding agent and types one line. The agent reads the codebase, decides what is needed, sometimes searches, installs a product and writes the integration code.
That surface is not in the measurement, and it does not track the chat surface:
| Finding | Why prompt monitoring cannot see it |
|---|---|
| Same request, one winner in 111 of 111 JavaScript sessions, 48 of 132 in TypeScript | There is no repository in a chat prompt |
| A written instruction lost 2 of 3 once a rival was installed | There is nothing installed in a chat prompt |
| The three coding agents disagreed on the leader in 10 of 18 categories | A chat model is a fourth thing again |
| Claude Code searched the web in 1.6% of decision runs | Content is absent from most of its decisions |
The equivalent metric for that surface is install share.
The reasonable position
Track AI visibility. It is cheap, it takes about an hour to set up, and knowing what an assistant says about you is worth knowing.
Do not treat it as a proxy for adoption. Measure the installs separately.
Common questions
What is AI visibility?
AI visibility is how often and how favourably AI systems mention a brand when they answer questions. It is measured by sending a fixed prompt set to chat models on a schedule and counting brand mentions, position and sentiment in the replies.
How is AI visibility measured?
Almost always by prompt monitoring: a tracked list of prompts is sent to several AI assistants at a regular interval, and the replies are parsed for brand names, position, sentiment and cited sources.
Is AI visibility a good metric?
It is a good metric for the chat surface and a poor proxy for anything else. For a developer tool it says nothing about whether a coding agent installed the product, which is a separate measurement.
What is a good AI visibility score?
There is no absolute scale. The number is only useful compared with your own history and with the competitors in the same tracked prompt set, because both depend entirely on which prompts were chosen.
Where this comes from
Armature ran 5,292 judged sessions with Claude Code, Codex and Cursor inside 51 realistic codebases, and published every run. The numbers on this page come from that work.