Synthetic users: what an AI audience can and cannot do
Synthetic users, also called AI personas, are AI characters that model an audience and react to copy, pages or ideas. That is fast and works before launch. What their verdict is worth depends on what they are based on and whether anyone measures their accuracy.
Last updated:
What are synthetic users?
A synthetic user is a character played by a language model, with a role, situation, goals and doubts. Several of them model an audience. You show them copy, a page or a concept and get reactions, objections and a choice between versions.
Providers take different angles. Synthetic Users presents itself on its homepage as a tool for interviews and market research, Uxia as usability testing with AI testers. 99takes checks ad and page copy. The directory SourceForge lists such tools in a separate “Synthetic User Tools” category.
What they are good for
- Shortlisting several versions before one of them costs real budget or real people’s time.
- Finding objections and unclear passages your own team no longer notices.
- Seeing different perspectives side by side: the managing director, the buyer, the sceptical specialist.
- Sharpening hypotheses before you run interviews or a market test.
Where they fail
Synthetic users show no real behaviour. UX studio had ChatGPT review prototypes and concluded it cannot replace usability testing with real people. AIMultiple sees the most practical use in hypotheses and early testing, not in final decisions.
On top of that, personas a model invents freely sound plausible. So check whether they really differ and whether they ever disagree instead of approving every version. How reliable they are depends on what they are based on and whether anyone measures their accuracy.
Five questions for every provider
- What are the personas based on? Real customer voices with a source, or just a description?
- How often were they right? Is there a measured hit rate, with the number of results and a range?
- How different are they? An audience of ten variants of the same person says little.
- Whose verdict counts? Copy everyone likes a little is not the same as copy that convinces the most important buyers.
- Do the personas know what should come out? If the jury learns which version is new, its verdict is no longer blind.
How 99takes answers these questions
- Personas backed by public customer voices, each with its source; competitor pages feed into the research. Without evidence an audience is labelled a hypothesis.
- A hit rate per jury from reported market results, with number and range, published from ten decidable results.
- Typically 10 to 14 personas with a wide spread, a run takes the whole audience (up to 24), and each persona also reads under the others’ jobs.
- The verdict follows the personas the offer is made for; the edge is reported separately (core and edge).
- Versions are checked blind: the jury does not learn which one is new.
99takes runs on Claude and has no language model of its own. It is a simulation before launch, not a market measurement.
Frequently asked questions
Are synthetic users the same as classic personas?
No. Classic personas are profiles a team reads. Synthetic users react to copy and pages themselves and can be questioned. How to build them with AI: create personas with AI.
Can synthetic users replace customer interviews?
Not entirely. They help sharpen questions and shortlist versions. Where real behaviour and real words matter, you need real people.
How many personas does an audience need?
At 99takes typically 10 to 14 that differ clearly even within the same role. A few uniform personas show an average rather than an audience.
What does it cost to try this with 99takes?
You start with 350 credits, no payment details. Setting up your own audience costs 15 credits, comparing two versions with 12 personas about 73.
Related
An audience backed by evidence, in Claude
99takes researches your competitors and public customer voices and builds your own audience from them.
Try it for free