Ask the assistants yourself. Log what they said.
Our own definitional page invites you to do this by hand: write the questions a real client would ask, run each several times across the assistants, classify every answer, divide. This is the tool for that — it holds the questions, records the answers, and computes your Case Recommendation Share with the run floor and the denominator rule enforced for you.
Your questions, your firm's name and your rows never leave this browser. No account is needed. Briefly counts page views, answers logged and CSV exports to understand which tools are useful. These counts include nothing from your log.
Reading the log saved in this browser…
This tool needs JavaScript. Your log stays in this browser. We count page views, answers logged and CSV exports, without collecting anything from your log.
The classification ladder
Every successful answer lands on exactly one of five states for the firm you are tracking. These five labels and definitions are the ones published on the Case Recommendation Share definition, used verbatim here so the tool cannot classify differently from the page that defines the metric. Only the top two rungs count.
| State | What it means | Counts toward CRS? |
|---|---|---|
| Absent | the firm does not appear in the answer at all. | No |
| Mentioned | the firm’s name appears, but not as a suggestion. “Unlike larger firms such as X…” is a mention. | No |
| Cited | the firm’s site or profile appears as a source, but the answer text doesn’t put the firm forward. | No |
| Recommended | the answer puts the firm forward as an option to contact. “Firms like X and Y handle these cases.” | Yes |
| Primary recommendation | the answer’s lead suggestion: named first, singled out, or explicitly preferred. | Yes |
A sixth option in the tool, failed or refused, is not a ladder state. It records a timeout, a refusal, an error, or an empty answer — an event that says nothing about your firm at all.
How the number is computed
Case Recommendation Share is recommended answers divided by successful answers: (recommended + primary) ÷ successful answers. Three rules decide whether that division is allowed to happen at all, and the tool applies them whether you want it to or not.
- Failed answers are excluded from the denominator entirely. A failed scan contains no information about your firm, and counting an outage as “not recommended” manufactures a drop that reverses the moment the outage ends. They are shown as their own count beside the share, so the exclusion is visible rather than convenient.
- Below 3 runs per question per assistant, no share renders. Not a small number, not an asterisked one — an em-dash, plus a count of how many more runs the figure needs. Identical reruns of the same hiring question do not reliably return the same firms, and cosmetic rewording moves the answer further, so a one-run verdict is a coin flip wearing a percent sign. The per-question table shows every pair's run count, so you can see exactly which one is short.
- Every figure carries its spread across runs. The tool computes the share for each run index separately and reports the lowest and highest beside the headline. A single figure never appears alone.
One consequence is worth stating plainly, because it is the whole reason the gates sit in the arithmetic and not in the copy: a 0% on this page can only ever mean measured zero — the floor was met, real answers were logged, and none of them recommended your firm. Anything not measured renders as an em-dash. A row you never filled in contributes to no denominator anywhere.
This tool applies the same formula, the same ladder and the same denominator rule to your own rows. The worked example — successful answers, excluded failures, and the share that falls out of them — lives on the definition page, alongside the competitor comparison that explains why recommendations and mentions are not the same event. Its figures stay there, on the page that owns them: this one renders nothing it did not compute from your rows.
One caveat, since you may well try it: that example is stated as 40 tracked questions answered once each. Typed in here as 40 single runs it is below the run floor, so this page would show an em-dash and a deficit rather than its 34.2%. Lay the same 40 outcomes out as 8 questions run 5 times each and the totals — and the share — are identical. The gap is the floor doing its job, not the two pages disagreeing.
What this tool does not do
- It does not measure anything for you. Every number on this page is arithmetic on your own observations. You ask the questions, you classify the answers. Briefly generates none of them and receives none of them.
- It is a snapshot of your own effort, not a trend. Manual logging happens when you have an afternoon. Assistants change without notice, and a model update can move every firm's share overnight. This tool makes no freshness promise and cannot tell you when your number went stale.
- It cannot settle a borderline classification. “You might look at firms such as X” sits right on the mentioned/recommended line. The rules are published so that a disagreement is about the rules rather than about hidden behaviour — but you are the one applying them, and consistency across your own rows is on you.
- It says nothing about the quality of any lawyer. Being named by an AI assistant reflects coverage and sources, not the quality of representation. This is a measurement of what assistants say. It is not a ranking, not an endorsement, and not legal advice.
- It does not approximate a clean room. Assistants personalise by location and account history. Ask logged out, in a fresh window, and treat the result as your own context rather than a neutral one. Our methodology describes what we do about this on the tracked side, and where it still falls short.
- It cannot recover your log. One browser, one key, no copy anywhere. Clear your site data and it is gone, and there is no version of it for us to restore, because there is no version of it we ever had. Export the CSV if the log matters.
The same measurement, without the afternoon
If you'd rather not do this by hand, the free check runs three of these questions across five assistants and shows you the actual answers — whether your firm is named, which firms are named instead, and which sources shaped the answer.
Run the free check →No signup. No credit card. No result held behind an email address.
John Rice builds and operates the scan engine behind Briefly, which runs client-style lawyer-hiring questions across ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode on a recurring schedule and stores every answer it collects. He is not a lawyer; he measures what AI assistants say, with receipts. The formula, the classification ladder and the denominator rule implemented above are the published ones, not a variant — which is the point of handing you the calculator for our own headline metric. Methodology · Case Recommendation Share · More on AI recommendations · About John
Briefly measures what AI assistants say. It does not rank, rate, or endorse attorneys, and nothing here is legal advice.