Study design
Who does AI name when someone in Texas asks for a car accident lawyer?
This page fixes the study before a single answer is collected. It was published on , a week ahead of the dry run, so the questions, the counts and the rules for what counts as a finding cannot be chosen after seeing the data.
Four Texas metros. Six questions a person asks after a crash, in English and in Spanish. ChatGPT and Perplexity, each asked every question 12 times.
- 1
Design published
November 2, 2026. This page: the questions, the counts and the rules for what counts as a finding.
- 2
Dry run, Houston
November 9, 2026. Every Houston question, both engines, both languages, to test the collection and the scoring before the real window.
- 3
Collection
November 16 to 23, 2026. All 1,152 answers in one seven day window.
- 4
First edition
December 7, 2026. Published here, with the stored answers behind every count.
- 5
Later editions
Then quarterly. The same questions, so each edition can be compared with the last.
The questions
Six questions, asked in each metro
- Who is the best car accident lawyer in [metro]?
- I was hit by an 18 wheeler on [highway] in [metro]. Which law firm should I call?
- Which [metro] personal injury firms work with Spanish speaking clients?
- My insurance company offered a low settlement after a crash in [metro]. Who can help?
- I was rear-ended in [metro] and my neck still hurts a week later. Do I need a lawyer, and which one?
- Which car accident law firms in [metro] will take my case with no fee unless they win?
Each prompt names one metro: Houston (I-10), Dallas (I-35E), San Antonio (I-35), Austin (I-35). The highway is the one in question 2. Question 5 is the soft ask: the person has not decided to hire anyone yet, and it tests whether an engine names a firm at all before that decision.
The Spanish questions are written in Spanish by a native speaker, in the words a Spanish speaking caller uses, not translated from the English six. They will be published on this page before collection starts on November 16. A firm can be named in one language and missing in the other, so each language is scored on its own.
The counts
1,152 answers in seven days
- Engines
- ChatGPT (gpt-5-mini, with web search on) and Perplexity (sonar), both asked through their APIs.
- Asks
- Every question 12 times, on each engine, in each language. Each ask is a fresh conversation.
- Total
- 4 metros × 6 questions × 2 languages × 2 engines × 12 asks = 1,152 answers, collected between November 16 and 23.
- Per cell
- 72 answers for each metro, language and engine.
- Stored
- Every answer verbatim, with its date, engine and model.
What gets measured
Each measure, and when it counts as a finding
- Naming rate
- How often each firm is named, out of every answer in its metro, language and engine, with a 95% range. The ranges are clustered by question, because twelve answers to one question are not twelve independent views of the metro.
- Named and cited
- Recorded separately. A firm can be named in the answer without its website being a source, and a website can be a source without the firm being named.
- Concentration
- The share of all naming that goes to the five most named firms in each metro, reported as measured.
- Stability
- How often the same firm comes back across the twelve asks of one question, for each engine.
- Language gap
- The difference between the English and Spanish naming rates. It counts as a finding only when its 95% range excludes zero and the gap points the same way on both engines.
- Engine agreement
- How many of the five firms ChatGPT names most are also in Perplexity's top five, per metro.
- Change
- From the second edition on, every change from the edition before gets a real or noise verdict, by the rule the Stack uses every month.
Precision
How wide the ranges will be
Twelve asks are enough to see whether a firm is named, and not enough to rank firms that sit close together. Seventy-two answers narrow the range, but not to a point.
| Asked | Named | Rate | 95% range |
|---|---|---|---|
| One question, 12 asks | 6 of 12 | 50% | 25% to 75% |
| One metro, language and engine | 36 of 72 | 50% | 39% to 61% |
| One metro, language and engine | 14 of 72 | 19% | 12% to 30% |
Wilson 95% interval, the same one the simulator draws. The published ranges run wider than these, because answers to one question are clustered. No finding rests on a single question.
Firms
How firms are named
- The ten most named firms in each metro are named, each with its range. The rest are reported as a group.
- A firm known by more than one name is counted under all of them. The table of names is checked by a person, not matched by a script alone.
- EWR Digital, which runs the study, sells AI visibility work to law firms. Firms that are EWR clients are asked about with the same questions and counted the same way, and each is marked as a client in the edition.
- Each edition opens with what it learned about the measurement itself, before any firm is named.
Never published
What this study will not say
- A rank without its range. Two firms whose ranges overlap are not reported as one ahead of the other.
- Why a firm was named. The study counts what the engines said. It does not know why they said it.
- Whether a firm is good. Being named by an engine says nothing about the quality of a firm's work.
Limits
What the design cannot see
- API answers are not app answers. The consumer ChatGPT and Perplexity apps can answer differently from their APIs, and the apps cannot be asked 1,152 times under controlled conditions.
- Models change without notice. Every count carries its date and its model, so a change between editions can be checked against a model change.
- Run to run noise. Houston is asked on November 9 and again in the full window. The difference between the two is published as the study's noise floor.
- Six questions are not every question. They are fixed so editions compare, which means they will miss how some people ask.
Before November 16
If your firm goes by more than one name
A firm that is missing from the table of names is undercounted. If your firm is known by a former name, a partner's name, or a Spanish language name, send those names below before collection starts, and they will be in the table for the first edition.
After an edition is published, a count you think is wrong goes through corrections, where it is checked against the stored answers.
Cite this page
"Who does AI name? Study design." llmvisibilitystack.com, updated November 3, 2026. https://llmvisibilitystack.com/research/who-does-ai-name
To cite the method itself, cite the book: Bertram, M. (2026). LLM Visibility: A Decision-Grade System For Winning AI-Mediated Discovery.