009 Agency is a law firm SEO agency built around Generative Engine Optimization: we make legal practices visible, cited and chosen inside AI answers across ChatGPT, Claude, Gemini, Perplexity, Grok and DeepSeek — and we re-measure it every two weeks against a logged baseline. For a law firm that means one question, asked the way a prospective client asks it: “who should I call about this?”
Legal is a smaller bench for us than crypto — three law-firm audits and one legal client on retainer — but it is the most consistent one. Every firm we measured had the same shape: a strong Google footprint, a real reputation, and an AI answer that named somebody else. The three findings below are from those audits, with the numbers.
why law firms lose in AI answers specifically
The general failure pattern holds everywhere: known but not recommended. In our published dataset 93% of brands were described accurately by AI and only 4.5% were named when a buyer asked who to pick. Law firms add three problems on top of it, and all three are about trust the engine cannot read.
ranking #1 on Google does not get you cited
One multi-state firm we audited has around 29,000 organic visits a month, 7,100 ranked keywords and 220 of them in Google’s top three. On the same questions, ChatGPT named the firm in 1 of 12 prompts and Perplexity in 6 of 12. For the posts where the firm ranks first on Google, ChatGPT cited the statute on Justia instead. The engines read the firm’s pages, understood them, and quoted the primary source they linked to. That is not a ranking problem; it is a citability problem, and no amount of traditional legal SEO fixes it.
the credentials exist, but not in a form a model can verify
The same firm scored 22 out of 100 on structured data: no LegalService or Attorney record on a site that gives legal advice, and every one of 1,089 blog posts signed “Webmaster”. A second firm, a certified estate-planning specialist, marked its lawyers up as Attorney rather than Person, with no bar number anywhere a crawler could reach. Engines answer legal questions with a warning label unless they can confirm who is speaking. In our audits the warning was there every time the credentials were images, PDFs or absent.
duplicate domains and hidden libraries hand your citations to someone else
The estate-planning firm’s site was also served, in full, from a second domain the firm had forgotten about. ChatGPT cited the firm through that domain, not the real one. The same audit found 17 older articles still live and answerable but outside the sitemap and the navigation — a library the engines could quote and the firm did not know it had. Neither problem shows up in a rankings report.
what the shelf looks like in our audits
| practice | what the buyer asked | what the engines answered |
|---|---|---|
| multi-state litigation and employment | “which firm for a landlord dispute in Colorado” and 11 similar prompts | named on 7 of 24 runs. The only prompt where both engines named the firm was the one practice where it had a dedicated “for landlords” page — which is the playbook. |
| estate planning, Northern California | everyday estate planning near Lake Tahoe, and “a tech executive with a second home in Tahoe” | named on 4 of 24 runs, all local. On the high-net-worth prompts, 0 of 10: the engines named five Bay Area firms with published partner bios and none of them had a Tahoe office. |
| crypto and licensing law, EU | “which law firm for a MiCA licence” | 0 of 5 engines at audit; 4 of 5 by day 17 of the program. the case, with the logs. |
The pattern across all three: the engines answer with the firm that has named humans, a verifiable record and a page for the exact question. Local-pack rank, review count and Google position did not decide it once.
what the first 90 days look like
The order is the same as everywhere, because it has to be: readable, then verifiable, then quotable.
- Audit. The prompts your prospective clients actually ask — by practice area, by location, by the situation they are in — run live through the engines and logged verbatim, plus crawler checks with real AI user-agents. You get a health score, the verdicts engines currently give about your firm, and the three most expensive gaps. Complimentary. what is inside →
- Make the credentials machine-readable. Person and LegalService records with bar admissions, practice areas and locations; one canonical domain; the hidden library back in the map. Shipped as files your web vendor applies. technical GEO →
- Publish what gets quoted. One page per situation a client is in, written to be liftable in one paragraph and signed by the attorney who handles it. The “for landlords” page that worked by accident, done on purpose.
- Get onto the lists. The directories, bar-association pages and legal roundups the engines actually cite for your practice and city — identified from your own panel, not from a generic outreach list.
- Re-measure. Same panel, every review, reported as a delta. Series, not screenshots.
proof, and its limits
One measured result we can state: a licensing and crypto-law firm went from 0/5 to 4/5 engines naming it on its category prompt by day 17 of the program (July 2026). The full log is on the case page. That is one measured run on one firm, not an average and not a promise — we re-test every review precisely because single runs are volatile.
The other two law firms on this page are audits, not clients. Their names stay out of this text; on the call we show the raw answers, including the ones where the engines were right to prefer a competitor.
one thing we will not do
We will not put credentials in your copy that are not on the record. Engines check a lawyer’s claims against bar databases, court records and third-party profiles, and a mismatch produces a warning in the answer, not a footnote. Everything we mark up is something a model can confirm elsewhere; if it cannot be confirmed, it stays out of the schema until it can.