Whether AI assistants recommend your firm is now a question every personal injury owner asks, and the market has responded with two bad answers: vendors selling "AI visibility scores" nobody can define, and screenshots of a single lucky prompt presented as proof. Both are theater. The honest answer is a protocol: a repeatable, logged, boringly consistent way of asking the machines what they tell injured people, run on a schedule, read with discipline. This piece is that protocol, start to finish, including the limits no one selling a dashboard will tell you about.
AI answers are generated fresh, shaped by phrasing, location, history, and model updates nobody announces. That makes them unlike rankings, which at least pretend to be a stable list. Any single answer is weather. What a firm needs is climate: the stable pattern across many askings, which only emerges from repetition under controlled conditions.
This is also why the honest version is cheap and the dishonest version is expensive. The dishonest version buys a proprietary score that cannot be audited. The honest version is a spreadsheet, a fixed set of prompts, and a monthly hour. The machines answer anyone who asks; the edge is asking systematically.
Fix the prompts. Write a set that mirrors the real query path, in the words claimants use: the quiet questions (do I have a case if I was partly at fault in your state), the process questions (how do contingency fees work), the comparison questions (how should I choose an injury lawyer in your city), and the direct asks (best car accident lawyer in your city; who should I call after a truck crash near a named landmark). Ten to twelve prompts. Freeze the wording; the test is only a test if the inputs hold still.
Fix the engines and the conditions. The major assistants plus the AI layer of search itself. Fresh sessions, no logged-in history where avoidable, your market's location context. Note engine and date on every run, because models shift and your log is only evidence if it can show when.
Run and record, verbatim. For every prompt: which firms were named, in what order, what was said about them, which sources were cited or leaned on, and whether any answer was simply wrong. Copy the text; do not summarize from memory. Memory launders.
Score three things only. Presence: was the firm named at all, per prompt, per engine. Framing: when named, was the description accurate and would it help a frightened reader act. Sources: what kinds of pages the answers leaned on in your market, because that list is your placement map, the record of where the machines go to form opinions about firms like yours.
Repeat monthly, same prompts, same conditions. The first run is a baseline. The pattern across runs is the finding. Anything that moved is a question worth asking; anything stable is your actual climate. Add a prompt only by versioning the set, old prompts keep running unchanged, new ones start their own history, so the log never compares answers to questions that quietly shifted underneath them. An unversioned prompt set is how teams accidentally study their own edits instead of the machines.
Three reading disciplines separate evidence from anecdote.
Never read one answer as a verdict. A single naming is luck; a single absence is noise. Patterns across prompts, engines, and months are the only unit worth reacting to, and the log exists precisely so your reactions have denominators.
Expect the directory pattern and read it correctly. In market after market, the generic asks get answered from aggregator-shaped sources rather than individual firms. The wrong reading is despair. The right reading is a map: those sources are where corroboration currently lives in your market, and the early-path prompts, the quiet questions, are where individual firms still get named on the merits of their own answers. The protocol shows you exactly which prompts are winnable and which are currently rented to aggregators.
And resist optimizing for the test. The goal is not to engineer one prompt's answer; it is to learn what record the machines currently corroborate about your market, then improve the record: the entity consistency, the question coverage, the named attorneys, the sources that mention you. Answers follow records. That is the entire mechanism this series keeps returning to.
The protocol doubles as competitive intelligence at zero extra cost: the same log that records whether you were named records who WAS. Read that column seriously. For every competitor the machines keep naming, go look at what record earned it: the pages the answers cited, the question coverage those firms hold, whether their attorneys exist as consistent, verifiable people across the sources the engines leaned on. You are not looking for tricks, because there rarely are any; you are looking at which parts of the public record the machines in your market actually corroborate from, demonstrated by the firms currently winning it. The gap between their record and yours, on exactly those surfaces, is the work list. And when a named competitor's citation turns out to rest on a directory membership or a stale aggregator page, that is intelligence too: it tells you how shallow the current winner's moat is, and how quickly complete coverage could rearrange the answers.
A log nobody reads is a hobby, so close the loop the same way the attribution chain closes it: one page, once a quarter, in front of whoever owns the marketing budget. The brief answers four questions in plain sentences. Where are we present that we were not last quarter, and what changed in the record to cause it? Where are we still absent, and which specific work item does each absence route to? What did the machines get WRONG about us, and has the source of the error been fixed at its origin? And what moved in the cited-sources list, because the places the engines draw from are themselves a moving market. Four answers, with the dated log behind them as evidence, and the firm has something no dashboard sells: a governed, auditable account of its standing in the channel everyone is guessing about. The brief takes an hour to write from a well-kept log, and it is the difference between monitoring AI visibility and merely worrying about it.