Personas
How AI personas are generated, configured, and kept realistic.
A persona is a reusable model of a kind of user: what they want, how much patience they have, and what they notice. Personas are the unit you re-run against, so results stay comparable across builds.
Anatomy of a persona#
Every persona carries:
- A goal: what they came to do.
- A temperament: how they behave under friction. Impatient, thorough, skeptical.
- A lens: what they're disproportionately sensitive to. Price, proof, accessibility, visual craft, activation friction.
The lens is what makes a persona useful. A persona with no particular sensitivity produces findings that read like a generic audit.
Grounded vs. stock personas#
Stock personas ship with Fraser and cover common archetypes. Good for a first run.
Grounded personas are generated from your own customer data, feedback, tickets, call transcripts. Fraser extracts recurring pain points, objections, and intents and builds personas that reflect them.
Grounded personas are strictly better when you have the data, because they encode complaints your real users actually made.
Grounding personas in your customer data#
Grounding gets sharper the closer the data sits to real behavior.
- Product analytics (PostHog). Connect PostHog and Fraser reads the product-analytics data you grant access to, the events, funnels, and drop-off points your customers generate, and builds personas from it. A grounded persona behaves like your real traffic, not just like the feedback people wrote down.
- Additional enterprise sources. For enterprise teams we can help set up persona generation from additional data, such as observed testing sessions with real people running your flows. Whoever gathers those sessions is responsible for collecting the data with a lawful basis and the participants' consent.
Your data is not used to train or fine-tune any model. Fraser reads the data you connect and generates personas from it at the time of the run; nothing you provide is used to train Lythe's, a shared, or any third-party model, and it is never shared with other customers. You are responsible for having the rights and consents to use any customer, analytics, or testing data you connect or provide. See Authorization & limits and the Privacy Policy.
Choosing a set#
Pick for contrast, not coverage. Three personas that disagree teach you more than eight that overlap. A reasonable starting set:
| Persona | Lens | Catches |
|---|---|---|
| Impatient multitasker | Speed, step count | Extra steps, slow states, friction before value |
| Skeptical evaluator | Proof, specifics | Vague claims, missing docs, thin trust signals |
| Screen-reader user | Accessibility | Focus traps, unlabeled controls, contrast |
| Budget decision-maker | Price legibility | Hidden pricing, unclear tiers |
| Design-critical reviewer | Craft | Hierarchy, spacing, inconsistent components |
Keeping personas honest#
A persona is a model, not a person. It's most reliable when it reflects behavior your customer data actually shows and least reliable when it's asked to speak for a group you have no data about. If you need to know what a segment thinks and you have no evidence from that segment, talk to them, don't synthesize them.
Was this page helpful?