AI insights

The Insights tab has two halves. Your ratings are counted the moment the page opens. An analysis is the other half: it reads every written response at once and hands back what they add up to — the mood, the topics that keep recurring, and what to do about them.

Ratings at a glance

If your study opens with a quick rating, the top of the tab already shows what those ratings came to. Nothing to generate and nothing to wait for — it is counted from the responses you have, and it updates as more arrive.

Each rating type is summed up the way that type is normally read, with the full breakdown underneath so you can see where the answers actually landed:

Thumbs: The share that came back positive, and the up / down counts behind it.
Rating scale: The average out of your scale’s top, plus how many picked each number. Your own end labels sit against the top and bottom rows.
NPS: The Net Promoter Score — the percentage of promoters (9–10) minus the percentage of detractors (0–6) — with promoters, passives and detractors split out and every score from 0 to 10 listed.
Reaction: The average, with each face and how often it was picked.

Ratings are never AI

These are counts, not judgement. They come straight from what people tapped, so they cost nothing and cannot be wrong about your data.

Generating an analysis

Open the study’s Insights tab and generate. It takes a few seconds. You need at least one written response — rating-only entries carry no words, so they are not part of the AI analysis, though they do count towards the rating summary above it and on your dashboard.

How many responses is enough?

A dozen written conversations already surface real patterns. An analysis covers up to the 200 most recent written responses; past that, the report says so at the bottom.

The report is written in the study’s language. If your visitors and your team don’t share one, turn on Write insights in another languagein the study’s Setup: the interviews stay in the visitors’ language and the summary, topics and actions come back in yours. Customer quotes are the exception — they are never translated, so what you read between quotation marks is always what the person actually typed.

What you get

A row of plain counts sits at the top — total responses, how many completed, the average number of answers per response, and the date range they were collected over. Those, like the rating summary under them, are measured rather than written by the AI. Below them comes the report itself.

An analysis report: counts for responses, completed, average answers and date range; a row of saved version pills; a one-line summary followed by bulleted takeaways and an overall sentiment bar; and recurring topics with a representative quote and response count each — The top of a report: the measured counts, the saved versions, the summary with sentiment, then the recurring topics.

Summary: The headline in one sentence, then three to five takeaways you can scan, plus an overall sentiment and the positive / neutral / negative split as a bar. Read this first: it is the answer to “how did it go?”
Common feedback: The recurring topics, each with how many responses hit it, a sentiment, and a representative quote. Open a topic to read the actual conversations behind it — clicking through jumps you straight to the matching line in a transcript.
Suggested actions: Concrete things to change, each marked high, medium or low priority. This is the part to bring to a planning conversation.
Questions to ask next: What the responses suggest you still do not know — usually a good starting point for the next study.

A list of suggested actions, each with a title, a short explanation and a priority chip reading high or medium — Suggested actions, each with a priority. Topics tell you what is happening; these tell you what to do about it.

Versions

Each run is saved rather than replacing the last one, so you can regenerate whenever new responses arrive and still look back at what you were reading a month ago.

Every analysis is stamped with when it ran and how many responses it covered.
Once there is more than one, a row of version pills lets you switch between them; the newest is marked Latest.
When responses have come in since the latest run, the report says how many are new — your cue to regenerate.
Delete this analysis removes that one version. The responses it was built from are untouched.

Reading it well

Topics tell you what repeats. A single vivid complaint can still matter — open the topic and read the quotes rather than trusting the counts alone.
The analysis only knows what people typed into this study. If the summary feels thin, the fix is usually a sharper goal or more context on the study itself, not another run.
Regenerating after a batch of new responses is cheap and often changes the ranking — priorities move as the sample grows.