Executive Summary
A law firm's visibility in AI answers is not a single rank. A system can mention the firm without recommending it, recommend it without linking to it, or cite a useful article without naming the practice. Those outcomes answer different marketing questions. A defensible measurement program records them separately and connects them to the client situations the firm is qualified to serve.
This paper proposes a repeatable evaluation framework for marketing directors, agencies, and SEO/GEO practitioners. It is a methodology synthesis, not an original market study. No law firms were tested for this article, and no improvement rate or customer outcome is reported. The practical recommendation is to maintain three parallel records: a controlled answer benchmark, platform-reported website visibility, and observed referral activity. Agreement strengthens a diagnosis; disagreement is information to investigate, not a reason to blend everything into a proprietary score.
Start with a modest, documented set of client scenarios. Preserve the exact prompts and test conditions. Define what counts before collecting answers, retain the underlying evidence, and report missing observations. The resulting report should let a skeptical partner understand what changed, reproduce selected tests, and decide which website or authority problem deserves attention next.
01. Define the decision you are measuring
Begin with the commercial question, not a tool subscription. An immigration practice might want to understand whether prospective clients can discover its family-based immigration services in the cities it actually serves. A personal injury firm may care about local representation after a serious collision. These are hypothetical planning examples, not client cases. Each involves different geography, urgency, qualifications, and information needs.
Separate discovery prompts from branded fact checks. Asking whether a named firm handles a service tests entity accuracy; asking which firms may fit an unnamed person's circumstances tests competitive discovery. Combining both can inflate a dashboard because a brand included in the question is easier to mention in the answer. Also separate informational queries from requests for a provider shortlist. A strong explanatory article can be cited without any attorney recommendation.
Document the unit of analysis as one prompt, one platform surface, one test condition, and one run. Do not treat a consumer ChatGPT answer, an API response, and a Google AI Overview as interchangeable observations. Google itself explains that AI Mode and AI Overviews may produce different responses and links. A useful report identifies the specific experience tested rather than assigning one universal position to “Google AI.” 1
| Evidence layer | What it supports | What it does not prove |
|---|---|---|
| Answer benchmark | Presence in the defined test panel | Population-wide preference |
| Platform reporting | Recorded website exposure | An explicit firm recommendation |
| Referral analytics | Identifiable recorded visits | All AI-influenced demand |
Original editorial framework. This exhibit contains no observed client results, performance statistics, or market estimates.
02. Build and freeze a scenario panel
Create scenarios with the firm's marketing and intake teams using nonconfidential patterns, not copied client narratives. Record practice area, service location, language, decision stage, and relevant constraints. Include situations the firm would decline as well as those it wants. A firm appearing for an unsuitable matter can create intake waste rather than commercial value.
For each scenario, write a neutral prompt in ordinary client language. Avoid supplying the desired answer, the firm's marketing slogan, or a competitor list unless the test explicitly studies those inputs. Review the panel for geographic and practice-area imbalance. Twenty variations of one profitable service do not create twenty independent coverage categories. The number of scenarios is a design choice to justify against budget and business coverage, not an industry benchmark.
Freeze a dated baseline panel before remediation. Keep later experiments in a separate exploratory panel. Otherwise, changing the questions after seeing disappointing results can manufacture apparent improvement. If a prompt needs correction, preserve the previous version and mark the comparison as interrupted. Set a review date for the panel itself so that business expansion, office closures, and changes in accepted matters are reflected deliberately rather than silently.
03. Record test conditions and preserve evidence
For every run, capture the prompt verbatim, date and time, platform, interface, visible model or mode where available, language, and relevant location settings. Note whether the session is signed in, whether memory or conversation history may influence it, and whether web search was used or visibly available. Some variables cannot be fully controlled; record that limitation instead of calling the test objective by default.
Use a fresh conversation when the protocol calls for an independent discovery test. Keep follow-up journeys as a separate evaluation because prior answers can shape later responses. Save the full answer, cited URLs, and a screenshot or export where permitted. Retain an evidence identifier in the results table so another reviewer can inspect the original response. Follow platform terms and applicable access restrictions; a measurement plan does not authorize automated scraping.
Repeat observations under a consistent schedule and preserve failures. A timeout is not a negative recommendation. A successful response with no firm mention is a valid absence under your coding rules. A refusal is another outcome. Decide in advance how each affects the denominator. Repeated runs help expose variability, but correlated answers from the same prompt should not be represented as independent consumers or as a representative sample of the legal market.
04. Use explicit metrics rather than an opaque score
A practical coding sheet distinguishes mention, qualified recommendation, owned-domain citation, third-party citation, factual error, and no relevant answer. Define a qualified recommendation narrowly enough for two reviewers to agree. For example, a firm named only in a warning about unsuitable representation should not count as a positive shortlist inclusion.
Calculate a mention rate as qualifying runs containing the firm divided by valid runs for the same reporting segment. Calculate a recommendation rate separately, using the recommendation rule. Show numerator and denominator beside every percentage. A citation rate should specify whether it concerns the firm's domain or any source mentioning the firm. If several firms appear in an answer, explain whether each receives one inclusion or whether a fractional allocation is used. There is no reason to label an arbitrary allocation as market share.
Report errors as their own work queue. Incorrect service descriptions, locations, attorney identities, or contact details can matter even when mention frequency rises. Do not hide them within a composite score. Where an answer presents an ordered list, preserve the order as an observed feature of that response; do not assume the order is a stable ranking or a measured preference across users.
05. Reconcile platform reports, answers, and visits
Google announced dedicated Search Generative AI performance reports in Search Console in June 2026, with a subsequent worldwide rollout note. Use the reports available in the firm's property and retain the report name, filters, date range, and export. A website impression reported by a platform is not automatically an explicit recommendation of the law firm. Keep this evidence separate from manually coded answers. 2
Website analytics answers a different question: which recorded visits arrived with identifiable source information? Review landing pages, referral domains, and observed campaign parameters. Validate actual incoming URLs before creating channel rules. Do not assume every AI-influenced visit carries a recognizable referrer, or that every visit from an AI product represents a new legal prospect. A colleague checking an article can produce traffic too.
Server logs provide another diagnostic layer. OpenAI distinguishes OAI-SearchBot, used for search, from GPTBot, associated with model training, and ChatGPT-User, associated with certain user-triggered actions. A crawler request is not proof that the firm was recommended or that a prospect visited. Keep crawler access evidence in the technical audit, not in a lead count. 3
06. Interpret change without overclaiming causality
Compare like with like: the same prompt version, platform surface, coding rule, and reporting window. Display changes by practice area and geography before producing a portfolio total. If the platform changes its interface or visible model, annotate the time series. If testing coverage expands, show the original panel separately so the denominator does not move invisibly.
Use a second reviewer on a defined subset of answers and resolve disagreements with written examples. This is especially valuable for implicit endorsements and ambiguous firm names. Preserve an unresolved category when the evidence cannot support a confident classification. An error bar computed from an inappropriate independence assumption does not make the underlying collection representative.
A before-and-after improvement following website edits is a temporal association. Competitor activity, platform updates, seasonality, and prompt variability can also explain movement. A matched comparison or staged rollout can strengthen an evaluation, but the assumptions still need disclosure. In a management report, prefer “observed in the repeated test panel” to “the market now recommends us.” The first claim describes the evidence; the second extends beyond it.
07. Methodology, scope, and a usable reporting package
This framework draws on the official documentation listed below and adds an editorial recommendation for operational measurement. It does not estimate platform usage, legal consumer behavior, or the effectiveness of 360AI Legal services. No benchmark dataset accompanies this draft. All proposed fields and workflows are examples for teams to adapt and approve.
A reproducible delivery package contains the scenario register, frozen prompts, collection log, full response evidence, coding definitions, reviewer decisions, calculation sheet, and a change log. Include a coverage statement listing untested platforms and inaccessible features. Record the reporting date separately from the date of each observation. An agency should make this package understandable without requiring the client to retain the agency's proprietary dashboard.
Before commissioning a larger measurement program, run a pilot to identify ambiguity and access constraints. Use it to improve the protocol, not to advertise a performance result. Agree who can see stored answers, how long evidence is retained, and how sensitive information is excluded. A public marketing benchmark should never require exposing an actual prospective client's account of a legal problem.
08. Key findings and business implications
The central analytical finding is that recommendation presence, website exposure, referral traffic, and acquisition are different measurement layers. None is a reliable substitute for all the others. A firm's report should state which layer it observes and which business decision it can support. Technical access problems call for engineering review; factual errors call for evidence correction; unsuitable inquiries call for positioning and intake review.
For a marketing director, the immediate priority is an auditable baseline and a small set of actionable defects. For an agency, it is a stable protocol and transparent client reporting. For a partner approving spend, it is a clear boundary between observed evidence and expected future benefit. No method can honestly guarantee that an AI system will recommend a particular firm.
A useful next step is to review the technical foundations that make the firm's public information accessible and coherent, then connect observed discovery to privacy-conscious intake measurement. Treat forecasts about more conversational discovery as planning hypotheses. Investment decisions should remain revisable as actual evidence accumulates, rather than depending on a prediction being presented as a verified outcome.
Sources & references
Official documentation reviewed November 13, 2026. Platform features and policies can change. The linked sources support the cited facts; proposed workflows are editorial analysis.
- Google Search Central: AI features and your websitehttps://developers.google.com/search/docs/appearance/ai-features
- Google Search Central: Search Generative AI performance reportshttps://developers.google.com/search/blog/2026/06/gen-ai-performance-reports
- OpenAI: Overview of OpenAI Crawlershttps://developers.openai.com/api/docs/bots



