There is no single AI recommendation list
An answer about divorce lawyers depends on the question, the product and the information available in that session. A web-enabled answer can retrieve current sources; an answer without live search may behave differently. No public document supplies a universal formula for getting a law firm recommended across every assistant.
For a family-law firm, the useful starting point is narrower: can a system access accurate information about the firm, and does that information answer the prospective client's actual question? Access is a prerequisite for many search features. It is not a promise of selection.
What the platforms actually document
Google's guidance explains that its generative search features draw on Search ranking and quality systems. Eligible pages must be indexed and available for snippets. Google recommends useful, original content and warns against mass-producing pages for query variations. This describes Google Search, not a universal rule for Claude or ChatGPT.
OpenAI's crawler documentation distinguishes search crawling from model-training crawling. Review OAI-SearchBot access when diagnosing ChatGPT search discovery; do not assume that a training-bot setting alone determines search access.
Anthropic documents separate bots for training, search and user-directed retrieval. Its guidance says restricting Claude-SearchBot can reduce search visibility. These controls govern access, not guaranteed recommendations. Review each provider's current documentation before changing crawler settings.
Start with a question a real client would ask
“Recommend a family-law firm” and “Who handles a contested divorce in my county?” ask for different help. The second contains a location and a matter type. A useful firm website explains the real jurisdictions served, the services offered and the people responsible for the work. It should not invent offices or capabilities to match a prompt.
Use two separate test groups. Discovery prompts should not mention your firm's name. Accuracy prompts should name the firm and ask about its lawyers, location or services. Combining the two can create a misleading report: an assistant repeating a brand you supplied does not prove independent discovery.
Build an evidence trail for each important fact
Our recommended working document is a simple fact register. Assign an owner, an authoritative source and a review date to each fact. It gives the marketing team a way to correct contradictions before turning them into more pages or videos.
- People: current lawyer roster, accurate biographies and verified credentials.
- Practice: matters the firm actually handles, with appropriate attorney review.
- Place: real offices, service area and consistent contact details.
- Process: what an inquiry or consultation involves, including any verified fees or eligibility limitations.
- Proof: authorized examples, awards with accurate context, and genuine third-party references.
For example, if a biography says a lawyer handles family law but a newer team page omits that lawyer, confirm the current fact with the firm before changing either page. More repeated text will not resolve an underlying uncertainty.
Separate accessibility, accuracy and selection
These are three different checks. Accessibility asks whether a crawler can retrieve the page. Accuracy asks whether the page and the answer describe the business correctly. Selection asks whether the firm appears in an answer to a relevant, unbranded question.
Record each separately. A page can pass a live crawl test and still be absent from an answer. A firm can be mentioned with an incorrect address. A source can be cited for general information without the firm being recommended. Each result calls for a different response.
A repeatable test record
Before editing the site, save a baseline using fixed prompts. For every run, record the exact prompt, date and time, product, visible model or mode, whether web search was used, relevant location context, and whether the conversation was fresh. Keep the full answer and its cited URLs, not just a favorable cropped screenshot.
- Mark whether the firm was mentioned without being named in the prompt.
- Record the answer's actual wording and whether it was a recommendation, comparison or passing reference.
- Check each claim about the firm against verified facts.
- Record links to the firm's site separately from third-party citations.
- Keep errors, unavailable products and login blocks in the log. Exclude unavailable runs from the valid-test denominator.
Report small samples honestly. Two mentions in ten completed runs describes those ten runs. It does not establish a market-wide ranking, a probability of future recommendation or ten new leads. Repeat comparable tests and disclose any changed conditions.
Our AI visibility research methodology
Method version 1 · October 20, 2026. This protocol defines how to collect and interpret observations. It does not report a completed study or measured client gains. We keep agency-discovery tests separate from tests of a law firm being discovered by prospective clients.
Set the questions before collecting answers
Freeze ten relevant, unbranded prompts for each audience and give each prompt a stable identifier. Include different buyer needs rather than ten minor rewrites of one question. Keep the exact wording unchanged during a comparison period. For a law firm, record the real location and matter type being tested. For an agency, test the services a firm is trying to buy. Branded fact-checking questions belong in a separate accuracy set.
Plan three fresh-conversation repetitions per prompt and platform where access permits. Record missing repetitions instead of quietly reducing the planned sample. Do not add the business name, its website or a favorable earlier answer to an unbranded discovery conversation. Record account, personalization and search settings as observed; do not assume a fresh conversation removes every source of personalization.
Keep a complete record for every attempt
- Conditions: prompt identifier and verbatim text; timestamp and time zone; platform; displayed model or mode; search setting; language; supplied location; account context and fresh-conversation status. Mark hidden or unknown settings as unknown.
- Evidence: the complete answer, a capture or accessible record, all cited URLs and the date each source was checked. Preserve unsuccessful attempts as well as favorable answers.
- Classification: whether the business was named, explicitly suggested as an option, or cited through its website; any third-party references; factual errors and the source used to check them.
- Run status: completed answer, no AI answer shown, refused, login blocked, rate limited, or technical error. Never substitute an invented answer for an unavailable run.
Score observations consistently
A mention names the business. A recommendation explicitly presents it as an option for the requested need. A first-party citation links to the business's own site. A third-party citation is recorded separately. These measures can overlap, but none substitutes for the others. A citation is not automatically an endorsement, and appearing first in one answer is not a stable ranking.
For each platform and fixed prompt set, report the number of completed answers and the counts of mentions, recommendations and first-party citations. A mention rate is answers mentioning the business divided by completed, usable answers to that unbranded set. Show the numerator and denominator alongside any percentage. Report blocked, refused and errored attempts separately. For Google searches, also report how often an AI answer appeared among successfully loaded searches; absence of an AI answer is a distinct observation, not a recommendation result.
Illustration only: if 30 attempts produce 27 usable answers, 3 login blocks and 6 mentions, report “6 mentions in 27 completed answers; 3 attempts blocked.” Do not call that six leads or a market-wide visibility score. Repeated prompts in one product are not independent samples of every prospective customer.
Review, compare and correct
Review every positive recommendation, ambiguous classification and factual discrepancy against the saved answer. A second reviewer should check published findings where available; record who reviewed them and flag single-reviewer work. Keep corrections dated, with the original observation retained.
Compare the same prompts, platforms and modes at the baseline and planned 30-, 60- and 90-day checkpoints. If a platform changes its model, search experience or access conditions, flag the break and show the affected results separately. Log website releases alongside test dates. A change after a release is an observation; it does not by itself prove that the release caused it.
Measure business outcomes separately through verified inquiry records and available analytics. Before publishing a case study, confirm the actual engagement, evidence, dates and permission to name the client. Our AI video concepts demonstrate creative formats and are not evidence of client performance.
Where useful video fits
A short, reviewed video can explain an approved answer in a format a prospective client prefers. Pair it with accurate page text or a transcript and clear context about who is speaking. An AI presenter should not be mistaken for a real attorney or client testimonial.
Video is a communication asset, not a shortcut to an AI recommendation. Our family-law video examples are labeled creative concepts. They demonstrate formats, not client outcomes. See our family-law AI video production approach for review and production scope.
For a step-by-step implementation checklist, read our practical AI visibility guide for family-law firms.
Choose the next fix from evidence
If access fails, investigate the technical cause. If facts conflict, correct the source and its legitimate profiles. If the page is accessible and accurate but offers little useful detail, develop an answer that reflects the firm's real process and experience. If a test cites an outdated third-party page, document the discrepancy and seek an authorized correction.
Our family-law AI visibility service brings these checks together. The combined visibility and video program connects reviewed written answers, supporting video and measurement. Neither promises a particular ranking or recommendation.
Editorial note: This guide combines the linked platform documentation with Kid in a Hoodie's recommended audit workflow. The workflow is a practical method, not a disclosed ranking formula or a claim of measured client results. Platform behavior and documentation can change.