The numbers
Seven runs, August to October 2026.
124 firms
Every Western New York personal-injury firm we could verify against its own website or bar registry, Erie and Niagara counties, August 2026. Satellite and virtual offices excluded.
24 to 37 named
Of the 124, the number named in at least one of a run's 40 answers. August runs: 29, 28, 24, 25. October runs: 32, 35, 37.
50 ever, 74 never
50 firms appeared in at least one of the 280 answers. 74 firms, three fifths of the market, were not named once in seven runs.
Run by run
What each run found.
Each run is the same 10 questions to the same four assistants, 40 answers. "Names extracted" is every business name the answers contained, including directories, out-of-area firms and firms outside the roster; "roster firms named" counts only the 124. The most-named firm's share shows how concentrated the answers are.
| Run | Names extracted | Roster firms named | Most-named firm, answers of 40 | Roster firms in 5 or more answers |
|---|---|---|---|---|
| 17 August 2026 | 44 | 29 | 28 | 11 |
| 31 August 2026, run 1 | 47 | 28 | 30 | 13 |
| 31 August 2026, run 2 | 45 | 24 | 33 | 12 |
| 31 August 2026, run 3 | 45 | 25 | 34 | 12 |
| 8 October 2026, run 1 | 56 | 32 | 29 | 10 |
| 8 October 2026, run 2 | 67 | 35 | 23 | 12 |
| 8 October 2026, run 3 | 68 | 37 | 27 | 11 |
Distinct roster firms named: 40 across the four August runs, 41 across the three October runs, 50 across all seven.
Read it carefully
Three things the numbers do and do not say.
The spread is the instrument's own
Three runs on the same afternoon, same questions, same assistants: 24 to 27 firms in August, 32 to 37 in October. AI answers move on identical inputs. That is why no single run is ever quoted as a trend, here or in any Janua report.
October is not "improvement"
The per-run count rose from the mid-20s to the mid-30s, but the assistants also produced far more names of every kind: Perplexity went from 17 or 18 distinct names per run to 31 to 33, Gemini from 12 to 14 up to 17 to 23. More talking, not clearer seeing. The firms that were never named in August were, with few exceptions, still unnamed in October.
The answers are concentrated
In every run one firm appeared in 23 to 34 of the 40 answers, 10 to 13 firms appeared in five or more, and most named firms appeared once. When a customer asks AI who to call, a handful of names do most of the answering.
Method
How it was measured.
- 01
The roster
A draft list of 145 firms was checked firm by firm against each firm's own website and the bar registry for personal-injury practice and an Erie or Niagara county office. 114 survived, 10 verified firms were added, duplicates and defunct, defense-side and out-of-area entries were removed, and borderline satellite or virtual offices were excluded. Result: 124. Directory-only listings were deliberately not counted.
- 02
The questions
Five phrasings of what a person actually asks, each put once for "Buffalo, NY" and once for "Western New York": "Who is a good personal injury lawyer in [place]? Please name specific businesses." "I'm looking for a personal injury lawyer near [place]. Which ones should I consider?" "Best personal injury lawyer in the [place] area, who are the top few?" "Recommend a personal injury lawyer near [place] or in the surrounding area." "If a friend in [place] needed a personal injury lawyer, who would you tell them to call?" None of the questions names a firm.
- 03
The assistants
ChatGPT, Claude, Gemini and Perplexity, each with live web search on, each asked all 10 questions: 40 answers per run, every answer stored in full with its date.
- 04
The matching
Every business name in every answer is extracted, then matched against the roster. Directories and review sites, defense and non-injury firms, firms outside the two counties and firms excluded from the roster are left out of the roster count. Spellings the assistants use for roster firms (a partner's name, a short brand) were checked by hand and mapped. Two ambiguous names were not counted.
Until 8 October 2026 Janua's AI Facts page said "44 of the 124" for the first run. That was the count of every name extracted, including directories and firms outside the roster, not roster firms. The roster-matched figure for that run is 29. The figures on this page replace it.
Why it matters
Readable is not the same as recommended.
Most of these 124 firms have working websites that an AI assistant can read. Three fifths of them were still never named when the question was asked the way a customer asks it. Being readable by AI is the foundation; being the name that comes back is a different measurement, and only this kind of measurement answers it.
The same method runs for any market: the questions change ("a family restaurant near me", "a plumber in Amherst"), the roster changes, the engines and the discipline do not. Restaurants and other service businesses are being measured the same way now.
What this page is not
Aggregate only.
No firm is named on this page, and Janua does not publish which firms the assistants name or do not name. A firm on the roster can ask Janua for its own position. Nothing here predicts what an assistant will say tomorrow, and no one can guarantee a specific AI answer, ranking, citation or recommendation.
Find out whether AI can read your business.
The free check reads your website live and shows every check it ran. It does not test whether AI names you; that is the measurement above, scoped and quoted with the founders.