“AI visibility” is not one number from one prompt. Ask an engine the same question twice and you can get two different answers, and the answer in a consumer app can differ from the answer through an API. A useful measurement therefore needs a method: fixed questions, fixed engines, the same analysis every time, and a trend instead of a snapshot. This guide describes that method. It is the one askvisibly uses, so where we cite numbers, they come from our own scoring rules.
Three signals per answer
For every answer an engine gives, record three things.
- Mention. Your brand name, one of its aliases or your domain appears in the answer text.
- Citation. At least one cited source is a page on your domain.
- Position. Your rank among the brand-like names in the answer, counting from 1. The first name gets the most attention, so being fifth is not the same as being first.
Keep mention and citation apart. A mention without a citation usually means the engine knows you from memory or from third-party pages. A citation without a mention means your page was used as a source but the answer never said your name, which is common for documentation and how-to pages. Each pattern calls for a different fix.
Build a prompt set that mirrors buyers
Write questions in the words a buyer would use, across five intents:
- Category: “best invoicing software for freelancers”.
- Comparison: “Tool A vs Tool B for small teams”.
- Alternative: “alternatives to Tool A”.
- How-to: “how do I send recurring invoices”.
- Brand: “is your brand good for agencies”, “your brand pricing”.
Leave your brand name out of the category, alternative and how-to prompts. That is the test that matters: does the engine bring you up on its own? Brand prompts answer a different question, which is whether the engine describes you correctly. Keep the set stable between measurements. If you change the prompts, you are no longer comparing like with like, so add new prompts as a separate group.
Count answers, not prompts
One prompt asked of one engine is one answer. Ten prompts on two engines make 20 answers, and the rates below are shares of those answers. Ask several engines, because they disagree. A brand that is named by one engine and missing from another has a specific problem to fix, and an average would hide it.
From signals to a score
You need one number to track, as long as you remember it is a summary. We use a weighted blend: 0.5 for the mention rate, 0.3 for the citation rate and 0.2 for position, scaled to 100. The position part gives 1 for first place and loses a fifth for each rank below, reaching 0 at sixth place. In a worked example with 10 prompts on 2 engines, a brand mentioned in 10 of 20 answers, cited in 4 and averaging 0.6 on position scores 43. The weights are a choice. We weight being named most because an unnamed brand cannot be recommended. The full formula is on the how it works page.
Read the results in slices
- By intent. Category prompts show unprompted visibility. Brand prompts show whether descriptions are accurate. A strong brand score with weak category scores means people who already know you find you, and people who do not know you cannot.
- By engine. Differences between engines point to where your presence is thin. Each engine uses its own index and its own rules for when to search, as we explain in how engines choose sources.
- By competitor. Share of voice is your mentions divided by your mentions plus your competitors’ mentions. It tells you who the engines name instead of you, which is the list of pages and products you are competing with in the answer.
- By cited page. Look at the URLs engines cite. They show which of your pages work, and which third-party pages the engines trust for your category.
Handle variance
A single run is a sample. The same prompt can produce a different list tomorrow, and engines change models, search providers and answer formats without notice. Scan on a schedule, compare the same prompt set over several runs and look for moves that persist before you react. A one-off drop is noise until it repeats. A drop that appears on several engines at once often has a cause you can find, such as a page that went offline or a crawler that got blocked. The AI crawler check catches the second one in seconds.
Pitfalls that make numbers meaningless
- Measuring only with prompts that contain your brand name.
- Counting every mention as good news. Read the sentence around it, because a mention can be critical.
- Changing the prompt set and comparing the new score with the old one.
- Mixing results from an API with screenshots from an app and treating them as one series.
- Chasing one engine’s quirk. Fix the causes that show up across engines first.
A routine that works
Scan weekly or monthly with the same prompts. Review the prompts you gained and lost since the last run. Pick one gap, such as a missing comparison page or an outdated pricing page, fix it, and scan again. If you want to see where you stand right now, run a free scan. If you want the repeating version with history, alerts and more engines, see the plans.