AI visibility audit for law firms

Not keywords. Each question is a client situation for your practice area, written the way a person describes it: what happened, how serious it is, where they are, what they are worried about, set in your own market. The set is fixed and versioned, so the same questions can be asked again later and the answers compared.

The question never mentions your firm, and no list of competitors is supplied. The assistant has to find the firms itself — which is the only way to learn whether it would find yours. Situations are available for 12 practice areas: Personal Injury, Family, Criminal Defense, Immigration, Estate Planning And Probate, Employment (Employee Side), Bankruptcy And Debt Relief, Real Estate (Consumer), Disability Benefits And Workers' Compensation, Small-Business Law, Tax Resolution, Tenant Rights. One scan measures one practice area.

Being mentioned is not the same as being recommended. An assistant can name your firm in a list and still send the client somewhere else, so the audit counts three things and keeps them apart:

AI Discovery Index

How often AI names your firm in the completed test answers.Answers naming your firm ÷ completed test answers × 100.

AI Selection Index

Of the test answers that name your firm, how often AI recommends it first.First recommendations ÷ test answers naming your firm × 100.

Recommended first overall

How often AI recommends your firm first across all completed test answers, not only the answers that name it.First recommendations ÷ completed test answers × 100.

Every figure is printed with the counts it was divided from, and a figure with nothing to divide by reads as not measured, never as zero. How each one is counted.

For every answer that did not recommend your firm first, the audit records who it did recommend. Across the scan that becomes the list of firms AI actually puts in front of your prospective clients — how often each was named, and how often each was recommended first. These are the assistants’ picks, not your guess at your competition and not a directory’s list.

An answer that recommends nobody — it declined, found nothing, or stopped at a list — is counted on its own, never as a loss to a rival.

A score you cannot check is an opinion. For every answer, the report keeps:

• the client situation exactly as it was asked;
• the AI model that answered, and when;
• the searches it ran and the source links it returned;
• the full answer, unedited;
• which firms it named, and which one it recommended first.

The headline measures open into the answers that produced them, and the sources are totalled by website, so you can see where the answers looked and whether your own site was among them.

A re-test runs the whole benchmark again — the same situations, the same population — so a before-and-after compares the same questions’ answers rather than two different samples. If the two scans did not ask the same questions, the report says so and refuses the comparison instead of printing one.

What a re-test reports is an observed change, with its counts. It never claims that your website edit caused it: assistants change, the web changes, and a small sample moves.

• What one client saw. The models are called through their APIs with web search available. That is not the ChatGPT, Claude, Gemini or Perplexity app: an app can add memory, personalization and its own instructions that an API call does not have, and, where it shows them, ads.
• Google AI Overviews or AI Mode. These are features of Google Search and are not measured here.
• A position. There is no “rank #3 on ChatGPT” to read and no 0–100 score here; there are answers, counted.
• A promise. No one can guarantee that an assistant will recommend a firm. The audit measures and explains; it predicts no score and no gain.

Many things are sold under this name, and some of them never ask an AI model anything. Whoever you use, ask:

1. Does it put questions to AI models, or only inspect your website?
2. Does it show you the answers and their sources, or only a score?
3. Does it separate being named from being recommended first?
4. Does it show which firms were recommended instead — found by the AI, not taken from a list you supplied?
5. Does every rate come with the count it was divided from?
6. Can the same questions be asked again later, and will it say when they were not?
7. Does it promise results? A measurement cannot.

AgentSignal answers each of these on its methodology page, including what it deliberately does not do.