We asked Google’s AI Overviews 1,000 legal questions. Here’s How It Responded

Every day, people in a need of legal guidance turn to platforms like Google to look for help. They may ask a question like “should I hire a workers’ comp attorney to speed up my settlement?” and instead of getting a list of 10 blue links, Google now serves at the top of their search result a summarized answer. That answer is the new #1 ranking.

Unlike the old search results page, a list of websites is no longer the first thing a users sees and you can’t buy your way into the top of the list. We wanted to better understand how Google’s AI Models respond to different user queries in the legal space so we looked at over 1,000 questions that users may ask in a specific geographic region. We’re most interested in learning when those questions lead to an AI model providing a list of recommended law firms.

The Test We Ran: 1,000 unbranded legal questions, three passes of measurement.

We built a corpus of legal-services queries reflecting how real people ask. These include personal injury, workers’ compensation, Social Security disability, and veterans’ claims, across shared metros in the Southeastern US. From it we drew a stratified random sample of 1,000 queries: 448 phrased as informational (“what evidence do I need to prove fault in a red light accident?”) and 552 as commercial (“best truck accident lawyer in Montgomery Alabama”).

None of the queries named a specific firm or attorney. This is a study of unbranded demand, the moment a potential client doesn’t yet know who you are. Each query then went through three passes:

1. Ask. Every query was sent to Google’s Gemini flash model with live Search grounding, likely the same retrieval substrate behind AI Overviews. We recorded the full answer, every source it cited, and every search the engine ran on its own behalf.
2. Extract. A second, independent model pass read each answer and pulled out every provider presented as an actionable recommendation: law firm, government agency, nonprofit or legal aid, directory, or other company. Entities merely discussed didn’t count.
3. Classify. Each of the engine’s 6,436 self-generated searches was classified as informational, commercial, or navigational.

Firms are anonymized throughout (“Company A,” “Company B”). The findings are about the playing field, not any one player. Here’s what came back.

AI answers are recommendation engines, even when nobody asked for one.

59% of all answers recommended at least one specific provider. Ask a commercial question and names are near-certain but almost never as a bare list: the overwhelming majority were mixed answers, substantive legal explanation wrapped around named firms. And nearly one in five people who only asked to learn got a shortlist anyway.

The lesson inside that asymmetry: the AI doesn’t switch between “teaching mode” and “selling mode.” It teaches, and it names providers while teaching. 18% of informational queries and 92% of commercial queries produced an answer naming at least one provider. Content that explains is the vehicle for being recommended.

The hidden search layer: one question in, six searches out.

Before answering, the engine fanned each query out into an average of 6.4 Google searches the user never sees. We classified all of them to find out whether those machine-generated searches keep the user’s intent or drift between learning and buying. The flow turns out to be one-directional.

Across the 6,436 fan-out searches in the sample: when AI Overviews fans out an initial informational question, 4% of the searches it runs are commercial (95 of 2,599). The engine probes for providers even when the user only asked to learn.

Drift runs far stronger the other way: 18% of the searches spawned by initial commercial questions were informational (684 of 3,837). Ask “who should I hire?” and the engine quietly researches deadlines, eligibility rules, and settlement mechanics to justify its recommendations. Another 24% were navigational, the engine googling specific firms by name, checking directories and review platforms before deciding who makes the answer.

If your commercial pages have no informational depth behind them, you’re likely to struggle to appear in AI Overviews for a fifth of the retrieval that decides commercial answers. And if your reputation surfaces are thin, you can lose the answer in a background check you never knew happened.

The shortlist of AI recommendations is long, roughly double a map pack.

When an answer recommends providers, it isn’t a winner-take-all snippet. Recommending answers that named any law firm named 7.6 on average. And firms aren’t only competing with each other: a meaningful share of slots goes to government agencies, legal-aid nonprofits, and directories — free resources holding the same real estate.

Nobody owns this AI Overviews market, yet.

Treating every law-firm mention as one “slot,” 936 distinct firms competed for 3,670 slots across the sample. If you expected a Google-style winner-take-most curve, this is the surprise: the market leader holds a share most industries would consider a rounding error.