How AI answer engines decide which law firm to cite

Retrieval, not ranking: what actually happens when someone asks for a lawyer

When a person asks an assistant for a lawyer, the model does not consult a ranked list. It runs one or more searches against an index (Bing for ChatGPT and Copilot, Google for Gemini and AI Overviews, Perplexity’s own crawl plus licensed sources), reads the pages that come back, extracts the entities those pages describe, and composes an answer from the entities it can support with more than one source. Google calls the multi-search step query fan-out: a single prompt such as “best divorce lawyer for a high-asset case in Orange County” becomes several sub-queries about divorce lawyers, high-asset divorce, Orange County family law and reviews, and the answer is assembled from the union of results.

The consequence is that the engine is grading corroboration rather than position. A page that ranks first for “Orange County divorce lawyer” gets read, but so does an Avvo profile, a Super Lawyers listing, a Google Business Profile with 140 reviews, a Reddit thread and an Attorney at Law Magazine feature. The firm that appears consistently across that set is the one the model is confident enough to name. Ahrefs measured how loosely this tracks classic rankings in March 2026: only about 38% of pages cited in Google AI Overviews rank in Google’s top ten for the query, down from 76% in mid-2025. Roughly a third of citations came from pages ranked 11 to 100, and a third from beyond 100.

Why the engines disagree with each other

Because they retrieve from different indexes with different licensing. Martindale-Avvo analysed its own legal query data in April 2026 and found ChatGPT’s answers overlap with Google’s top ten results less than 25% of the time, Perplexity and Claude about 75%, and Gemini about 50%. ChatGPT reads Bing and its own crawler; Gemini and AI Overviews read Google; Perplexity retrieves live and has licensing deals that weight certain sources. A firm strong in Google will usually be visible in Gemini and AI Overviews and may be invisible in ChatGPT.

The one thing all engines agreed on in Citorian’s June 2026 test is the top of the list. Across ChatGPT, Perplexity, Claude, Gemini and AI Overviews, all ten of the most-named personal injury firms were named by all five engines. Consensus at the top and divergence below it is exactly what corroboration-based retrieval produces: firms with overwhelming records get named everywhere, and the marginal picks depend on which index the engine happened to read.

Which sources carry the citations for law firms?

Directories first, then Google, then community sources. Citorian’s five-engine personal injury test logged every source the engines cited across 359 answers. Super Lawyers carried 33% of citations, Justia 20%, Google 12%, Reddit 8%, Attorney at Law Magazine 7%, Best Law Firms 6%, Avvo 5% and YouTube 5%. 5WPR and Haute Lawyer’s April 2026 Legal AI Visibility Index reached the same conclusion at the top of the market: across eight practice areas, seven directories (Chambers, Legal 500, Super Lawyers, Best Lawyers, Martindale, Avvo, Justia) owned the citation layer, and even Am Law 100 firms ranked below the directories for queries about their own practice areas.

Two nuances. Reddit is read far more often than it is cited: Ahrefs data reported by Lexgro shows ChatGPT pulling Reddit pages in 67.8% of retrievals while naming Reddit as a source in 1.93% of answers. And the firm’s own site is almost never the citation for a “who should I hire” prompt. It is where the engine confirms what the directories said. A practice-area page that contradicts the Avvo profile (different office, different focus, a different attorney named) is a corroboration failure, not a ranking problem.

The three tests a firm’s record has to pass

Consistency, specificity and corroboration. Every engine applies them in some form, and a firm that fails one is filtered before the answer is composed.

• •Consistency. The firm name, address, phone, attorney names and practice focus have to match across the directories, Google Business Profile, the firm site and LinkedIn. A firm listed as “Smith & Jones Injury Lawyers” on Avvo and “Smith Jones LLP” on Justia reads as two weak entities rather than one strong one. SOCi’s 2026 index found consistent NAP data and a review response rate of at least 5% among the traits of the 1.2% of locations ChatGPT recommends.
• •Specificity. Engines answer specific prompts with specific firms. Citorian found a named firm in 79% of specific-injury prompts against 23% of general question-led prompts. A firm with one “personal injury” page competes for the general prompt; a firm with pages on truck accidents, rideshare crashes and spinal injuries is retrievable for the prompts where the engine actually names someone.
• •Corroboration. Independent sources have to agree. Reviews (recency more than volume; SOCi’s ChatGPT-recommended locations averaged 4.3 stars), directory listings, legal press and community mentions each add a vote. The Princeton generative engine optimization study (Aggarwal et al., KDD 2024) found that adding statistics, citations and quotations to a page raised its visibility in generative answers by up to about 40%, with statistics alone adding 22 to 41%. The engines reward pages that look like evidence.

Why a firm that ranks first can still be invisible

Because the passage the engine needs is not on the page, or the record around the page does not back it up. Ranking first for a query means Google’s classic algorithm scored the page highest. Being cited means an engine found a self-contained passage that answers the sub-query it generated and found other sources that agree with it. A ranking-first page written in the conventional law firm style (“At Smith & Jones, we understand how devastating an accident can be…”) has no extractable answer to “what is the statute of limitations for a car accident claim in Florida”, so the engine lifts the sentence from a competitor’s page that states it in one line.

The corroboration failure is subtler. A firm can rank first with a thin Google Business Profile, an unclaimed Justia listing and an Avvo profile last updated in 2021. The engine reads all of them, finds the record inconsistent and stale, and names the firm two spots down whose profiles agree. Forward Push’s informal test of ChatGPT recommendations in one city found the named firms shared depth per practice area, hundreds of Google reviews and a coherent identity across LinkedIn, YouTube, the directories and local news. None were the biggest advertisers.

What the engines do with attorneys, as opposed to firms

They resolve people. A client asks for a lawyer, and the engine tries to return one: a name, a firm, a practice focus, a reason. That means the attorney has to exist as an entity the engine can corroborate, separately from the firm. Bar admission records, a directory profile per attorney, a bio page with the same credentials in the same words, publications, speaking, and where the state allows it, matter history, all contribute. Firm AEO’s experience across engagements is that the attorney-level record is usually the half that is missing: the firm has profiles, the individual lawyers do not, and the engine names a solo practitioner down the street whose personal record is complete.

This is also where bar rules shape the work. Everything published about an attorney has to be demonstrably true under ABA Model Rule 7.1 and the state’s own advertising rules; specialist claims require the certification; past results are restricted or conditioned in several states. The compliant version of an attorney entity is credentials, focus and verifiable facts, which is also the version the engines corroborate most easily.

How to test your own firm in an afternoon

Run the prompts, log the names, read the sources. This is the baseline Firm AEO builds on the first call, and a firm can do the rough version itself.

• •Write five prompts a real client would type: one direct vetting prompt (“best [practice] lawyer in [city]”), two specific-matter prompts, one question-led prompt and one situational prompt.
• •Open ChatGPT, Perplexity and Google (for AI Overviews) in private windows with location set to your market. Run each prompt three times per engine. Log every run: which firms were named, in what order, and which sources the engine cited.
• •Open every source the engines cited. Note which directories, which review platforms, which press. That list is the retrieval pool for your market, and it is what you have to be present and consistent in.
• •Compare your own record against the named firms’ records on each source. The gaps are the work.
• •Repeat monthly. A single run is a sample, not a verdict; appearance rate across repeated runs is the only number that means anything.

What Firm AEO changes, in order

Entity repair first, because inconsistency is a filter and everything else is wasted until it is fixed. Then the directory and Google Business Profile layer, because that is where more than half of citations originate. Then answer-shaped practice-area pages, one per matter type, each opening with a self-contained paragraph that answers the sub-query an engine will generate. Then the review cadence, compliant and steady, because recency is a signal the engines read every month. Then third-party corroboration: legal press, community presence, comparison capture. Every asset is reviewed against Rule 7.1 and the state rules before it goes live under the firm’s name.

The measurement is an appearance rate: each of around fifteen agreed client questions run repeatedly every month across ChatGPT, Perplexity and Google AI Overviews, every run logged. The guarantee is written against it. If the firm is not named across the set by day 90, that period’s fees are waived. It is a guarantee about visibility and about Firm AEO’s fee, never about cases or clients, and Firm AEO will not write one that is.