Open text questions
How InviziPoll handles long-form answers: character limits, recurring-word themes, cohort filters, on-device AI summaries, why verbatims never leave the workspace, and how to word the prompt.
On this page
Free text is the highest-risk question type in any anonymous survey — not because of how it is stored, but because of what people write in it. The controls below exist for that reason.
An open text question collects a paragraph: "Anything else you'd like to share?" The answer is encrypted in the respondent's browser like every other answer, and you read the verbatim text after unlocking results.
Open text vs. word cloud
They look similar and behave very differently.
| Open text | Word cloud | |
|---|---|---|
| Length | 200–5,000 characters (default 1,000) | Up to 50 characters |
| What you read | The full paragraph | Aggregated tokens only |
| Best for | "Tell us what happened" | "One word for how you feel" |
If you want sentiment at a glance, use a word cloud. If you need the story, use open text — and accept that you are asking people to write something identifying if they are not careful.
Writing the question
Two fields matter:
- The prompt (up to 200 characters). Ask for the thing you will act on. "What is one thing that would make next quarter better?" produces more usable answers than "Any other comments?"
- Helper text (optional, up to 200 characters). This is where you set the safety expectation.
Please avoid naming individuals.is short, and it measurably changes what people write.
Set the character limit deliberately. The default of 1,000 characters is roughly 150 words. Raising it to 5,000 invites the long account of a specific incident — which is sometimes exactly what you want, and sometimes the thing that makes an answer identifiable in a team of nine.
Reading the results
The results panel shows three things:
- How many people wrote an answer.
- Recurring words, counted by how many distinct people used each. A word appears only when at least two different respondents used it — one person's vocabulary is not a theme, and surfacing it would point straight at them.
- The verbatims themselves, in a shuffled order that has nothing to do with when they were submitted, with a CSV download of the answers for your own records.
Two more things apply on top:
- Cohort filters narrow the list. If the poll has demographic dimensions, filtering to a group shows only that group's answers, and a group below the five-response floor locks the whole panel rather than showing a short list. See Demographic cohorts and thresholds.
- An optional summary, computed on your device. On a desktop browser with WebGPU you can download a small language model once and ask it for themes, a short summary, and a sentiment split. The answers never leave your browser to produce it. Full account: On-device AI summaries.
If you turned on anonymous replies for the poll, each verbatim also carries the control to reply to its writer without learning who they are.
Verbatims never leave the workspace
If you share a poll's results publicly, the theme summary and the response count are included. The written answers are not.
A public results link is forwardable and cannot be revoked once shared. Paragraphs written by employees who were promised anonymity do not belong behind a URL that can end up anywhere. This is not a setting.
The same rule holds on every other surface that leaves the admin's browser: the aggregate results CSV and the PDF export carry themes and counts only, which is what keeps the PDF's "Aggregate data only" footer honest (the verbatim CSV is a separate download that exists only in the signed-in results view); a shared slice for a manager is built from a format that has no field for verbatim text; and a series' trend snapshot archives counts, never words. An on-device AI summary is excluded from all of those too, and lands in a PDF only if you tick a box for that one export.
Anonymity in practice
Encryption protects the answer in transit and at rest. It cannot protect an answer whose content identifies the writer — "as the only person on the team who was at the Berlin office in March…" is self-identifying no matter how well it is encrypted.
Reduce that risk by using helper text to ask people not to name individuals, keeping open text out of surveys aimed at very small groups (or pairing it with a higher unlock threshold), and preferring a word cloud when you only need the temperature.