How to Rank in ChatGPT

1. Find the prompts you actually lose

Do not start with "AI SEO best practices". Start with the specific buyer prompts where your firm is invisible.

Build a versioned panel of unbranded buyer situations, run it across the assistants in scope, and repeat enough observations to report counts rather than anecdotes. The appropriate sample depends on the decision and the uncertainty you can tolerate; a magic “10 runs” rule is not a methodology.

A useful prompt sounds like a buyer, not a category. Not "who are the best divorce lawyers in NYC" but "my spouse controls all the accounts and I'm trying to leave — who should I talk to in New York?" The first is a directory lookup; the second is what AI assistants are actually asked.

Group losses by buyer situation and model. The hard part is deciding which gaps are commercially important and which observable evidence is missing. That diagnosis — not a giant prompt spreadsheet — is where the work begins.

2. Rewrite the losing pages in answer-shape

AI search products can retrieve pages or passages and then synthesize an answer. Give that pipeline clear, self-contained evidence to work with. “Complex matrimonial matters” says little. A plain description of the matters handled, jurisdiction, process, and relevant proof says considerably more.

Three useful editorial checks:

• Category labels → buyer situations. "Family Law" is a Google label. Buyers don't type it. "Divorce with hidden assets" or "custody with an out-of-state parent" is what they type.
• Connect credentials to fit. Experience becomes useful when the page explains which matters it applies to and provides support a reader can verify.
• Boilerplate → specific detail. Describe scope, geography, process, limitations, and evidence. Do not invent impressive numbers merely to create something quotable.

Every page targeting a buyer prompt needs at least one paragraph an assistant can quote verbatim without embarrassment.

3. Make evidence easy to find and interpret

A short answer block can help when it improves the page for a human reader. An FAQ can help when buyers genuinely ask the question. Neither format is a published universal preference signal, and FAQPage markup is not a ChatGPT ranking switch.

1. The question sounds like a real buyer prompt.
2. The answer is a single, quotable paragraph — not a link dump.
3. Any structured data accurately matches what is visible on the page.

Also verify the boring machinery: public access, clean HTML, sensible canonicals, internal links, consistent business identity, and crawler rules. Technical access cannot manufacture evidence, but it can prevent good evidence from being considered.

4. Add citation-worthy proof

A recommendation is easier to justify when relevant claims have support. Prioritize proof that helps a buyer establish fit:

• Verifiable specifics. Use numbers only when they are defined, supportable, current, and ethically appropriate to publish.
• Named work. Landmark cases, published opinions, media appearances, professional recognitions — with sources.
• Live disclosures. Fee structures, response times, intake protocols, and other operational specifics that other firms bury.

Owned claims and third-party corroboration do different jobs. Which source gap matters most depends on the prompt, model, category, and competitors already being cited. That is a diagnosis, not a universal checklist.

5. Re-run and prove the delta

The final move separates AI ranking work from theatre. Verify the change is public and accessible, then re-run the matched panel on a predeclared schedule. Compare recommendation presence, model coverage, and cited sources. One new mention is a clue, not a victory parade.

Anyone selling AI ranking work who cannot show you the before-and-after prompt set with matched sample sizes is guessing. That is why Viclaro publishes methodology-versioned Atlas snapshots and why the audit product ships with a re-run built in.