Synthetic users: what AI personas can and can't tell you

"Synthetic users" are AI-generated stand-ins for real people: a language model given a profile (a role, a situation, goals, doubts) and asked to react to something as that person would. Teams use them to get quick reactions to a page, a message or a product idea without recruiting anyone.

We build a product on this idea, so we have a stake in it. That's also why we want to be clear about the limits. Used well, AI personas are a fast way to find likely problems. Used badly, they're a way to feel validated by a model telling you what you wanted to hear.

What synthetic users are good at

Catching the obvious things you've stopped seeing. After weeks on a page, you can't read it as a stranger. A persona who has never seen the product will ask what it does, who it's for and why they should believe you. Those are the questions most pages leave half-answered.

Taking a perspective consistently. A model told to read as a skeptical buyer will look for proof, and one told to read as a budget holder will look for the price, every time. That makes it easy to see a page from several angles quickly.

Comparing versions. If the panel, the instructions and the scoring stay the same, a change in the result is more likely to come from the change in your page. That makes synthetic feedback useful for "is version B clearer than version A?", even when its absolute scores mean little.

Speed and cost. Minutes instead of weeks, and you can run it on a design before anything is built. That makes it practical for pages that would never get a user test.

Mechanical checks. Contradictions between sections, a call to action below the fold, an ad promise missing from the page. These are checks a model can do reliably and a busy human often skips.

What synthetic users can't do

They aren't your customers. A persona is a model's idea of a type of person, built from what it learned in training and what you told it. It doesn't know your market's specific history, your competitors' latest move, or the one objection that only comes up in your industry.

They can't predict conversion rates. A score from simulated buyers is a diagnostic, not a forecast. A page that scores well can still underperform, for reasons no persona can see: traffic quality, timing, price, brand.

They can be too agreeable, or invent problems. Language models tend to give an answer even when there isn't a good one. Without grounding, a persona may praise a weak page or raise a concern that no real buyer would have.

They share blind spots. Five personas from the same model are not five independent people. If the model misunderstands your category, all five can be wrong in the same way.

They don't replace discovery. If you don't know who your buyer is or what problem they have, synthetic users can't tell you. They react to what you give them.

How to use them well

Ground the personas in your real audience. The more specific the profile (role, company size, current tool, budget), the more useful the reaction. Generic personas give generic feedback.

Look for patterns, not quotes. One persona's complaint is a hint. The same objection from four of five personas, from different angles, is worth acting on.

Keep the setup fixed when comparing. Same panel, same rubric, same instructions. Otherwise you're measuring the prompt, not the page.

Check what can be checked. When a persona says a claim is missing or contradicted, verify it against the page. Good tools do this automatically: in Klar, for example, every quote from a consistency conflict is looked up in the page text, and claims that can't be found are dropped.

Treat the output as hypotheses. Use synthetic feedback to decide what to fix and what to test. Then test it: with real buyers in interviews, with a five-second test, or with an A/B test when you have the traffic (our sample size calculator tells you if you do).

Where synthetic feedback fits

A sensible order for most teams:

1. Synthetic first read: fix the obvious problems on the design or staging page.
2. Real buyers: test the open questions, such as positioning and price, with interviews or user tests.
3. Real behaviour: once live, watch analytics and run A/B tests on the big ideas.

Synthetic users make step 1 fast and cheap, so steps 2 and 3 can focus on the questions only real people can answer. For a comparison of all the methods, see how to get honest feedback on your landing page. For how Klar compares with specific tools, see our comparisons.

How Klar uses AI personas

Klar builds a panel of five buyer personas from the audience you confirm. Each has a different archetype and stance: a skeptic, a budget holder, a busy professional, a quality seeker and an early adopter. Each is interviewed on a screenshot of your page and scores, on an anchored 0–10 rubric, how likely they are to take the next step. You can re-run the same panel after a change, so scores are comparable over time, and test new headlines against it in Copy Lab.

We label the personas as simulated wherever reports are shared. We think that's the only honest way to sell this: as a fast first read that helps you decide what to check with real buyers, not as a substitute for them.