Who do AI engines trust when someone asks for a car accident lawyer in New York? 9,049 answers, three source diets
Written by
Odin Lund
Category
Law firms
Date
Share this article
What we measured, and how
When someone in New York asks an AI engine who to call after a car accident, the engine does not search the way Google used to. It retrieves a small set of pages it trusts for that question, reads them, and writes an answer that names two or three firms. Which pages it trusts decides which firms get named. We measured that for one practice area, one state and 30 days.
The data set is 111 questions that people actually ask after a crash in New York, from “who should I call after a car accident in New York City” to “how does no-fault insurance work” and “is it worth hiring a lawyer for a minor accident”. We run every question daily on ChatGPT, Gemini and Google AI Overviews for William Mattar Law Offices, the New York motor vehicle accident firm we manage, tracked against 13 competing firms. Between 18 September and 17 October 2026 that produced 9,049 answers. For each one we recorded which firms were named and in what order, and every page the engine retrieved and cited.
Two limits. It is one state and one practice area, so the specific domains will differ elsewhere, although the pattern has held in the other categories we track. And “retrieved” means the engine fetched the page while writing the answer, which is not the same as citing it. We count both and say which one we mean.
Three engines, three source diets
The headline finding is that the three engines do not share a source list. Each one has a diet, and the diets barely overlap outside the firm’s own website.
ChatGPT reads the state first. nycourts.gov was retrieved in 28% of its answers, ny.gov in 27%, nysenate.gov in 15% and the state bar in 9%. It then names firms from directories: Avvo in 15% of answers, Justia in 11%, Expertise in 9%, lawyers.com in 9% and Super Lawyers in 9%. William Mattar was retrieved in 15% of ChatGPT’s answers, more than any competitor, and when it was retrieved ChatGPT cited it four times on average.
Gemini goes straight to law firm websites and skips most third parties. William Mattar’s site led at 15% of answers. After that come Super Lawyers at 6% and then a long tail of individual firm sites at 3% to 6% each. Gemini is also the heaviest citer: when it retrieves a page it cites it 2 to 4 times, so a firm whose own pages are retrieved gets most of the credit in the answer.
Google AI Overviews behave differently from both. YouTube was the single most retrieved source, in 23% of answers. Competitor firm sites follow at 10% to 19% each, and then the social layer: Facebook in 6%, Reddit in 5%, Instagram in 5%. Directories matter far less here (Avvo 5%, Justia 6%) and government sources appear in only 7% of answers. AI Overviews are built from video and social proof, not from the state and not from the directories.
Same firm, same questions, three different results
Because the source diets differ, the same firm gets a different result on each engine for the same questions.
On ChatGPT William Mattar holds 55% of all firm mentions and is named in 22% of answers, first of the 14 tracked firms. On Gemini it holds 57% and is named in 20%. On Google AI Overviews it holds 21% and is named in 10% of answers. It is still first on AI Overviews, but by a thin margin, and the competitor closest behind is the firm whose site AI Overviews retrieves most, at 19%.
The explanation is in the diets. The work done on the firm’s own site and on its directory profiles is exactly what ChatGPT and Gemini read, so those engines reward it. AI Overviews reach for YouTube, Facebook, Reddit and Instagram, where the firm has a smaller footprint than the competitor that leads there. For most firms this is the open flank: the engine that sits on top of Google itself is the one least influenced by the website and the directories, and most influenced by video and community sources.
Average position barely moves between engines (2.0, 1.9 and 2.3). Once a firm is in the answer, it tends to be named early on all three. Getting into the answer is the hard part, and it is decided by the sources.
What a law firm should do with this, and sources
The practical reading of the data is a short list, in the order it pays off.
Answer the process questions on your own site. ChatGPT currently answers “how does no-fault work”, “what counts as a serious injury”, “how long do I have to file” from ny.gov and nycourts.gov, and then names firms from directories. A firm page that answers the same question plainly, with the statute and the numbers, gets retrieved next to the official source, and that is often the first time the engine connects the question to a firm.
Make the directories agree with the site. Avvo, Justia, Super Lawyers, Expertise and lawyers.com are where ChatGPT finds firm names. Practice areas, cities served, fee model, consultation terms and case results must read the same on every profile as on the site. Where they disagree, the engine hedges or picks another firm.
Build the video and social layer for AI Overviews. Short YouTube explainers on the exact questions above, with the firm name and city in the title and description, and consistent Facebook and Instagram profiles. This is the only source group that moves AI Overviews, and it is the one most firms have left to their intake team.
Measure by engine, not in aggregate. A blended visibility number hid a 36-point gap between ChatGPT and AI Overviews for this firm. The engine breakdown is what tells you where to work next.
This is the work SourceNode does for firms: we run the questions daily, read which sources each engine is using, correct and publish the sources, and match the visits that follow to signed cases. The companion piece lists the 111 questions themselves. How an engagement runs for a firm is on our law firms page, the documented result is on the results page, and a call shows what the engines say about your firm today.
Sources
SourceNode tracking: 111 questions, 14 New York motor vehicle accident firms, ChatGPT, Gemini and Google AI Overviews, 9,049 answers, 18 September to 17 October 2026. Retrieval and citation recorded per answer.
Engine shares: ChatGPT 3,330 answers, Gemini 3,330, Google AI Overviews 2,389.
Related: Which sources AI models cite (the 46-domain concentration finding) and How AI search builds an answer.
Share Blog
