Persona Feedback.

Four to six personas — diverse in tech proficiency, demographic, and goal — independently rate the page across five dimensions and write their reactions in first-person. Surfaces friction a single reviewer can't see: a power user spots inefficiency, a cautious first-timer spots missing trust signals, a thumb-driven mobile user spots layout problems.

What it does runPersonaFeedback()

Step one: the LLM generates N personas (default 4, tunable to 8) appropriate for the page. Each persona has a name, role, background, list of goals, and list of concerns. Personas are written once per audit so they're consistent across the report.

Step two: for each persona, the LLM looks at the screenshot from that persona's viewpoint and returns a structured rating:

• Usability ★★★★★ — how easy is it to do what I came here to do?
• Visual design ★★★★★ — does this look credible and current?
• Content ★★★★★ — is the copy clear, scannable, accurate?
• Trust ★★★★★ — do I feel safe entering my data / making a payment?
• Accessibility ★★★★★ — can I perceive and operate this with my abilities + setup?

Each rating is a 5-star score with a 2–3 sentence comment, a specific-issues list (what bugged them), and a specific-praise list (what worked for them). The overall persona rating is the average of the five.

Who shows up in a typical run

Personas are generated to match the page. Common archetypes (with example names — the LLM picks fresh ones):

• Maria — Cautious first-time visitor. Wants reassurance, clarity, trust signals. Hates jargon. Notices when the privacy policy is hidden.
• Greg — Power user. Wants efficiency, keyboard shortcuts, advanced filters. Hates extra clicks.
• Sam — Mobile-thumb user. One-handed iPhone navigation. Wants obvious taps, no horizontal scroll, readable text.
• Robin — Keyboard + screen-reader user. Tab-orders the page, listens for landmarks. Notices missing alt text and focus traps.
• Kelly — Price-conscious shopper. Compares plans side-by-side, hunts for hidden fees, reads return policies first.
• Diana — Skeptical evaluator. Procurement / IT-buyer voice. Wants compliance docs, security pages, SOC 2 status, references.

The mix adapts to the site. A B2C consumer page tends to draw Maria + Sam + Kelly; a B2B SaaS dashboard draws Greg + Diana + Robin. Add a customPrompt to steer ("focus on enterprise procurement personas") or to swap the persona set for your audience.

What it surfaces

• "Maria — 4/10 usability: the hero says 'Reimagine your workflow' but I have no idea what this product is. I'd bounce in 3 seconds."
• "Greg — 6/10 usability: I want to bulk-action 50 rows but the only option is to click each one individually."
• "Sam — 3/10 mobile: the sticky cookie banner covers the 'Sign in' button on iPhone 13. Can't dismiss it without first tapping Accept."
• "Robin — 2/10 accessibility: I tabbed to the menu but the dropdowns don't open with Enter. They only respond to mouse hover."
• "Kelly — 5/10 trust: the pricing table says 'starting at $29' but the actual checkout shows $42. Where did the extra come from?"
• "Diana — 6/10 trust: there's a SOC 2 badge in the footer but it links to a 404. Real auditors will check that."
• Praise lists, too: "Greg — visual design 9/10: keyboard shortcuts are documented in a tooltip on every button. Excellent."

Coverage

• These are simulated personas, not real users. They surface high-signal hypotheses, not statistical confidence.
• No real-user metrics — no analytics, no conversion data, no session recordings. Pair with your own RUM tools.
• The screen-reader persona (Robin) is heuristic — for legal/conformance review use the dedicated WCAG audit plus a real screen-reader pass.
• Personas don't navigate the site — they react to a single screenshot. For interactive UX issues, see Exploratory.
• Doesn't substitute for usability testing on sensitive launches (pricing changes, major redesigns) — use it for triage, not final sign-off.

Sample finding

// One persona entry from report.personaFeedback.personas[]
{
  "personaName":    "Kelly — Price-conscious shopper",
  // Star ratings shown to the viewer.  Stored as 0–10 in JSON so half-stars
  // (e.g. 2.5/5 = "★★☆☆☆ ½") are representable when averaged across personas.
  "overallStars":  2.7,    // 5.4/10 → ★★★☆☆ (2.7/5)
  "ratings": {
    "usability":      3,    // ★★★☆☆
    "visualDesign":   3.5,  // ★★★½☆
    "content":        2.5,  // ★★½☆☆
    "trust":          2,    // ★★☆☆☆
    "accessibility":  2.5   // ★★½☆☆
  },
  "comments": "The pricing page promises 'starting at $29' but when I clicked through\nto checkout the total was $42.  The $13 I didn't expect was a 'platform fee'\nin small text below the line items. I'd close the tab and go look at\ncompetitors. Make the all-in price obvious on the pricing card.",
  "specificIssues": [
    "Hidden platform fee at checkout",
    "No annual-vs-monthly toggle on the pricing card",
    "Compare-plans button is below the fold on a 13-inch laptop"
  ],
  "specificPraise": [
    "Free tier is clearly labeled, no credit card required to try"
  ]
}

See also