Personas and goals
Send an AI agent to your live site as a kind of visitor, trying to get one thing done, and read how it went.
An audit tells you what a search engine sees. A persona run tells you what a visitor goes through. An AI agent opens your live site in a real browser, acts as one kind of visitor, and tries to do one thing. Afterwards a second model reads what happened and writes a report: a score, where the visitor struggled, and three changes worth making.
Personas and goals
A persona is who the agent pretends to be. It has a name, a profile describing that person, and a device: Phone or Desktop.
A goal is what the agent tries to do, in one sentence, such as "Send an invoice to a new customer." A goal can also say what success means, and it can sign in with a test account when the task needs one.
Personas and goals belong to a project, so they cover all of a product's sites, like its marketing site and its app. Each goal starts on one of the project's sites, and a run can carry on to the others. Any persona can be sent at any goal, so three personas and five goals give you fifteen pairs to try.
What's on the page
Open a project and choose Personas. The buttons at the top are Run once, New persona, New goal and Suggest. Below them are the tabs:
| Tab | What it holds |
|---|---|
| Past runs | Every run in this project, newest first, 20 at a time, with its outcome and score. |
| Scores | A grid of each persona's latest score at each goal. |
| Run once | Start a run now and watch it. |
| Schedules | Runs that repeat daily or weekly. |
| Personas | The site's personas. |
| Goals | The site's goals, and the test account logins they can sign in with. |
Past runs lists each run's goal, persona, outcome, number of steps and score. A run that's still going says Running. Open a run to read its report.
Before you start
A run needs a verified site, at least one persona and at least one goal. If you have an audit from the last week, suggestions can draft personas and goals from your own pages.
Runs act on your live site
The agent clicks, types and submits forms on the real site, the same as a visitor would. Guardrails stop it before it pays for anything or deletes anything, but anything else it does really happens. Use a test account, not a real one.
Scores are a model's judgement
A run's score comes from a model reading the run's log and a few screenshots. It's a considered opinion of one run, not a measurement, and two runs of the same pair can score differently. Look at the friction points and the steps behind a score before acting on it.
Google picked a different canonical page
Google treats another page as the original of these, so it shows that one in search instead, ignoring the canonical the page names or the page itself.
Adding personas and goals
Write the visitors the agent plays and the tasks it tries, change them, and delete them.