Simulated customers aren’t real customers. Here’s why we still use them.

The most reasonable objection to PersonaQA is also the easiest one to agree with: an AI-generated persona is not a real person.

It has not spent its own money. It does not have a lived disability, a difficult procurement committee, a mortgage depending on the decision or years of experience with the product category. It cannot tell you how all customers will behave.

If simulated-customer testing is sold as a cheaper replacement for people, the objection wins.

I built PersonaQA around a different proposition: simulation is a fast, structured way to challenge a website, identify plausible friction and preserve evidence before or alongside live customer data.

Why this is not just a persona prompt

Asking a general chat model to “act like a cautious buyer and review this website” can produce useful commentary. But the method changes with the prompt, the model may only see selected content, the result is free-form prose and there is usually no clear record of what was actually tested.

A PersonaQA run constrains the task:

• the persona has defined goals, information requirements and evaluation criteria;
• allowed browser actions and journey budgets are bounded;
• the agent navigates the real website in Chromium;
• steps, screenshots and structured datasets are preserved;
• findings must connect to evidence encountered during the run;
• observation is kept distinct from interpretation and recommendation.

The model is still involved. The difference is that it operates inside a testing system rather than an open conversation.

What simulated customers can do well

Challenge a journey before traffic exists

A staging site, new campaign or early product cannot provide historical behaviour. Simulation can still expose missing information, broken paths and avoidable uncertainty.

Apply several defined perspectives quickly

A first-time buyer, mobile user, executive evaluator and accessibility specialist notice different things. Convergence across independent runs is more useful than one generic review.

Generate better hypotheses

Instead of “the pricing page performs badly,” a run can suggest a mechanism: customers can find the plans but cannot map them to their team size or predict the next commitment.

Create a repeatable challenge layer

The same journey definition can be rerun after a release. When methodology and evidence remain comparable, teams can investigate meaningful changes.

What they cannot establish

Simulation cannot tell you the actual proportion of customers who feel something. It cannot establish that a recommendation will increase revenue. It cannot reproduce every cultural, emotional, situational or accessibility experience. It cannot make a legal or compliance judgement on behalf of a qualified reviewer.

Those boundaries should appear in the product, not only in the terms and conditions.

• Use “the simulated customer interpreted” rather than “customers think”.
• Use confidence levels based on evidence rather than certainty language.
• Mark sparse or contradictory real-world data honestly.
• Do not turn a score movement into a regression alert without comparable methodology and confirmation.
• Do not label an automated assessment as regulatory approval.

The evidence model

PersonaQA separates four things that AI products often blur together:

1. Observed: the page, action, error, screenshot or journey step that objectively occurred.
2. Inferred: how this simulated perspective interpreted what it encountered.
3. Corroborated: whether other personas, repeat runs or connected real-world signals support the concern.
4. Recommended: the change suggested in response to the evidence.

This does not eliminate model uncertainty. It makes the reasoning easier to inspect and challenge.

Use simulation with—not instead of—real customers

The strongest workflow uses different evidence for different questions.

• Analytics shows what happened at scale.
• Session replay shows where real sessions encountered visible friction.
• Human research reveals lived experience and unexpected needs.
• Experiments can establish the effect of a defined change.
• Simulation challenges the journey early and makes hypotheses easier to locate and repeat.

When the sources disagree, that is useful. It means the simulated explanation should be downgraded or investigated, not defended.

The standard I want PersonaQA to meet

I do not want the platform to produce impressive-sounding certainty. I want it to produce useful, reviewable evidence.

That means being precise about which persona ran, which pages were visited, what the goal was, what happened, where the interpretation begins and how confident the user should be. It means preserving limitations. It means allowing a reviewer to reject or annotate a finding rather than silently rewriting the evidence.

Real customers should remain the source of truth about real customer outcomes. Simulated customers give teams a way to fail earlier, ask better questions and direct scarce research or traffic towards the places most likely to matter.

That is a narrower claim than “AI replaces user research”. I also think it is a far more useful product.

Methodology note: PersonaQA produces directional behavioural forecasts and evidence-backed diagnostic findings. It does not claim to reproduce population behaviour or replace representative human research.