Which AI model is best for each UX research job? We're testing all of them.
A new frontier model ships almost every week, and they don't perform the same on research work. We're running every one through the jobs researchers do, from writing a screener to synthesising interviews, and publishing the results with our take.
New: The first head-to-head results drop soon. Sign up and they land in your inbox first.
Tania ClarkeAuthor, running the experiment
Harri ThomasChief of Staff, grading the models
Mark SzaboAI product manager, running the model evaluations
Get the first results
Why we're running it
At Great Question, we want to know which AI model is best at which task. So we're putting them all to the test on common UX research jobs. We're also testing a harder question: how well can an AI model simulate a researcher's thinking? We're as curious as you are to see where this goes.
Watch Harri, our Chief of Staff and a former researcher at Meta, introduce the series below.
The models in the test
Seven models from three labs: Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 from Anthropic, GPT-6.1 Terra and GPT-6 Luna from OpenAI, and Gemini 4 Argon and Gemini 3.8 Flash from Google.
What we're testing Eight research jobs, plus one harder question
Every model gets the same brief for each job. Then we see whether it can think like a researcher.
Research plans
Turning a product question into a plan and brief.
Screeners
Questions that find the right participants.
Discussion guides
Interview questions and follow-ups.
Survey questions
Clear questions that don't lead the answer.
Transcript clean-up
Tidying transcripts without changing what was said.
Tagging
Applying a codebook to hundreds of quotes.
Synthesis
Findings across many interviews, each backed by a quote.
Stakeholder readouts
A short summary of what the research found.
Thinking like a researcher
The harder question: Can a model reason through a study the way a researcher would?
How the experiment runs
Sign up once and each stage lands in your inbox.
- Now
Testing underway
Every model is working through the same jobs.
- Next
First results, dropping soon
Every model on every job, with our take.
- Then
The final picks
The best model for each job, and how well each one thinks like a researcher.
Method How we're testing
Real research jobs
Discussion guides, screeners, synthesis, readouts. Not coding or trivia.
Same brief, every model
Each model gets the same task and material, so the model is the only difference.
Graded by people
Our team scores every output by hand.
The AI model live experiment: common questions
Which AI model is best for UX research?
None wins every job, which is why we're testing job by job. On published evidence so far, long-context models lead on synthesis, small fast models on tagging and clean-up at volume, and the most natural writers on anything a participant reads.
Is there a benchmark for AI interview synthesis?
Not a published one. General leaderboards measure writing, coding and factual recall, not whether themes from forty interviews are true to the transcripts. That gap is the reason for this experiment.
Do AI models make things up when analysing research?
Sometimes, and the newest models still do. On Artificial Analysis's hallucination test, GPT-6.1 Terra guesses 38% of the time when it doesn't know the answer, and Gemini 4 Argon 15%. For synthesis, ask for the session behind every finding and spot-check a few before you share anything.
When do the results come out?
Over the coming weeks, job by job. Sign up and each one lands in your inbox as it publishes.
Get the blog in your inbox
Latest insights for customer-obsessed product builders. Once a month, no noise.