Questions we get, answered.
Including the awkward ones. If something here is missing, ask us and we will add it.
How is this different from just asking ChatGPT or Claude directly?
A general purpose model answers from its training data and gives you one voice with no weights, no segments, no sources and no error bar. userken builds the audience from a specific, inspectable corpus of public reviews, samples each persona several times, weights the result by how large that segment actually is, grounds every reply in real quoted text, and records the review ids used. We also measure the difference: our accuracy work reports an ungrounded LLM baseline, which is exactly the 'just ask the model' answer, next to ours.
What data is it built from?
Public app store reviews today: Apple App Store and Google Play, across the news and browser categories, with more categories being added. Reddit, Hacker News, Bluesky and your own CSV uploads are next. See /sources for live coverage numbers.
How accurate is it?
We publish it rather than claim it. /accuracy explains the two metrics we report (1-MAE and NDAM), how we hold data out by time, the baselines we score against including an ungrounded LLM, and the human ceiling from split half agreement. Until the first benchmark run completes, that page shows a clearly labelled pending state rather than a number.
Can I use my own data?
That is on the way. CSV upload of support tickets, survey verbatims or transcripts will let you build private audiences, mix your data with public sources, and run accuracy reports on your own audiences. Public data is what makes the product work without it.
Does this replace real research?
No, and we would not sell it as that. It replaces the guess you were going to make anyway, and it absorbs the quick questions that never make it into a research wave. Use it to narrow options, find segments and pre-test wording. Use a real sample for anything you will publish as a statistic, anything regulated, and anything where being wrong is expensive.
How do I use it inside Claude?
userken is an MCP server, so you add its URL once in Claude Desktop or on claude.ai and the tools appear alongside everything else you use. Then you ask in plain language and Claude picks the tool: build an audience, ask it a question, run a concept test, open a focus group. /docs has the setup steps and the full tool list. The web app and MCP run the same engines, so a run you start in one is the same kind of object in the other.
What is a persona here, exactly?
A cluster of semantically similar reviews, named and described by a model from the language inside the cluster, carrying its keywords, themes, typical rating range, a real representative quote, and a weight for how large the cluster is relative to the audience. It is a summary of how a group of real people wrote, not an invented character.
How many people are in an audience?
An audience is defined by filters over real reviews, and its size is whatever those filters return. We show that count, and we refuse to build personas from clusters below a minimum size instead of inventing a segment out of a handful of reviews.
Why reviews rather than a survey panel?
Reviews already exist, cost nothing to collect, cover competitors you could never survey, and go back years so you can look at change over time. The tradeoff is that reviewers are self selected and skew toward people with a grievance. We say so on /accuracy and design around it by favouring comparisons over absolute levels.
Can I ask about a competitor's users?
Yes, and it is one of the main reasons to use this. Because the corpus is public, you can build an audience from a competitor's reviews or from a whole category, which is the audience your own customer list cannot give you.
How long does a question take?
Seconds to under a minute, depending on how many personas and samples the run uses. A moderated focus group is interactive: each turn comes back in a few seconds.
What question formats are supported?
Single choice, multi select, Likert scales, ranking, and open ended. Templates wrap those into concept tests, A/B/n tests, name tests, message tests and Van Westendorp pricing sequences. Open ended questions return themes plus verbatims rather than a distribution.
Do I get the underlying quotes?
Yes. Every result carries verbatims per segment and the grounding review ids, so you can go and read the original text. An answer you cannot trace is an answer you should not brief on.
Is it reproducible?
Audiences are defined by readable filters and the runs record the model and the grounding ids, so the same question against the same audience is comparable over time. Sampling is stochastic by design, because a distribution is the point, so expect small run to run variation rather than byte identical output.
What about languages and countries?
Whatever is in the index, which /sources lists. Coverage is currently strongest in English language reviews. An audience cannot represent a market we have not indexed.
Who can see my audiences and runs?
Your own work is scoped to your account, and to your workspace once shared workspaces ship. Uploaded data is private to your workspace. The public audiences built from public reviews are, by design, available to everyone.
How do I get an account?
Accounts are admin provisioned while we are in early access. Request access from the pricing page and we will get in touch.
What does it cost?
See /pricing. The numbers there are introductory and subject to change, and billing is not switched on yet.
What does it cost you to run, and why does that matter to me?
Most of the cost is model tokens, because embedding and clustering run locally. We say so because it explains the pricing shape: usage is metered in credits that map to real compute, rather than priced by how much we think your budget is.
What are you not going to do?
We are not going to publish an accuracy number we have not measured, invent a customer logo, or tell you a synthetic audience is a substitute for a representative sample. The list of things this method does badly is on /accuracy, on purpose.
Still have a question?
Request access and put it to us directly. We would rather answer it than have you guess what this does.