Prepared by Nadali with AI assistance.
A shop asks AI for a customer survey and receives: “How satisfied are you with our fast, convenient collection service?” The sentence assumes the service is fast and convenient before a customer has answered.
This guide is for small-business owners and customer-service teams preparing a short, ordinary customer-experience questionnaire. The deliverable is a reviewed questionnaire, a question worksheet and a record of pilot revisions. It is not a validated scale, a representative market study or a report of measured business improvements. The shop and all examples below are fictional editorial illustrations.
1. Give the draft one job
Imagine a stationery shop deciding what to investigate before changing its click-and-collect instructions. It wants to distinguish difficulty understanding the collection message from difficulty finding the collection point. It does not need an advertising testimonial or a prediction of future sales.
Write a brief before asking AI for questions:
- Decision: which part of the collection journey needs further investigation?
- Respondents: adults who personally collected an order during a specified period.
- Reference: their most recent collection within that period.
- Channel: a self-completed web form.
- Exclusions: deliveries, purchases made entirely in store and collections by somebody else.
- Output: wording, answer options, routing and unresolved design questions.
Use exact start and end dates in the actual form. Keep customer names, contact lists and order histories out of this drafting task. A fictional description of the service is enough.
2. Ask AI to criticise before rewriting
Ask for a diagnosis of each item, then a proposed replacement. That keeps the original problem visible instead of letting a rewrite silently change what you measure.
Our opening sentence has three problems: praise embedded in the question, an unspecified occasion and multiple aspects bundled into one rating. A customer might find collection fast but the instructions confusing. One score would not tell the shop which mattered.
Pew Research Center’s question-writing guidance warns against asking respondents to evaluate multiple concepts in one item. It also explains that wording and earlier questions can affect answers.
Try this editorial prompt:
Draft a short questionnaire for the attached decision and respondent group. First flag leading wording, assumptions, vague time periods and questions that combine separate concepts. Explain what each item would measure. Propose a replacement and complete answer options, including relevant non-experience and uncertainty options. Preserve IDs and show every change. Do not invent customer responses, validation, completion times or business results. Do not send invitations or create a live form. Mark missing information for human review.
This is a writing brief, not a technical access control. Use a drafting environment without customer-contact actions enabled.
3. Protect the meaning of the answer options
For a difficulty rating, offer both directions with equivalent intensity. The Australian Bureau of Statistics describes this balance and recommends explicit uncertainty or inapplicability options where valid.
In our proposed scale, “Neither difficult nor easy” is a substantive middle answer. “I do not remember” means the respondent cannot recall. “I did not read the message” means there is no experience to rate. Do not merge them into “Neutral” to make a chart tidier.
Our worksheet assigns separate codes to these categories and to a skipped item. The codes are labels, not numbers to average. Distinguish skipped questions from questions not shown or never reached. AAPOR also recommends allowing respondents to skip a question or choose an explicit response, such as “don’t know” or “don’t want to answer”, especially for difficult or sensitive questions.
Keep ordered scales in a meaningful sequence, rather than shuffling individual points. For this draft, retain negative-to-positive order throughout. Do not remove the middle answer merely to force a preference; equally, do not add a neutral midpoint to a factual yes/no question.
4. Build a small questionnaire worth testing
The following is a draft, not a ready-to-send instrument. Replace the date placeholders, confirm the shop’s terminology and test the complete form.
Introduction: “We are reviewing how customers collect online orders. Participation is voluntary. Think about your most recent collection between [start date] and [end date]. You may skip questions. Do not include names, order numbers or other personal details.” Add an accurate data-use notice and a completion-time estimate only after checking the real process. Do not promise anonymity unless the collection settings support it.
For this example, recruit only people aged 18 or older. If recruitment does not ensure this, add Q0 before Q1: “Are you 18 or older?” Options: Yes; No; Prefer not to answer. Only Yes continues to Q1; other answers and a skipped item end the questionnaire politely. Do not ask for a date of birth.
- Q1: “Did you personally collect an online order from this shop between [start date] and [end date]?” Options: Yes; No; I do not remember. Only Yes continues. Other answers or a skipped item end this questionnaire politely.
- Q2: “Overall, how difficult or easy was collecting that order?” Options: Very difficult; Somewhat difficult; Neither difficult nor easy; Somewhat easy; Very easy; I do not remember.
- Q3: “Before collecting that order, did you read the shop’s collection message?” Options: Yes; No; I do not remember. Yes goes to Q4. All other responses, including a skip, go to Q5.
- Q4: “How difficult or easy was it to understand the collection instructions in that message?” Use Q2’s six options plus ‘I did not see any collection instructions’.
- Q5: “How difficult or easy was it to find the collection point when you arrived?” Use Q2’s six options.
- Q6, optional text: “Is there anything else you would like us to know about collecting that order?”
Q2 is deliberately an overall judgement, followed by separate diagnostic items. Q6 appears after those topics, so do not describe its comments as wholly unprompted. This form will miss people who attempted collection but never completed it. Investigate that experience separately instead of claiming coverage.
5. Test understanding, then the working process
AAPOR recommends pretesting with people similar to intended respondents, followed by a pilot of survey procedures. These are different checks: understanding the wording does not establish that routing and exports work.
For the first check, ask participants to explain what “collection message” and “that order” meant to them. Ask how they chose an option and whether an answer that fitted their experience was missing. Avoid explaining the intended meaning before they answer.
Then run the actual form through its different routes. Check mobile display, keyboard use, skipped items, completion time and exported codes. The Government Analysis Function’s guidance distinguishes cognitive and usability testing and recommends testing the questionnaire in its intended mode. That page was marked under review when checked.
There is no sample size in this guide that certifies validity. Plan testing around the people, language versions and paths the questionnaire must support. A tiny, convenient pilot cannot establish population percentages or eliminate selection bias.
6. Keep a question worksheet and revision log
Use one row per item with these fields:
- Stable question ID and version.
- Decision it supports and concept measured.
- Exact wording, time reference and eligible respondents.
- Options, codes and routing, including skipped answers.
- Potential misunderstanding and planned test.
- Observed problem, source of the observation and proposed change.
- Reviewer, approval date, retest status and remaining limitation.
For Q4, the worksheet codes can be DV, DS, MID, ES, EV, DK and NO_INSTRUCTIONS, matching the options above. Keep SKIPPED, NOT_SHOWN and NOT_REACHED separate from those answers.
An illustrative revision entry, not a real finding:
“Q3, v0.1: ‘Did you read our notification?’ Possible pilot observation: a participant interprets notification as the payment receipt. Proposed v0.2: ‘Before collecting that order, did you read the shop’s message saying it was ready for collection?’ Decision pending: confirm that this describes the message the shop actually sends. Retest Q3 and Q4 together before approval.”
AI can format the log but must not turn a possible observation into a completed test. Preserve the old text. Record whether a revision changes meaning, answer choices or routing; do not silently pool responses from materially different versions. Keep development responses separate from the final collection unless a qualified analysis plan justifies otherwise.
Common questions
Can AI replace the pilot with simulated customers? No. Synthetic answers may help inspect a route, but they are not evidence of how real customers interpret a question.
Can we translate the questionnaire automatically? Use AI for a draft, then review meaning and test each language. Pew’s translation process aims for equivalent meaning rather than word-for-word matching. Keep IDs, scale direction and routing aligned.
Does a neutral questionnaire make the results representative? No. Who was invited, who could respond and who actually replied still matter. Report those limits rather than calling respondents “the market”.
Before sending
Have a person approve the exact wording, options, routing, privacy information and current version. Resolve ambiguous pilot findings or record why they remain. Keep sensitive research and decisions about individuals outside this lightweight workflow.
The useful result is a questionnaire the team can explain and revise, with an honest record of what was tested. AI helps prepare the draft; it does not supply the customer evidence.
Sources checked on 26 October 2026. The prompt, questionnaire and worksheet are original Nadali editorial suggestions, not a validated research method.
