What is AI visibility?
AI visibility is how often, how prominently and how accurately a brand appears in answers that AI engines give to the questions its buyers ask.
The metrics
| Metric | Definition | Formula |
|---|---|---|
| Mention rate | How often answers name your brand | answers naming you ÷ all answers |
| Citation rate | How often answers cite your domain as a source | answers citing your domain ÷ all answers |
| Recommendation rate | How often you are presented as an option to choose | answers recommending you ÷ all answers |
| Share of voice | Your part of all brand mentions in the tracked set | your mentions ÷ mentions of you + tracked competitors |
| Position | Where you appear when the answer lists more than one brand | average rank among listed brands |
| Accuracy | How often the facts stated about you are correct | correct mentions ÷ all mentions |
We report every metric per engine and per market. We do not blend engines into one score, because a blend hides where the problem is.
How do we build the prompt set?
- Sources: your sales and support questions, your search console queries, reviews, and community threads in your category.
- Intents: category discovery, comparison, problem, brand and local prompts (see the audit method).
- A fixed core set that never changes within a reporting period, so trends are comparable.
- A smaller exploratory set that rotates to catch new questions.
- Versioned: every change to the set is logged with a date and a reason.
How often do we run it?
| What | Default frequency |
|---|---|
| Free audit | 30 prompts × 5 engines × 3 runs = 450 answers, once |
| Core prompt set (50 to 150 prompts) | Weekly |
| Full prompt set, all engines | Monthly |
| Runs per prompt per engine | 3 |
| Full re-run after a model update | As soon as the update is detected |
How do we keep runs comparable?
- Clean sessions with no personal history, where the platform allows it.
- A fixed language and region per market.
- The date, platform, surface and model name (where the platform shows it) are stored with every answer.
- The same classification rules every time, with a sample checked by a second person.
What do we do when a model is updated?
- Detect. From the vendor's release notes, or from a sudden shift across the core set.
- Re-run. Run the full prompt set on the new model.
- Mark the break. The trend chart shows a break, and we do not compare numbers across it without saying so.
- Re-map sources. Check whether the new model cites different domains for your topics.
- Adjust the plan. Change next month's priorities if the source map changed.
Limitations
- Answers vary between runs and between users; our rates describe a sample, not every answer anyone sees.
- Personalisation and logged-in history can change answers in ways a clean session does not show.
- We cannot see inside the models. When a number moves, we separate what we can show from what we suspect.
Measurement FAQ
Why run the same prompt 3 times?
AI answers are not deterministic. The same question can name different brands on different runs. One run can mislead; a rate over 3 runs is steadier, and we can raise the number for volatile prompts.
Do you use the APIs or the consumer apps?
We record which surface each answer came from. Consumer apps and APIs can answer differently, so we do not mix them in one metric.
Can I see the raw answers?
Yes. Every report includes the prompt list and the raw answers behind each number.
What if our numbers go down?
We report it, with the likely reasons we can support with data, and say plainly where we do not know.