Law Firm Website AI Visibility: It's an Architecture Problem

If your firm ranks on page one for “estate planning attorney [your city]” and still never gets named when someone asks ChatGPT the same question, stop rewriting the copy. Law firm website AI visibility is decided by architecture, not prose. A machine has to identify who you are and parse what the page actually claims. Then it has to do something once it arrives. Most firm sites fail at the first step and never get graded on the third.

Legal feels this harder than almost any other industry. Martindale-Avvo’s 2026 State of the Legal Consumer report found legal queries trigger AI Overviews at close to 78%, the highest rate of any vertical, and that consumers have moved from “search and click” toward “ask and chat.” That is one vendor’s sample. Treat 78% as a direction, not a constant.

Now the shallow fix, which you have probably already been pitched. Add FAQ schema. Publish an llms.txt. Commission a few “AI-optimized” blog posts. Skip it. Ahrefs analyzed 137,000 sites and found 97% of their llms.txt files were never fetched at all, and Google has said plainly that no AI system currently uses the file. That is a hack aimed at what a machine says about you. The real work changes what a machine can read and do on your site, which is the part you own.

The frame I use for that work is Slobodan Manić’s Machine-First Architecture: four pillars in a fixed order. Identity, structure, content, interaction. The order is the point, because each layer only works if the one under it holds.

Why law firm website AI visibility breaks before the content does

There are two different machines reading your site, and they have different capabilities.

Googlebot renders JavaScript. The AI crawlers largely do not. Vercel’s analysis of AI crawler behavior found GPTBot and ClaudeBot fetch JavaScript files but never execute them. Glenn Gabe documented the same pattern in a case study of client-rendered sites losing visibility across ChatGPT and Perplexity while their Google rankings stayed flat.

That one fact explains most “we rank but we’re invisible” cases. If your attorney bios, practice descriptions, fee explanations or intake form paint after hydration, Googlebot sees a law firm and GPTBot sees an empty shell.

You can check this yourself in thirty seconds, no vendor required. Run curl -sL https://yourfirm.com/practice-areas/dui/ and read what comes back. If your practice copy is not in that raw HTML, it does not exist to the systems your prospective clients are now asking.

Pillar one: identity, or being one entity instead of five

A retrieval system will not cite what it cannot resolve. Before anything else, it has to decide that Smith & Associates on your homepage, Smith and Associates PLLC on your Google Business Profile, The Smith Law Firm on your bar listing, and whatever spelling sits in a directory nobody has logged into since 2019 are all the same organization.

Most firms quietly fail this.

Drop the Attorney type

schema.org deprecated the type in favor of LegalService, which it calls “more inclusive and less ambiguous.” Markup using the deprecated type still validates and buys you nothing. Changing it is a twenty-minute job on most templates.

Define the firm once, then point at it

One LegalService node with a stable @id. A real sameAs array pointing at the bar profile, the Google Business Profile, LinkedIn and the directories you actually maintain. Then a Person node per attorney with worksFor pointing back at that @id. Repeating a fat firm block on every page blurs which node is canonical, and a model reading five slightly different firm descriptions has no way to pick the true one.

Consistency is itself the signal

Search Engine Land’s work on structured data and local visibility makes the same point from the local side: structured data now functions as evidence that your business facts are stable enough to reuse. A firm whose address and practice scope contradict themselves across the web is a firm a model has a reason to skip. This is also where LocalSEO and AI visibility stop being separate projects. They read the same entity graph.

Pillar two: structure a retrieval system can lift from

Once you are identifiable, the question becomes whether a page yields clean, quotable units.

Server-render what matters

Not because rendering is fashionable. The crawlers with no JavaScript engine get exactly one shot at your HTML.

Headings and tables are where the chunks break

Keep one h1, then a heading hierarchy that describes the content underneath it rather than decorating it. Put fee structures and deadlines in real HTML tables, never in a screenshot. Retrieval systems chunk pages before they rank passages, and your headings decide where those cuts land.

Two things I refuse to let a firm ship

PDF-only content libraries, because your best explanatory material should not be locked in a format that chunks badly. And practice-area pages that exist only inside a filterable component, which is a very expensive way to be invisible.

Pillar three: content that is actually quotable

The archetype you have read a hundred times

Our firm has decades of combined experience aggressively fighting for injured Floridians. We treat every client like family. Call today for a free consultation.

Count the extractable facts. Zero. There is no jurisdiction, no statute, no filing deadline, no fee model, no description of what happens in week one. A model asked “how long do I have to file a claim in Palm Beach County” cannot use that paragraph for anything, so it uses someone else’s.

What earns a citation is boring

Open the page with a direct, plain answer to the question the page is named after, in the first two sentences, before the credibility paragraph. Name the governing statute and link it. State the deadline and date-stamp it, because a model weighting recency has no way to know whether your 2019 page is still accurate. Explain the fee structure in the terms a client would use. Spell out the first three business days after someone calls you.

None of that is a guarantee of being cited. I cannot promise a model will name your firm, and nobody honest will. You can only be legible or not.

The constraint I will not talk a firm out of

Every claim has to survive bar scrutiny. Answer-first structure is not license to write comparative superlatives or outcome promises that ABA Model Rules 7.1 will not carry. Attorney review stays in the pipeline. In the firms I have worked with that has cost roughly a day per page, which is my experience and not an industry benchmark. Worth it.

Pillar four: interaction, or what happens when an agent tries to hire you

The fourth pillar is the one nobody audits, and it is arriving fastest. Agents do not only read. They fill in forms and book consultations on someone’s behalf.

Most firm intake cannot survive contact with one

The form sits behind a JavaScript widget or a scheduler on a third-party domain. Add a CAPTCHA and the path is simply closed. Cloudflare now lets site owners treat agent traffic as its own category, governed separately from search crawlers, and as of September 15, 2026 new ad-bearing domains block the agent category by default.

Decide the policy, do not inherit it

Allow the retrieval crawlers and keep a server-rendered contact path that works with scripting off. Then make sure whatever comes through it lands in a human’s hands fast. That last part is intake, not marketing, and it is where most of this value leaks back out.

What a principal-led audit actually changes

Twenty-five years across marketing and software engineering is the only credential I will lean on here, and it is on the author page if you want to check it. What I have watched is one pattern repeating. Print directories lost to search. The firms that moved early did not win because their copy improved.

Here is the sequence I work in, and why it is not the order an agency usually proposes.

1. Fetch the raw HTML for three money practice pages and one attorney bio. Diff it against the rendered page. This alone tells you whether you have a content problem or a rendering problem, and it takes an afternoon.
2. Reconcile the entity. One canonical firm name, one address format, one LegalService node, attorney Person nodes wired to it, sameAs cleaned of dead profiles.
3. Rebuild the heading and table layer on the money pages so each page has extractable units rather than one long undifferentiated column.
4. Rewrite openings, not whole pages. My working heuristic is that the first 60 words of a practice page do most of the retrieval work. That number comes from my own testing, not from published research, so treat it as a shape rather than a constant. Full rewrites are usually agency busywork.
5. Test the intake path with scripting disabled. If the form does not submit, you have a conversion bug that predates AI and is costing you human clients too. The full list is in the conversion checklist.
6. Set the bot policy on purpose. Which crawlers you allow, which agents you allow, logged and reviewed.

Notice what is absent: a content calendar, a package tier, a junior running a crawl tool and mailing you the PDF. Steps one through three are engineering work on the site itself. Machine-first work is website work. If the person selling you AI visibility cannot open the template and change how it renders, they are going to sell you blog posts, because blog posts are what they can produce.

If this is your problem

Pull the raw HTML on your three most valuable practice pages. If your best arguments are not in it, you have found the thing that is costing you. Write us at hi@carlosarias.com and tell us which pages you pulled. You will get a straight read from the person who would do the work.