A prospect with a case does not always open Google any more. A growing number of them open ChatGPT, describe what happened in their own words, and ask what kind of lawyer they need and who they should call. The assistant answers in a paragraph or two, names a handful of firms, and the conversation moves on. If your firm is not one of the names, you were never in the running, and no ranking report will tell you it happened.
Getting named is not luck, and it is not a setting anyone can switch on. It is the product of four things an assistant can actually check: whether it can identify your firm as a distinct entity, whether it can find independent sources describing you, whether your pages are written so a relevant passage can be lifted cleanly, and whether anything it finds contradicts anything else it finds. This is what that work looks like, what gets measured, and how long it takes.
What ChatGPT is actually doing when it names a firm
Two different mechanisms produce a law firm mention, and they behave differently.
The first is retrieval. When a prompt is specific or local, the assistant runs a search, reads what comes back, and writes an answer grounded in those pages, usually with citations. This is the mechanism you can influence in weeks rather than years, and it is the one that matters for "personal injury lawyer in San Diego" style prompts.
The second is the model's own parametric knowledge, the impression of your firm formed during training. That moves slowly, it is not directly editable, and it is downstream of how widely and consistently you are described across the open web. You do not optimize for it so much as earn it, and the same work that wins retrieval today is what shows up in the next training pass.
Nearly every practical gain in the first few months comes from the retrieval path. The goal is narrow and concrete: when the assistant searches on a prompt that matters to you, your page is in the result set, it is legible, and the sentence the assistant needs is sitting near the top of it.
Step one: become an entity the model cannot confuse
Language models resolve names to entities before they reason about them. If your firm's name is ambiguous, if it is spelled three ways across the web, or if two unrelated practices share it, the model has to guess, and a guessing model tends to name somebody else.
The fix is unglamorous. Pick one canonical firm name and use it exactly, everywhere, including the legal suffix or excluding it consistently. Make the name, address, and phone number identical on your website, your Google Business Profile, your bar listings, and every directory that carries you. Claim the profiles you have abandoned. Kill the duplicates. Where an old name or a merged practice still appears, make sure something on your site explains the relationship in plain words, because an assistant reading two names will otherwise treat them as two firms.
On the site itself, this means one Organization or LegalService node with a stable identifier, the firm's real founding details, and links out to the profiles you control. It is the same entity discipline that underpins Google SEO, which is why the two programs are not separable in practice.
Step two: be described by sources that are not you
An assistant weighs what other people say about you far more heavily than what you say about yourself. A firm whose only description exists on its own website is a firm with one source, and one source is not enough to be named with confidence.
What counts here is coverage that carries a description, not a bare link. A directory entry that lists your practice areas, a bar association profile, a local news piece that quotes you on a case type, a trade publication that credits your commentary, a podcast page with a real summary. Each of these gives the model an independent sentence about what you do. Several of them agreeing with each other is what turns a plausible answer into a confident one.
This is the part of the work most firms underinvest in, and it is why editorial placement does double duty. Coverage earned for credibility with human readers is the same coverage an assistant uses to decide whether you are worth naming.
Step three: write pages a passage can be lifted from
Retrieval-based answers are assembled from passages. A page that buries its answer under four paragraphs of throat-clearing loses to a page that answers in the first two sentences, even when the buried page is better.
In practice that means a page per question people actually ask, a heading phrased the way they ask it, and a direct answer of roughly forty to sixty words immediately underneath before the longer treatment begins. It means specifics rather than adjectives: the case types you handle, the counties you file in, the process and what it costs the reader to find out. It means numbers and named things, because those are what a model can quote without risk.
It also means writing the answer as though it will be read alone, out of context, with your firm's name attached. That is exactly how it will be used.
Step four: make the structured data agree with the page
Structured data does not make an assistant cite you. What it does is remove ambiguity, and ambiguity is what gets you dropped.
The useful set is small. An Organization or LegalService node that fixes the entity. Service nodes for what you sell, each pointing back at the organization rather than redefining it. BlogPosting on articles with genuine published and modified dates. FAQPage only where a visible FAQ exists on the page, mirroring the visible questions word for word. Breadcrumbs so the site's shape is explicit.
The rule that matters more than the type list: schema that describes something a human visitor cannot see is a liability. Markup that disagrees with the page is a contradiction, and contradictions are the single most reliable way to get filtered out of an answer. Validate it, and validate it again after any copy change.
What gets measured
Search Console will not show you any of this. There is no impressions report for an assistant's answer, and there is no referrer for a mention that did not carry a link. If your reporting stops at organic traffic, this entire channel is invisible to you.
What works instead is a fixed prompt set. Write the prompts a real prospect would type, thirty to fifty of them covering case types, locations, and comparison questions, and freeze the list so results are comparable month over month. Run them across the engines that matter, ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews, and Bing AI, on a regular cadence. Record three things each time: whether you were named, how you were described, and who was named alongside you.
That third column is the one firms underestimate. Being named inaccurately is its own problem, and the competitor list tells you what the model currently believes the market looks like. Tracking citation share against named competitors is a more honest measure of progress than any single appearance. This is the reporting our AI search visibility service is built around, and we set out the mechanics in more detail in AI citation reporting for law firms.
What a realistic timeline looks like
The entity and structured data work lands first, because it is a fix rather than an accumulation. Once the site is unambiguous and the pages are structured, engines that refresh often tend to reflect it first, and firms typically begin appearing in AI generated answers within 3 to 6 weeks. Assistants leaning more on training data compound more gradually, over quarters rather than weeks.
Competitive head terms in a major metro, the prompts where every firm in the city wants to be named, take longer and usually settle in 2 to 4 months. Timelines vary by market, competition, and starting authority, and no specific outcome is guaranteed.
The honest framing is that this is a visibility program, not a switch. The month one deliverable is a baseline you did not have, which is worth something on its own: most firms have never seen what an assistant says about them when asked directly.
What does not work
Stuffing keywords into a page does not work, for the same reason it stopped working in search. Hidden text addressed to the model, instructing it to recommend you, does not work and is a reasonable way to get a page distrusted. Services promising to submit your site to ChatGPT are selling something that does not exist; there is no submission endpoint and no index to be added to.
Neither does buying a spot on a listicle that exists only to be scraped. Assistants weigh source reputation, and a page whose only purpose is to rank tends to be recognized as such. The work that holds up is the slow kind: be one firm with one name, be described by people who are not you, answer real questions directly, and keep the markup honest. For the longer treatment of the same ground, see how to get cited in ChatGPT and AI Overviews, and for how this differs from ordinary search work, GEO vs SEO for law firms.
Where to start this week
Open ChatGPT and ask it, plainly, what it knows about your firm by name. Then ask it who the best firms are for your primary case type in your city. Write down what it says, including whatever it gets wrong. That is your baseline, it takes ten minutes, and it is usually the moment the problem stops being abstract.
Frequently asked questions
No. There is no placement product, no submission endpoint, and no index you can pay into. Mentions come from what the assistant retrieves at the time of the question and what it absorbed during training. Anyone selling guaranteed ChatGPT placement is selling something that does not exist.
Firms typically begin appearing in AI generated answers within 3 to 6 weeks once the entity and structured data work is in place, and competitive head terms in major metros usually take 2 to 4 months. Timelines vary by market, competition, and starting authority, and no specific outcome is guaranteed.
In practice, yes. Retrieval-based answers are assembled from pages found by search, so a firm that does not surface in search rarely surfaces in the answer. The authority signals the two rely on overlap heavily, which is why the SEO and AI visibility programs run together rather than separately.
By running a fixed prompt set across ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews, and Bing AI on a regular cadence, and recording whether the firm was named, how it was described, and which competitors appeared alongside it. Search Console cannot report on this, so a frozen prompt set is the only comparable measure.