Updated . By the Headcount Labs data team.

What we built

We set up ten exact-match domains. Nine cover a legal question, and each is a single page that ranks real firms, with the facts on it sourced. The tenth is a control in a field we already know. All ten are published by Rolesource, an independent publisher. Rolesource also operates Headcount Labs, so read what follows with that in mind. None of the sites is a law firm.

How the assistants searched

We gave Claude and ChatGPT the kinds of prompts a person might type, and we wrote down what they did.

  • A bare prompt gets rewritten to "best <practice> lawyers <place> 2026".
  • A story prompt without a picking verb often triggers no search. In Claude, 11 of 11 legal story prompts, such as "a semi rear-ended me", ran no web search at all. The answers came from memory, with general advice and a pointer to the state bar referral service. ChatGPT searched on every prompt.
  • When the assistant does search, the query is the user's own adjective plus the specifics, then "lawyer", then city and state. "Which car accident lawyers in LA are top rated?" became "top rated car accident lawyers Los Angeles 2026". "An 18 wheeler hit my husband outside Dallas. Find me a lawyer" became "18-wheeler truck accident lawyer Dallas Texas". No adjective in, no adjective out.
  • ChatGPT usually named three to five firms with phone numbers. The list and directory pages it cited were Expertise.com and Avvo.

Day 0

On the first day, none of our ten pages showed up in any result. We expected that, since the pages were hours old. One result stood out. On Brave, the result set for the bare query "best truck accident lawyer" was almost all spam: 19 of 20 results were hacked .edu and .org pages. We will run the same prompts again at week two and week four.

The ten pages

More guides

Try it on real data

100,000 records free every month, no card.

Get started Pricing MCP setup