Why AI's “best lawyer” answer keeps changing
The short answer: an AI engine's favorite lawyer is not a ranking that holds. It is a snapshot of a moving system. We asked ChatGPT, Gemini, and Google AI Overviews the same Brooklyn personal injury question every day for six weeks and counted the firms each answer named. The most-named firm changed on every engine, by margins far too large to be sampling noise. Three separate mechanisms move the answers: new models, new answer formats, and shifting sources. Any number about AI visibility is only honest with a date range attached.
Watch six weeks of drift
The chart below is the raw story. Each panel is one engine; each line is one firm's share of the sampled answers, pooled over a trailing 7-day window. Step through it: every stage is one thing an engine did to its own recommendation while no firm did anything.
1 of 5One question, three engines, six weeks
We asked the three engines the same Brooklyn personal injury question every day from July 8 to August 17, 2026 and counted which firms each answer named: 670 answers in all. Each line is one firm's share of the answers in a trailing 7-day window. If an AI recommendation were a stable ranking, every line would be flat.
2 of 5ChatGPT rewrote its shortlist
Jodré Brenecki went from named in 44 percent of ChatGPT's answers to 88 percent, most of the jump landing in a single week around July 20. Over the same stretch, two firms that opened above 75 percent faded into the pack, and only Bendix Law held its spot through all six weeks.
3 of 5Gemini never settled
Gemini's most-named firm changed hands repeatedly. Rosenberg & Rodriguez peaked at 74 percent in early August, then Bendix Law, at 23 percent when the window opened, surged to 74 percent by the end. The vertical rule marks July 21, when Google made Gemini 3.6 Flash the app default; the mix of firms and sources stepped that same day.
4 of 5AI Overviews changed leaders twice
Shares here are of the 149 searches in the window that produced an AI Overview. The firm that opened on top, named in 81 percent of those answers, ended at 25. Bendix Law surged to 91 percent in early August and slipped back. In the final week Jacoby & Meyers took the lead at 66 percent.
5 of 5Three engines, three different movies
Bendix Law is the one firm all three engines kept naming, and each engine told a different story about it: steady on ChatGPT, a late surge on Gemini, a rise and a slip on AI Overviews. None of these firms rebuilt their websites mid-window. The engines moved. Any AI visibility number is only honest with a date range attached.
See the numbers behind the chart
ChatGPT
| Firm | Jul 14 | Jul 21 | Jul 28 | Aug 4 | Aug 11 | Aug 17 |
|---|---|---|---|---|---|---|
| Firm A | 92% | 92% | 80% | 91% | 73% | 57% |
| Bendix Law | 100% | 92% | 83% | 100% | 100% | 97% |
| Jodré Brenecki | 45% | 68% | 100% | 97% | 90% | 89% |
| Firm B | 77% | 77% | 94% | 91% | 90% | 34% |
| Firm C | 12% | 12% | 6% | 3% | 43% | 74% |
| Firm D | 51% | 38% | 34% | 26% | 23% | 40% |
| Firm E | 38% | 18% | 29% | 46% | 13% | 0% |
| Rosenberg & Rodriguez | 3% | 3% | 3% | 0% | 10% | 37% |
| Firm F | 20% | 29% | 29% | 0% | 3% | 0% |
| Answers in window | 65 | 66 | 35 | 35 | 30 | 35 |
Gemini
| Firm | Jul 14 | Jul 21 | Jul 28 | Aug 4 | Aug 11 | Aug 17 |
|---|---|---|---|---|---|---|
| Rosenberg & Rodriguez | 40% | 27% | 46% | 74% | 69% | 57% |
| Bendix Law | 23% | 33% | 6% | 40% | 54% | 74% |
| Firm A | 54% | 45% | 37% | 11% | 9% | 26% |
| Jacoby & Meyers | 37% | 32% | 54% | 46% | 34% | 23% |
| Firm B | 38% | 45% | 23% | 11% | 9% | 17% |
| Firm C | 26% | 37% | 26% | 31% | 40% | 17% |
| Firm D | 3% | 7% | 6% | 37% | 26% | 6% |
| Firm E | 11% | 23% | 31% | 11% | 11% | 14% |
| Answers in window | 65 | 60 | 35 | 35 | 35 | 35 |
Google AI Overviews
| Firm | Jul 14 | Jul 21 | Jul 28 | Aug 4 | Aug 11 | Aug 17 |
|---|---|---|---|---|---|---|
| Firm A | 81% | 79% | 37% | 42% | 50% | 25% |
| Bendix Law | 50% | 43% | 74% | 92% | 88% | 54% |
| Firm B | 71% | 50% | 0% | 4% | 13% | 25% |
| Jacoby & Meyers | 17% | 21% | 53% | 33% | 50% | 67% |
| Firm C | 50% | 29% | 32% | 21% | 21% | 33% |
| Firm D | 4% | 0% | 0% | 42% | 29% | 8% |
| Answers in window | 48 | 14 | 19 | 24 | 24 | 24 |
Each column is the trailing 7-day window ending that day. AI Overviews counts only the searches where an Overview appeared.
How we measure
Ayo asks the three engines the questions New Yorkers type when they need a lawyer, every day, several times per question, and records every firm each answer names and every page each answer cites. For depth, we also run the same question up to 100 times per engine in a sitting, which puts error bars on every share we report. When we say a change is real, we mean it cleared those error bars. Firms shown by name in the chart are the window's risers and holders; firms whose share fell are kept anonymous by policy.
The deep samples are how we first caught this. In early July 2026 we sampled a set of Brooklyn personal injury questions 100 times per question per engine, then ran the identical questions again three weeks later. On one question, the firm ChatGPT had named in 97 of 100 answers appeared in 79 the second time, while The Stein Law Group was named in 96 of 100 and two more firms crossed 90: the slot went from one clear favorite to a four-way contest. The market's monthly snapshots tell the same story: on our public Brooklyn personal injury leaderboard, the firm that sat in eighth place in the July window, Silberstein & Miklos, leads the window ending August 11.
Mechanism 1: model releases move answers overnight
On July 21, 2026, Google made Gemini 3.6 Flash the default model in the Gemini app. In our tracking, the mix of pages Gemini cites for lawyer questions stepped that same day: one Brooklyn firm's share of Gemini citations went from around 2 percent before the release to over 14 percent after. The answers' style and length metrics stayed flat, which points at the model's retrieval changing rather than its writing. That is the vertical rule in the Gemini panel above. A firm watching its numbers that week did nothing differently; the engine changed under it.
Mechanism 2: the answer format itself changes
Around July 20, ChatGPT began attaching a map pack to lawyer questions: a map with pins and star-rated business cards alongside the written answer. Within that week it appeared on 29 percent of our sampled ChatGPT answers for these questions. Days later, OpenAI's deal to bring Yelp listings, ratings, and reviews into ChatGPT became public. A new surface changes who gets seen without any firm's website changing: a map pin is a maps ranking, not a written recommendation, so we count the two separately rather than letting the widget quietly inflate anyone's share.
Mechanism 3: the sources rotate underneath
Between releases and format changes, the ground still moves. In the ChatGPT panel above, Jodré Brenecki climbs from under half of sampled answers to nearly nine in ten, most of it in one week, and a national brand that was a fixture in early July faded out of our samples entirely. Independent data says this churn is structural: SISTRIX measured 74 percent weekly citation churn on ChatGPT Search, against about 5 percent on AI Overviews, across 1.5 million snapshots. The engines do not just disagree with each other; each one keeps disagreeing with its own past self, at its own speed.
What this means for a firm
- Distrust any undated AI visibility number. "ChatGPT recommends us" is a claim about a window, whether the person saying it knows that or not. Our leaderboards state their window and show movement against the prior one for exactly this reason.
- A screenshot is not a measurement. One answer on one day proves almost nothing (why a single spot-check misleads). Share of many sampled answers over a stated window is the number that means something.
- Drift cuts both ways. The same churn that can quietly erase a firm's visibility is what lets a firm that does the right work show up within weeks, not years. Which work, per engine, is its own story: the three engines reward different things, and ChatGPT's levers are the most concrete.
See your own numbers
Ayo tracks how often ChatGPT, Gemini, and Google AI Overviews recommend your firm, question by question, updated daily, with the specific changes that move it. 7-day free trial, no card required.