Is Your Law Firm Website Ready for AI Agents?
Google's Lighthouse now has a category for AI agents. What an agent reads on a law firm website, the four things a site can ship for it, and what we changed on our own site when we checked.
Wolfe Services · · 5 min read
Updated
Someone hurt in a crash last night may not open a search page at all this morning. Some will ask an assistant to find a lawyer, compare two firms, or check whether a firm handles their kind of case. We can’t tell you what share of a firm’s prospects do that today, and neither can anyone selling a fix. The assistant does read the firm’s website, though, and it reads it differently from a person.
In May 2026 Google added a category to Lighthouse, the audit behind PageSpeed Insights, for exactly that reader. It’s called Agentic Browsing, and it’s still marked as under development. It’s the first time Lighthouse has asked whether software acting for a person can use a site.
What an agent actually reads
A person sees the design: the photo of the partners, the phone number in red, the reviews carousel. An agent gets the page as data. It reads the HTML, the accessibility tree that browsers build for screen readers, and any structured data the page declares about itself. If it can, it reads a plain-text version and skips the layout entirely.
That changes what counts as a good page. A homepage that loads 400 KB of script before the text settles is slow for a phone and slower for an agent working through ten firms at once. A menu built from unlabeled icons is a guessing game. A firm described only in marketing paragraphs, with no machine-readable statement of what it is, where it practises and what it handles, leaves the agent to infer those facts. It may infer them wrong.
a person sees
A photograph, a headline, a red phone number, a reviews carousel, and a form that emails an inbox. Twelve scripts and three tags behind it.
an agent reads
- index.html script weight, blocking scripts, image formats
- accessibility tree named buttons, labelled fields, a structure to follow
- JSON-LD · LegalService what the firm is, where it practises, what it handles
- llms.txt · index.md a plain-text index and a Markdown copy, read without parsing the layout
What Lighthouse checks
When we ran Lighthouse 13.5 in October 2026, the Agentic Browsing category scored three things and reported them as a count, such as 3/3, rather than a 0 to 100 score:
| Check | What it looks for |
|---|---|
| Accessibility tree | Interactive elements have names, roles and a sensible structure an agent can navigate |
| Layout stability | Content doesn’t move after it first renders, so an agent doesn’t act on the wrong element |
| llms.txt | If the file exists: it has a title, enough content, and links |
It also lists audits for WebMCP, a proposed way for a page to offer an agent named actions like “book a consultation”. Those didn’t count toward the result in our runs. A site without an llms.txt file is marked “not applicable” for that check rather than failed. That may change as the category develops.
Four things a firm website can ship
A clean, labelled page
This overlaps with accessibility law and good practice, and it’s the cheapest to fix: real buttons with real names, headings in order, forms with labels, no content that jumps after load. It’s also where most firm sites fall short. Fourteen of the 25 homepages in our scorecard failed Lighthouse’s accessibility-tree check.
Structured data that says what the firm is
Schema.org has a LegalService type. A firm’s homepage can state its name, offices, practice areas, attorneys, hours and service area in JSON-LD, the format search engines already read. In our October scorecard of 25 plaintiff-firm homepages, 15 did this in some depth, 2 had no structured data at all, and the rest declared something thinner.
An llms.txt file
Jeremy Howard proposed it in September 2024: a short Markdown file at the site root that tells AI systems what the site is and links to the pages worth reading. It costs an afternoon, and 11 of the 25 homepages we measured already had one. No search engine or AI company has told us it ranks or cites a page because of it, so treat it as basic hygiene rather than a growth lever.
A Markdown copy of each page
The same proposal suggests offering each page as plain Markdown at a predictable address, for example /docket/index.md beside /docket/. An agent then gets the text without the navigation, scripts and decoration. None of the 25 homepages we measured offered one.
What we found when we checked our own site
We built the scanner behind our readiness score to test client sites, and pointed it at ours first. The result was uncomfortable. wolfeservices.co passed on speed, images and caching. It had no llms.txt, no Markdown copies, and no Content Security Policy, the browser-enforced list of scripts a page is allowed to run.
We fixed all three this week, and none of them is maintained by hand. Each time the site builds, a final step reads the finished pages and writes a Markdown copy of each one, an llms.txt index of those copies, and a security policy that allows exactly the inline scripts that build produced. Because the policy is generated from the finished pages, it can’t drift out of step with them, and an inline event handler that the policy would block stops the build. Our Field Notes are served from the Markdown they’re written in.
The law firm site we operate in New York already had an llms.txt file and deep structured data, and passes all three Lighthouse agent checks. It doesn’t yet publish Markdown copies of its pages.
What this does not prove
None of this shows that an AI assistant will recommend a firm more often. We haven’t measured that, and we don’t know of a reliable way to measure it yet. A clean page for agents is still a clean page for people. Most of the items above also make the site faster, more accessible or harder to tamper with, which is why we think they’re worth doing before anyone can prove the agent case.
Some firms block AI crawlers in robots.txt on purpose, to keep their content out of training data. That’s a legitimate choice, and it can also keep the firm out of those assistants’ answers. Make it deliberately; a plugin shouldn’t make it for you.
Check your own site
- Run your homepage through PageSpeed Insights and open the Agentic Browsing result.
- Visit your-domain.com/llms.txt. Is there anything there?
- Ask whoever built the site which structured data type describes the firm, and where it lists your practice areas.
- Check robots.txt for GPTBot, ClaudeBot and PerplexityBot. Did someone at the firm decide that?
- Run the Wolfe readiness score to see all of it next to a site we built, measured at the same moment.
Next step
The diagnosis comes before the engagement.
If this describes your firm, request a private Architecture Review of your website, search, attribution, vendors, and intake system. One document, written for the partner who signs.