AI search for law firms

There are two distinct Google features in play here.

• AI Overview. If you're unfamiliar with the term: an AI Overview is the AI written summary with a few cited links that Google puts above the ordinary results.
• AI Mode is a separate conversation tab on the Google results page.

ChatGPT and Perplexity both do a lot of the same answering outside Google. If you're a law firm you're either named in the answer, or you're missing from it entirely.

How often this matters in legal

Whitespark used 540 searches across Houston, Phoenix and Denver to check for AI Overviews, but did all of the searching by hand. Personal injury lawyers were one of the six industries tested. Their searches came from outside each city, so their own location did not skew it¹.

If you look at the intent behind each search, it is definitely not uniform¹:

• 92% of research questions returned an AI Overview
• 97% of in-between questions like "should I get a lawyer after a car accident"
• Only 15% of ready-to-hire searches like "personal injury lawyers in Phoenix"

So the AI answer owns the early stages of the hire. The moment a client decides to call is the moment the map pack takes over. This page is not about the map pack.

The most useful finding: ranking is still the way in

On 12 October 2025, an AI Overview showed up on 362,000 US desktop searches seoClarity ran. They pulled the citations for each answer and checked them against the top 20 for the same search. That's 5.1 million citations to work with².

There are three lessons in this study.

It mostly looks like the AI picking a ranked page and quoting it. 94% of those AI Overviews have cited at least one page from the top 20².

Position still earned a lot of the citations.

Nearly half of the citations are from elsewhere (outside the top 20). The top 20 gets only 56% of all 5.1 million citations².

You'll see the same effect in ChatGPT, following its own search results. AirOps and Kevin Indig went through ChatGPT's retrieval flow and logged the 16,851 queries and 353,799 pages. A page at the top of ChatGPT's own search results has a 58% chance of getting cited. By the tenth result it comes down to 14%⁵.

Two things being sold to law firms that the evidence does not support

A file called llms.txt, sitting on your site with a list of your content for AI systems, is not going to get you cited.

Cyrus Shepard limited his review to 55 studies, experiments, patents and case studies and scored each proposed factor out of ten and marked llms.txt down hard.

One of the ranking pages for this subject was a legal SEO agency linking one from its own footer.

You don't need any special file to show up in AI Overviews. Google's own documentation does say this plainly:

"There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." And: "You don't need to create new machine readable files, AI text files, or markup to appear in these features."³

What decides whether your page is the one quoted

Once a page is in the candidate pool, the writing on the page matters.

The hard part is separating the cited 15% from the rest.

1. The answer has to be near the top of the page

AI answers quote passages, not pages. Dan Petrovic says Google's AI will lift content picked specifically to build an answer, known as "grounding snippets". It takes your copy often verbatim from the page it is grounding on. If the answer sits in your opening paragraph, the AI will very nearly lift the whole thing⁷.

Most legal pages open something like this:

"For over 30 years, our attorneys have fought for injured people in Michigan."

It answers a question that no client asked. So a client just skims past it, and an AI system will discard it. I could make a similar argument about how the deadline, the fee and the test for whether they have a case should come first. This would shift the history of the firm from the opening, where it currently lies, to the end of the page.

2. Length works against you past a point

Petrovic also looked at 7,060 queries and 883,262 grounding snippets. Each answer takes about 2,000 words across all sources. That works out to an average of about 540 words per page⁸.

So a 4,000 word page is just not four times as visible as a 1,000 word one. Use enough words to answer the question and its obvious follow-ups, then stop there.

3. Most citations point at sites you do not own

Wix went through 75,000 AI answers, and counted 1,056,727 citations across ChatGPT, Google AI Mode and Perplexity⁹.

Let's think about that for a legal search. Ask an AI system for a personal injury lawyer in your city and you get a list-heavy, directory-heavy response. The firm websites themselves are largely in a supporting role.

This is why the review and directory profiles you've ignored for years matter again.

The ultimate guide trap, which is the most useful thing on this page

AI systems (Google's, OpenAI's) generally do not rest on one single search. They run a fresh search for several sub-questions, and they gather different pages for each one. Google themselves say this and call it query fan-out³.

The obvious response is one enormous guide that is fully targeted to your sub-questions. This idea is the ultimate guide playbook, pure and simple and AirOps tested it across their 353,799 pages. They scored each page for covering fan-out questions with its subheadings.

Not only that, the 4.6 flatters broad pages for what is really their closer match to the main query. If you control for query match, broad coverage will have almost no effect⁵.

What their data shows is they cite the pages whose headings match the question that was asked⁵.

There is one interesting exception in the AirOps data, which is Wikipedia. Its 59% citation rate scores highest of all the domains in the dataset we talked about. Oddly few Wikipedia pages ranked well in the retrieval step, though. Its pages average 4,383 words with 31 lists and have 6.6 tables⁵. Your firm will likely never get Wikipedia's trust, so do not copy this playbook.

Where legal sits, and where I have to be straight with you

BrightEdge took a deep dive into AI Overview citations over sixteen months, and they measured the overlap with organic rankings. Over that time the overlap went overall from 32% to 54.5%. The spread between fields is fairly wide¹⁰:

• Healthcare, 75.3%
• Education, 72.6%
• B2B tech, 71.0%
• E-commerce, 22.9%

The vast majority of firms would primarily see AI Overview visibility as a function of ranking success. With a firm that cannot rank for a practice-area question, a quote in the answer above it would not be possible either. What I've described is a working assumption and I would treat it as such.

That missing legal category is also why the Whitespark numbers carry extra weight on this page. Now the Whitespark study has personal injury lawyers in it, so that 70% figure is real legal data¹.

What being cited is worth, and what it costs you

The cost side is pretty well measured.

About 60% of the searches in Google now end with no click to another site¹⁴.

But the return side is a little bit less known.

Like almost all vendor data, take the 120% figure with a grain of salt. I was curious about Google's claim that the people still clicking through are more likely to spend real time on the site³. Google says AI Overviews drive over 10% more use of Google on the queries where they appear, but how do we even know¹⁵?

You accept the click loss, get quoted in the answer and you're part way back. For a firm with cases worth thousands, that's a big deal.

The 18% problem, which nobody is selling a fix for

Profound went through around 730,000 cited ChatGPT conversations over three months trying to see where the citations came from¹¹.

The other 82% are really just inside the model's memory. No search, no citation, just what the model remembers and nothing in Search Console or any tracking tool.

If ChatGPT turns the search off, the answer about the best personal injury firms in Tampa comes from training data. That depends on whether you've been written about often and consistently across the web.

What I would do in the first week

None of this stuff needs a new tool or a package.

1. Write down the ten questions your clients actually ask, in their words, from your intake notes and your attorneys. That list is the primary raw material for everything below.
2. Check whether your site answers each one, and where on the page the answer sits. If the answer is in paragraph six or nowhere, that sets your to-do list.
3. Move the answers to the top of their pages. Two specific sentences should be up front in your page about each question.
4. Turn your section headings into the questions themselves. The key is heading match with rank accounted for⁵.
5. Split the monster pages. That's one question per page and your monster page gone.
6. Add the specifics that can be quoted: the statute and its number, the deadline, the fee arithmetic, the court. Sentences with no facts can't be used as an AI answer.
7. Claim and align every third-party profile. Somebody else's website already gets four in five citations⁹.
8. Record twenty short video answers to those same questions, with real titles and transcripts. Ahrefs analysed 75,000 brands and found that YouTube mentions tracked AI Overview visibility most closely. YouTube is the most cited domain in the answers¹⁶.

I'm a big fan of writing for clients, not search engines. In one sense, however, a page written purely to be extracted reads like it must have been written for a machine. Google's own page on helpful content, it asks very specifically, whether a page shows first-hand expertise, whether it's obvious who made it, and whether any automation is self-evident to visitors¹⁹.

Two warnings about every number on these pages

The platforms change first, the studies second.

In Semrush's data ChatGPT moved within weeks from citing Reddit in about 60% of responses to about 10%. Meanwhile Reddit's share on other platforms kept holding steady¹⁷. Lately I feel like any playbook for one AI system is going to have a short shelf life.

One check is generally not good evidence.

The content of the same AI answer constantly changes. Sample the same questions across repeated runs and engines and check the result holds before saying anything. That is the baseline the academic work on AI visibility starts with¹⁸.

Getting a screenshot of your firm in an AI answer is not exactly "a result". The question is whether the share of fixed questions naming your firm is higher than the share naming other firms. And if I'm a firm measuring this, I would sample at least two engines, run after run, on the same questions.

What this section still has to come

Nine further pages are written and cover each part in depth. There will be more of them here as they are checked:

• How AI answers are built: Fan-out, retrieval, grounding: the pipeline behind a cited firm.
• AI Overviews and AI Mode: How closely Google AI answers track the rankings you already have.
• ChatGPT and Perplexity: When they search, what they cite, and how fast that changes.
• Writing pages AI cites: Answer first, headings that match the question, facts that can be quoted.
• Authority and brand in AI: The studies disagree about links. Here is what that means for a firm.
• Getting named off-site: Most AI citations point somewhere else. Directories, reviews, Reddit, YouTube.
• Clicks and leads: Zero-click search, the CTR drop, and what it means for enquiries.
• Measuring AI visibility: What Search Console and Bing will tell you, and what they will not.
• What the studies say: Fifty-five studies, read against each other, applied to legal.