AI writes the survey draft. You fix the questions
Describe a survey and Zunoform's AI returns a structured draft — question types, scales, matrix batteries, option lists, and a theme — in your workspace, ready to edit. It is on the Business plan at $24.99/month, and it has a consistent set of blind spots in survey wording that you have to correct by hand. This page shows you exactly which ones.
What you get from a prompt
Give it something like "post-purchase survey for an online plant shop, focus on delivery condition and repeat intent" and you get a form with an opening statement, a handful of rating and choice questions, a matrix if several things need the same scale, an open comment box, and an NPS question when the purpose calls for one. Options are written as real answers, and the whole thing is themed.
That is a real head start on structure. Ordering is usually sensible, question types are usually right, and the tedious part — typing out eight option labels three times — is done. What it is not is a survey instrument. The wording will contain the same handful of errors nearly every time, and those errors quietly change your results rather than obviously breaking anything.
Where this sits in the plans
AI generation requires the Business plan at $24.99/month. Building surveys by hand does not: unlimited surveys, 500 responses a month, matrix questions, rating scales, conditional logic, and 133 templates are all on the free plan with no card.
The four wording problems to fix by hand
These come up in AI-drafted surveys constantly, because a language model optimizes for a fluent, agreeable question — and a fluent, agreeable question is often a biased one.
Leading questions
Models write toward positive sentiment. "How much did you enjoy our fast delivery?" presumes enjoyment and presumes speed, and it will pull your average up. Strip the adjective and the assumption: "How would you rate the delivery?" with a neutral scale. Watch for loaded words — easy, helpful, excellent, generous — sitting inside the question stem.
Double-barrelled questions
"How satisfied are you with our pricing and support?" cannot be answered by anyone who likes one and not the other, and you will never know which they meant. Every time you see an "and" or an "or" joining two different objects in a question, split it into two questions. This is the single most common defect in generated surveys.
Unbalanced or unlabelled scales
Generated scales skew positive — excellent, very good, good, fair — which is four positive-to-neutral points and one weak negative. A balanced scale has as many negative points as positive with a true midpoint. Also decide whether every point gets a label or only the ends; mixing the two across a survey makes answers incomparable.
No escape hatch
AI drafts assume every respondent has an opinion and a valid answer. Real ones do not. Add Not applicable, Prefer not to say, Haven't used it, or Other where honest, and be careful about which questions you leave required — a forced answer is a guess recorded as data.
The pass to run before you send it
Ten minutes on the draft. In our experience this changes more about the quality of the data than anything else you can do to a survey.
- Split every question containing "and" that asks about two different things.
- Delete evaluative adjectives from question stems — fast, friendly, improved, helpful.
- Balance every scale, and use the same scale direction throughout so people are not flipping their mental model mid-survey.
- Add Not applicable or Prefer not to say wherever a respondent could legitimately have no answer.
- Move demographics to the end. AI drafts habitually open with them, and that is where people drop off.
- Check that no early question primes a later one — asking about a specific problem first will color the general satisfaction score after it.
- Cut the survey to what you will act on. A 30-question generated draft usually has 12 questions of real value in it.
- Test it on one colleague before it goes out. Ambiguity is invisible to the person who wrote the question.
What the AI cannot set up
Sampling and routing are yours. The generator may add one to three simple rules — ending early for someone who is not a customer, skipping a follow-up — but a real screener, quota logic, or a branch that sends three audiences down three paths is something you build in the logic panel. It is free, unlimited, and per-question rather than section-level, so a single answer can hide or reveal anything downstream.
The same applies to anything past submission: notifications, Google Sheets export, Zapier, webhooks, and CRM handoff are configured on the form after the draft exists. And if your survey is scored — a maturity assessment, a readiness index — the point values and the grade bands are yours to set, though scoring itself is free and evaluated on the server at submission.
One thing worth knowing if you are replicating an established instrument: do not let a model reproduce a validated scale from memory. Wording and scale anchors in published instruments are load-bearing, and a paraphrase invalidates comparison with published norms. Type those questions from the source, and use the generator for the parts around them.
Surveys already written by hand
Free on every plan, and the wording problems above are already dealt with.
Buying habits, priority ranking, and feature importance for a market study.
A five-question satisfaction check after any customer interaction.
eNPS, satisfaction score, what's working, and an optional follow-up opt-in.
Stance, strength of feeling, and light demographics in one minute.
Questions people ask
Is the AI survey generator free?
No — it is part of the Business plan at $24.99/month. Creating surveys manually is free, including matrix questions, rating scales, NPS on Pro, conditional logic, and 500 responses a month.
What is a double-barrelled question?
One that asks about two things at once, like "How satisfied are you with the price and the service?" Respondents who feel differently about each are forced into a single answer, and you cannot tell which half the answer refers to. Split it into two questions.
Why do AI-generated surveys feel positive?
Language models are tuned to be agreeable, which shows up as complimentary adjectives in question stems and scales with more positive points than negative ones. Both nudge answers upward. Removing adjectives and balancing scales fixes most of it.
Can the AI generate a matrix or rating battery?
Yes. It can emit matrix questions with row labels and a column scale, star ratings, 1-10 scales, and NPS. Check that the scale it chose is balanced and that it is consistent with the other scales in the survey.
How long should a survey be?
Shorter than the draft you get. Ask what decision each question would change; anything that changes nothing is costing you completions. Most generated 25-question drafts do their job at 10 to 12.
Keep reading
Better forms.
Better data.
Build your first form in under 60 seconds. Free forever for personal use, no credit card required.