How we measure AI search visibility

Our measurement process, from prompt panel construction to share-of-voice scoring. Open process, no black box. If you are evaluating any AI visibility service, these are the questions worth asking.

Why measurement is harder than it looks

AI assistants do not publish rankings. There is no equivalent of a Google SERP position or an impression count you can pull from a dashboard. The only way to know whether a brand gets recommended is to ask the questions buyers ask and read the answers.

That makes the prompt panel the core of any credible AI visibility measurement. What questions you ask, how you phrase them, which engines you run them against, and how you interpret the results all determine whether the output is useful.

Step 1: Building the prompt panel

We also include a smaller set of branded prompts, which ask directly about the client's brand by name. Branded prompts are useful for understanding reputation and recall. But organic prompts are the ones that determine new customer acquisition, which is why we weight the audit toward them and report the two groups separately.

A panel typically includes 50 to 100 prompts. The mix depends on the category: in local services, geographic variations matter and we generate prompts for multiple query framings. In B2B software, use-case and company-size variants are more important.

Step 2: Running the panel against four engines

We run every prompt against four AI engines, each using live web search:

Web grounding is important. Without it, AI engines answer from training data alone, which may be months or years out of date. With it, they pull current information from the web before forming their answer, which is how most users actually use these tools for buying decisions.

We do not use consumer-facing interfaces. We use the APIs directly, which gives us repeatable calls and the explicit citation data that is hard to pull out of a browser session.

Step 3: Scoring each response

For every response, we record four things:

• Was the brand mentioned? (Any appearance, including passing references)
• Was the brand recommended? (Named as a specific suggestion or top option)
• Which competitors were mentioned or recommended?
• Which URLs did the engine cite as sources?

From these, we calculate share-of-voice metrics for each brand in the comparison set. Mention rate is the percentage of prompts where the brand appeared. Recommendation rate is the percentage where it was specifically recommended. Both are reported for the full panel and for organic prompts separately.

Step 4: The organic vs branded split

Every prompt is tagged as organic (no brand name in the query) or branded (the client's brand name appears in the query). We report all metrics separately for each group.

This split matters because branded and organic queries serve different purposes. A buyer who asks "Is Company X good?" already knows Company X exists. They are checking a specific option. An organic buyer-intent query comes from someone who has not decided yet, which is the higher-value moment to appear in.

We have seen a brand post an 80 percent-plus branded mention rate and near-zero organic visibility in the same audit. Without the split, the headline number looks acceptable. With the split, the gap is right there.

Step 5: Citation source analysis

For every competitor-win response (where the target brand is absent), we extract and categorize the cited sources. This produces a ranked table of the URLs that are driving competitor recommendations in the category.

This is the most actionable part of the audit for most clients. The citation table directly answers the question: "Why is my competitor getting recommended and not me?" Usually the answer is that the competitor appears on specific aggregator or directory sites that AI engines treat as authoritative for the category, and the client does not.

What we report

Every audit produces:

• Overall share-of-voice vs up to five competitors, with organic and branded splits
• Prompt-level breakdown: which specific questions you lost, which you won, and which engines made the difference
• Citation source table: which URLs drove competitor wins, ranked by frequency, with the gap between competitor-win citations and your own citations per source
• Priority action list: specific recommendations tied to audit data, ordered by expected impact

There is no generic advice unconnected to the data. Every recommendation cites the specific finding that motivates it.

Limitations we disclose

AI engine results are not deterministic. The same prompt can return different recommendations on different runs, especially across engines. We run each prompt once per engine, which means our results are a snapshot rather than a statistical certainty. Panels of 50 or more prompts reduce variance enough to identify real patterns, but a brand appearing in 60 percent of prompts and a brand appearing in 58 percent may not be meaningfully different.

We also note the gap between API and consumer interfaces. API calls with structured parameters can differ from what a user sees in a browser, particularly for engines where the consumer interface includes features not exposed in the API. We use the API because it gives us structured, reproducible output, but results can vary from what a buyer sees when they type the same question in their browser.

We disclose both of these limitations in every report. If an AI visibility service does not mention them, treat the numbers with caution.