What the scores actually measure
A trust score of 70% means the AI persona rated trust at 0.7 on a 0–1 scale. It does not mean 70% of your customers trust the page.
Think of these as a short survey of the AI persona. At each analyzed page or step, the model assigns six ratings while reviewing the interface from that persona’s perspective. The ratings help you spot a reaction worth investigating. The findings and screenshots help you understand it.
How a rating is assigned
The model receives the captured screenshot, the persona’s characteristics, and the available task or goal context. Its review instructions ask it to judge the visible experience and return a number for each of the six dimensions below. Persona characteristics guide that judgment; they are not numerical weights in a formula.
The model chooses the rating. Flock checks that the response contains the required numbers in the allowed ranges. It does not calculate these ratings by counting findings, adding severity penalties, or measuring how long someone hesitated.
There is an important limit here: the current instructions define what each dimension means, but do not give the model a calibrated rubric that distinguishes, for example, 0.6 clarity from 0.8 clarity. The number is a synthetic judgment. A five-point increase on the dashboard does not, by itself, establish a meaningful improvement.
The six ratings
Clarity
How understandable the experience is. The model assigns 0–1; higher means easier to understand.
Trust
How believable the experience feels. The model assigns 0–1; higher means a more trusting impression. This rating is not a security assessment.
Engagement
Whether the persona wants to continue. The model assigns 0–1; higher means more interest in continuing. It is not observed retention.
Recommend
Whether the persona would send someone else here. The model assigns 0–1; higher means more willingness to recommend. This is not NPS.
Frustration
How painful the experience feels. The model assigns a whole number from 1 to 10. Higher means more frustration; lower is better.
Happiness
How positively the experience makes the persona feel. The model assigns a whole number from 1 to 10. Higher means a more positive reaction.
These dimensions can disagree. A page can be easy to understand while making a skeptical persona uneasy about its claims. A useful feature can hold someone’s interest while a difficult form frustrates them. There is no formula requiring trust to match clarity or happiness to be the inverse of frustration.
What the percentages mean
The dashboard puts all six ratings on a percentage display. Clarity, trust, engagement, and recommend are multiplied by 100. Frustration and happiness are divided by 10, then multiplied by 100.
For example, 0.7 trust becomes 70%, and 6/10 frustration becomes 60%. The lowest native frustration or happiness rating is 1/10, displayed as 10%. The percentage is a way to display a rating, not a probability or a share of real users.
For a review containing several persona runs, each run with an available rating contributes equally to that metric’s average. A run with more findings does not get more weight. The gauge’s marker shows the mean, the lighter band shows the observed minimum and maximum, and the darker band shows the middle 50% of ratings. Those bands show disagreement among ratings, not a statistical confidence interval. With no available samples, the gauge says No samples.
Page reviews and multi-step flows
A single-page review produces ratings for that captured page. A multi-step flow produces ratings during its analyzed steps; it is not just one survey at the end of the journey.
For the review summary, Flock uses a run’s overall ratings when present. When they are absent, it averages the available step ratings for that run before combining it with the other runs. Each run still contributes one value per metric.
The individual run’s overall-impression panel can show the latest step’s ratings. That panel and the review summary can therefore differ: one describes the latest analyzed screen, while the summary can include earlier steps.
Use the ratings to decide what to inspect
Start with the dimension that changed, then open the relevant persona’s findings and evidence. Look for a concrete explanation: confusing labels, unclear next steps, a commitment that is hard to understand, or a visible error. These are examples of evidence to investigate, not scoring rules or fixed deductions.
Suppose trust falls after a checkout change. Inspect the screenshots and findings for what changed around price, commitment, and reassurance. The useful outcome is identifying a specific problem you can fix and verify. The number alone cannot tell you the cause.
For comparisons, keep the personas, goal, starting page, authentication, and flow bounds as consistent as possible. A different persona or a different point in the journey can change the ratings even when the interface is unchanged. Repeated runs can help you investigate whether a pattern persists; Flock does not currently attach a confidence interval or significance test to these ratings.
Ratings, findings, and abandonment answer different questions
Ratings describe the persona’s overall impression. Findings describe specific issues, with their own severity, confidence, and evidence status. See findings and evidence for those definitions. A finding’s confidence does not establish confidence in a trust or clarity rating. High ratings do not rule out a serious finding.
Abandonment risk is a separate model judgment of low, medium, or high risk. Would abandon is a yes/no judgment. A review’s abandonment rate is the share of available run-level decisions that say yes. It is not measured customer churn or a forecast of your conversion rate.
The current scoring path has no human-survey calibration that turns these six ratings into predictions of customer behavior. Use them as directional feedback alongside the evidence, your product judgment, and research with real users.
Ask your agent to explain a result
With the Flock MCP connection, ask: “Read Flock’s score documentation, then explain the trust and frustration ratings in this review. Show me the findings and evidence that might explain them, and say what remains uncertain.”
The agent can retrieve this guide through get_product_docs with topic: "scores", or read the flock://docs/scores resource. Documentation explains the scoring method; the agent must retrieve the review separately to discuss your actual result.