How to Measure AI Search Visibility

The first instinct is to open an assistant, ask "what are the best tools for X", and screenshot the answer. That tells you almost nothing. Ask again an hour later and the list changes. Measurement of answer engines has to be statistical from the start.

Build the prompt set first

Your measurement is only as good as the questions. A usable set has three properties:

• Neutral. No brand names. "Best AI visibility tool" is a fair question; "is Reflyma good" is a leading one and will flatter you.
• Realistic. Phrased the way a buyer speaks, in full sentences.
• Spread across intent. Category discovery, comparison, use-case fit, pricing, integration, and problem-first questions where your category is the answer but not the subject.

Thirty to sixty prompts per topic is enough to see structure. Fewer than fifteen and your numbers swing on noise.

Sample properly

Run every prompt across every engine you care about, multiple times, on a fixed schedule. Record the full response text, the engine, the model version, the timestamp and any cited URLs. Without the raw response you cannot audit the metric later, and an unauditable metric will not survive its first challenge from a stakeholder.

The core metrics

Mention rate

The share of runs where the brand appears at all, counting aliases and common misspellings. This is the base rate everything else conditions on.

Average position

Where in the answer the brand first appears — first recommendation, mid-list, or trailing alternative. Position separates "we are visible" from "we are the default suggestion".

Citation rate

How often your own domain is linked as a source. Mention without citation means the model knows of you from elsewhere; citation means your pages are in the retrieval set. These are different problems with different fixes.

Engine coverage

How many of the tracked engines mention you at least once. Concentration in one engine is fragile.

Share of voice

Your mentions as a proportion of all brand mentions across the prompt set. This is the number that tells you whether you are gaining ground or the whole category is.

Detecting mentions without fooling yourself

Naive substring matching produces both false positives (a brand name that is also a common word) and false negatives (possessives, casing, alternate spellings, product names used instead of the company name). A reliable pipeline does deterministic alias matching first, then uses a model to classify sentiment, position and context, and finally validates the model's output against the deterministic result. Where they disagree, the deterministic match wins and the case gets flagged.

Composite scores: useful, with a caveat

A single visibility score is good for tracking a trend and bad for diagnosis. If you use one, publish the formula and the weights, and always show the components next to it. A score nobody can decompose is a score nobody will trust when it drops.

Reporting that survives scrutiny

1. Report rates with the run count, not raw counts.
2. Segment by engine — the aggregate hides the interesting part.
3. Link every number back to the underlying responses.
4. Compare period over period on an unchanged prompt set. Changing prompts mid-quarter invalidates the trend.
5. Track competitors on the same prompts, or you have no baseline for what "good" is.

Done this way, AI visibility stops being anecdote management and becomes something you can put in a board deck.