How consistently does Google’s AI Overviews recommend the same law firms?

Ask Google’s AI overviews for a personal injury lawyer twice and you may get two different lists. Ask it 2,500 times across 50 cities, and the randomness resolves into something measurable: which firms are locked into the answers, which markets still reshuffle, and what the model reads before it decides. We ran a test to identify patterns in how Google’s AI Overviews recommended personal injury law firms when people searched for a personal injury lawyer in 50 different cities.

The test we ran: one question, 50 cities, 50 fresh answers each.

AI answers are non-deterministic. One answer is an anecdote. Fifty answers are a measurement. We asked Gemini, the model family behind AI Mode and AI Overviews, the exact same question 50 separate times in each of the 50 largest US cities. Search grounding was on, so for every answer we captured three things:

• the searches the model chose to run behind the scenes,
• the sources it cited,
• and the firms it recommended, in order.

Fifty passes per city gives every firm a recommendation rate with real error bars: appear in 40 of 50 answers and you are an 80% firm, give or take about 6 points. It also lets us define a locked firm (80% or higher, a favorite the model returns to almost every time) and an open slot (list capacity the favorites do not consume). Firm name variants were merged before any counting; 168 merges were reviewed by hand.

In many cities AI recommends the same law firms no matter how many times you ask.

How consistent is the AI? Consistent enough that most markets have a permanent cast. Forty-nine of our 50 cities have at least one locked firm. The median city has three. And 30 of the 50 cities have at least one firm that appeared in 100% of that market’s answers (40 firm-and-city pairs in all). In San Francisco, for example, one firm appeared in all 50 of the market’s answers, every single time in first position. (As a reliability check we also ran our home market, Salt Lake City, which sits outside this 50-city panel, with a different question phrasing on a different day; the same top five firms came back.)

In the square grids, each square is one of the 50 cities. Hover any square for the city behind it.

Here is the full spread. Most cities carry two to four locked firms; the extremes are where strategy changes:

Number of locked firms per city across the 50-city panel. Only one city, Portland, has no locked favorites at all. Four cities have five. Hover a bar to see which cities are in it.

Counting locked firms only tells half the story, because list length varies too. A city with three locked firms and nine-firm answers still has six open slots; a city with three locked firms and six-firm answers is mostly decided. Plotting both puts every market on one map:

Each dot is a city, colored by how locked its market is. Dashed lines mark the median city on each measure. Hover any dot for its numbers.

The corners of that chart are worth naming:

Open slots are the average number of positions per answer not consumed by locked favorites.

Read this as a market-entry map. In Portland the AI has not made up its mind, and a firm is competing for seven genuinely available slots per answer. In Boston the realistic brief is displacing an incumbent with a 90%+ recommendation rate. Same model, same question, completely different competitive situation. Any AI visibility engagement that does not start by classifying the market is guessing.

There is a flip side to the locked firms. When a firm shows up in an AI answer at all, does it keep showing up? Usually not. We took every firm-and-city combination that appeared in at least one answer, 1,827 of them. That works out to 36.5 distinct firms recommended per city, a long tail of contenders competing under the incumbents. Then we counted how often each one came back across the 50 asks:

How often a recommended firm reappeared, counted by how many of the city’s 50 answers included it. Share of all 1,827 firm-and-city pairs that appeared at least once.

Six in ten of the firms that appear at all (60.0%) surfaced in five or fewer of the fifty answers. They are drive-bys: the model mentioned them once or twice and moved on. Only 8.4% of pairs are locked in. This is why a single favorable AI answer means very little. Seeing your firm once is closer to sampling luck than to a position, and knowing whether you are genuinely visible requires asking the question dozens of times, not once.

One more layer of recommendation math worth having on hand: national concentration is almost nonexistent while local concentration is extreme. The most-recommended firm in the entire panel holds 1.2% of the 18,713 slots, and the top five firms combined hold 3.7%. But zoom into a single city and the picture inverts: about one in twelve recommended firms is locked somewhere, and 40 of those firm-and-city pairs are at a perfect 100%. This market is not won nationally. It is won city by city, and in many cities, it has already been won.

Recommendation strength fades fast below a market’s top firms.

Rank each city’s firms by recommendation rate, from most recommended on down, and compare the same rank across all 50 cities. The ladder is steep. The typical market leader appears in 100% of answers. The typical fifth firm appears in two thirds. The typical tenth firm is down to one answer in four, and by rank 20 a firm is appearing about once in every 25 asks:

Recommendation rate by rank within a city, compared across all 50 cities. The tick is the median city at that rank; the shaded band is the full spread from the lowest city to the highest.

Two things follow from this shape. First, the top of the ladder is crowded with near-permanent firms: ranks one through three all sit above 85% in the typical market, which is why displacing an incumbent is a different job than getting mentioned. Second, mid-pack visibility is worth less than it sounds. Moving from a market’s 20th firm to its 10th makes you about six times as visible, and the biggest single jump on the whole ladder is the climb into the top five.

Position wobbles too, and how much depends on where a firm sits. For every firm that appeared at least twice in a city, we measured how far its spot in the list drifted across its appearances. The locked leaders hold on tightly: centered near rank 3, drifting about 1.2 spots either way. Everyone below them floats in a wider band, roughly 1.5 to 1.7 spots of drift around a deeper average:

Average list position and typical drift for firms in each recurrence tier, across all firm-and-city pairs with at least two appearances. Part of the leaders’ tightness is simple geometry, since a firm near rank 1 has nowhere higher to drift.

The practical read: below the locked tier, a firm does not own a spot in the list. It owns a band, and the goal of the work is to move the whole band up.

Rank and recommendation rate travel together, too. Firms that ever reach first position average a 65.3% recommendation rate across the whole panel; firms that never break into the top three average 7.5%. There is no meaningful consolation prize for fourth place. For a marketing program, the goal isn’t “get mentioned,” it’s get mentioned first.

Google’s fan-out queries for Personal Injury Law Firm Recommendations fall into the same seven topics.

Before writing an answer, the model runs three to five of its own Google searches. These are called fan-out queries, and they are the retrieval layer that decides visibility: whatever ranks for them is what the model reads. We captured 9,308 of them, 4,362 unique, and the pattern is remarkably uniform. Strip out the city names and just 29 templates cover 58% of all search volume. The single most universal query, the only one that ran in all 50 cities, is not a rankings query at all:

Clustering every query semantically produces seven topics, and their shares barely move between markets:

The seven topics behind every fan-out query. Shares are occurrence-weighted: queries the model reaches for repeatedly count more.

Two details matter for content planning. First, 18% of all fan-out volume is year-stamped: the model appends “2026,” sometimes “August 2026,” to its own searches. Content with visible freshness answers the literal queries being run. Second, the topic proportions are nearly identical everywhere. The dominant topic’s spread across 50 cities is only 1.3 times what sampling noise alone would produce. Build content for these seven topics once and you have built it for every market you operate in.

The sources Google’s AI model references are mostly law firms.

Every answer cites its sources, 26,235 citations in our data. We classified all 1,387 cited websites:

Every citation behind the panel’s answers, classified into a 10-type source taxonomy, grouped here for readability.

The headline is the top row. The model’s reading list is mostly firms’ own websites. Whatever AI optimization means in legal, it starts on your own domain, because that is what gets read before the shortlist gets written.

The rest of the layer has sharp edges worth knowing:

• Two directories carry the directory layer. Super Lawyers and Justia account for over three quarters of all directory citations. After those two, directory ROI falls off fast.
• Award sites are read as authority. The third most-cited website in the entire study is an apply-to-be-recognized lawyer awards organization. It was cited more than every bar association and court combined. Official sources, bars and .gov domains, total 0.8% of citations.
• Parasite SEO is measurable. 437 citations flow through 34 sites whose core topic has nothing to do with law: an arts blog, a restaurant site, lifestyle magazines, all hosting lawyer listicles. Small, but not zero. One cited “firm site” appears to be an outright template spam site.

The third-party legal sites doing the heaviest lifting with citations by Google’s AI Overviews

Set the firm sites aside and rank what’s left by citations, and the non-firm layer turns out to be a short list with very different shapes. Two national directories are cited in every single city. The awards site is cited almost as heavily but in only 15 cities. And two of the biggest remaining sources are not legal sites at all: YouTube is cited in 47 of 50 cities and Reddit in 34, which means video content and forum threads are part of the reading list nearly everywhere.

The most-cited domains outside firms’ own websites, across all 2,500 answers. The right column shows total citations and how many of the 50 cities cite the domain at least once.