An AI-written survey can sound polished while steering people toward praise, combining two different experiences, or leaving out the answer they need. Before sending it, check what each question asks people to recall and whether their actual experience fits the choices.

This walkthrough shows how to check AI-written survey questions using a fictional shop-feedback survey. The examples are original teaching drafts, not a validated questionnaire, a live customer study, or evidence that any AI product has passed a test.

Start with the decision the survey should support

Imagine Willow Corner, a fictional neighborhood shop, wants to identify problems with finding products, getting staff help, and waiting to pay. It plans to ask people who visited in person during the past 30 days about their most recent visit.

That brief rules out questions about delivery, shopping visits from years ago, or broad claims about whether the brand is “the best.” Before reviewing the AI output, write down the audience, visit period, and intended use. Delete questions that do not help with that purpose.

1. Remove praise and assumptions from the question

Before: “How easy was it to find everything you needed in our beautifully organized shop?”

After: “On your most recent visit, how easy or difficult was it to find the products you wanted?”

“Beautifully organized” supplies a favorable judgment before the customer answers. The revision removes that judgment and explicitly allows difficulty. It also avoids assuming the person found everything. AAPOR’s survey best practices advise against wording that pushes respondents toward a particular answer.

Now inspect the answers too. Offering only “Very easy,” “Easy,” and “Fine” would undo the improvement. Include a balanced range and a way to indicate that the task did not apply.

2. Separate experiences that could receive different answers

Before: “How satisfied were you with the staff’s helpfulness and the speed of checkout?”

After: Ask one question about staff help and another about checkout waiting time.

A customer might have received useful advice and then waited a long time. One satisfaction score cannot tell the shop which experience needs attention. Pew Research Center’s question-writing guidance explains why questions covering multiple concepts are difficult to answer and interpret.

Splitting the item does not mean you must retain both parts. If checkout is outside the survey’s purpose, remove it instead. Here it matters, so the revised five-question draft below measures it separately. Notice that waiting time and satisfaction with waiting are different measures; choose the one the shop actually needs.

3. Check every answer choice, including the edges

Before: “How long did you wait before being served at checkout?” Choices: “Under 5 minutes,” “5–10 minutes,” “10–15 minutes.”

After: Use non-overlapping intervals, add waits of 15 minutes or more, and provide separate choices for not checking out and not remembering.

Exactly 10 minutes fits two original choices; a 20-minute wait fits none. AAPOR’s survey best practices recommend response choices that cover all reasonable answers without overlapping.

Test the draft with invented edge cases: a visitor who only browsed, someone helped by staff who bought nothing, and someone who cannot recall the wait. These are design checks, not simulated customer findings. Never add invented answers to real survey results.

A revised five-question shop-feedback example

Example introduction: “We would like feedback about your most recent in-person visit to Willow Corner within the past 30 days. Participation is optional. You may skip any feedback question. Please do not include names or contact details in your comments.”

This wording is an illustration. Before use, add a truthful explanation of who receives responses and how they will be used. Do not promise anonymity merely because the form has no name field; check the form’s actual collection settings.

  1. Have you visited Willow Corner in person within the past 30 days?
    Choose one: Yes; No; Not sure.
    If No or Not sure, end with a thank-you rather than asking for a visit rating.
  2. On your most recent visit, how easy or difficult was it to find the products you wanted?
    Choose one: Very easy; Somewhat easy; Neither easy nor difficult; Somewhat difficult; Very difficult; I was not looking for particular products; I do not remember.
  3. On that visit, how helpful or unhelpful was the help you received from staff?
    Choose one: Very helpful; Somewhat helpful; Neither helpful nor unhelpful; Somewhat unhelpful; Very unhelpful; I did not receive help from staff; I do not remember.
  4. On that visit, about how long did you wait at checkout before being served?
    Choose one: Less than 5 minutes, including no wait; 5 to less than 10 minutes; 10 to less than 15 minutes; 15 minutes or more; I did not use the checkout; I do not remember.
  5. What one thing, if anything, would you change about that visit?
    Optional short-text answer. Please leave out names and personal contact details.

This small draft has limits. Question 3 does not establish whether someone wanted help but could not get it. If that is an important decision, revise the design to capture it. Question 4 collects an approximate remembered wait, not a measured checkout duration.

Use AI to find possible defects, then inspect its edits

A focused review prompt is more useful than asking whether a survey is “good.” Try this with fictional or appropriately cleared material:

Review this draft for leading wording, multiple concepts in one question, unclear time periods, missing or overlapping choices, and broken skip instructions. For each issue, quote the item, explain the possible confusion, and propose one revision. Preserve the audience and purpose. Flag assumptions. Do not invent responses, validation results, or research citations.

Compare every revision with the original brief. A rewrite may sound smoother while changing what is being measured. Use the AI assistant acceptance-test worksheet to record expected behavior and observed errors. If the output introduces research claims, follow a separate source-to-claim citation check.

Test the form with people before sending it widely

The U.S. Census Bureau describes cognitive interviewing as a way to learn how people interpret and answer draft questions. For a small feedback form, ask suitable testers to explain what each question means and how they chose an answer. Watch for hesitation, missing choices, and incorrect assumptions.

Test the actual form on a phone, including skipping a question, returning to a previous page, and the No/Not sure exit. Revise confusing items and test again. An AI critique alone does not establish that customers understand the form.

Finally, clear wording does not guarantee an unbiased or representative sample. Who sees the invitation and who chooses to answer still matter. AAPOR treats sampling, nonresponse, and questionnaire design as separate concerns. Describe the responses you actually collected, and avoid presenting voluntary shop feedback as the view of every customer.