What the hand-off from interviews to a survey is for
The site's research guides recommend one sequence: explore, then measure. The sequence is set out in the comparison of qualitative and quantitative research. Alongside it, the survey design guide says interviews should usually come first and supply the survey's options and wording, and the interviews tutorial ends by proposing a follow-up, often a survey, for any finding the business needs measured before it commits.
None of them shows the stretch in between, from a table of themes to a first draft of questions. It is a short stretch with three hazards. Findings get lost, because a questionnaire written from the team's assumptions looks the same on the page as one built from the interviews, and nothing records which it was. Findings get over-used, because every theme becomes a question and the survey grows and loses completions. And findings get mistranslated, because an answer about why someone acted becomes a rating question about the same why, which a survey measures badly. The customer research section holds the interviews tutorial and the pretesting tutorial; this tutorial starts where the interviews' synthesis ends and finishes once the draft has gone to the pretest and the explaining step is planned.
Two things follow from the sequence. Interview findings are hypothesesA testable statement of the change to make, the effect expected on a named metric and the evidence-based reason for expecting it. about what people do and what stops them, and a survey built from them is a test that has to be able to come back with the answer that a theme is rare. And the survey's respondents are drawn separately from the interviewees, so the words that worked in conversation have to be checked on people who were not interviewed before the survey goes out. What the survey adds is the sample sizeThe number of users or sessions an experiment needs to detect a given effect reliably. and the make-up that a handful of conversations lack. The design guide holds the question types and the wording rules, and the pretest tutorial holds the testing; this page covers the choices that only arise when the questions come from interviews. The sort into four questions, the word bank, the trace table, the working arrangements, the triggers for going back to conversations and the order of the steps are general practice, this tutorial's own method, and not a rule from the linked guides or the sources they cite.
Before you start
Have the interviews' findings in synthesised form: each one written as an observation, its evidence and what it implies for the decision, with the participants it came from named by code. The evidence table that the AI research synthesis guide describes is the shape to aim for, with a row for each theme holding who raised it, the excerpts for and against and what it means for the decision, whether a person or a model did the coding. If the synthesis was never done, do it first. A survey written from memory of the interviews carries the words of whichever participants were remembered, and the interviews tutorial's synthesis step asks for patterns that cross participants, not the story that stuck.
Have the decision the survey informs, and the line written in the interviews tutorial's first step about what finding would change your mind. As general practice, that line helps you see which findings the survey has to measure and which it can leave. Have the analysis planA survey's plan, written before the questionnaire: the decision, the figures and comparisons the report needs, which question feeds each, and small-group rules. open or ready to write: the outputs as sentences with a gap for the figure, the groups to compare and the question-to-output map, as the survey analysis plan template lays them out. The plan comes before the questions so that a question has to earn its place.
Keep the transcripts or notes to hand with participant codes. As general practice, work from a copy in which names and anything else that identifies a person has been removed. Check what the participants agreed to. The consent tutorial starts by listing every use a study will make of what people say, and reusing their words to write survey options is a use that list has to cover. Where the interviews are already done, whoever owns data protection in the business settles whether that use is covered, and the survey writer does not. The Market Research Society's professional code, as the site's research guides describe it, rests on participation that is voluntary and informed, which is why that list is checked before the words are reused. As general practice, wording that goes into a questionnaire is a plain phrase with no name or detail that would let a respondent or a colleague guess who said it.
Agree who does the work. The person who ran the interviews and the person who will write the survey should sit together for the middle steps, because the first knows what the participants meant and the second knows what the analysis needs. A second reader checks the trace table at the end, without repeating the work.
The steps
Write what the survey must tell the decision that the interviews could not
State in a sentence or two what the decision needs that conversations cannot give: how many customers met a problem, how large a group is, or whether two groups differ. Put each as an output in the analysis plan, written as the sentence the report will contain with the figure missing. If an output could be answered by the interviews alone, because the question is about how or why for a handful of cases, it does not belong in the survey. The point of this step is to stop the survey becoming a second round of interviews with a larger audience and a worse conversation.
Mark each finding with the question it raises
Add a column to the evidence table and mark every finding with one of four questions. How many: how common is the problem among the people the decision concerns. How big: how large is the group of people who were in the situation the finding describes. How different: does the finding hold for one group and not another. Or why: what lay behind it for the person. The first three are questions a survey can answer for a population, and the fourth is one a survey measures badly. The second question, how big, is where the jobs-to-be-done framework earns its place: it says quantitative sources come after the interviews to confirm and size what they found, and, as general practice, a survey can do that for a situation once it has been described in the customers' own terms. The site's diagnosis of interviews that yield no insight already asks the brief to say which questions concern why and which concern how many; this step does the same for each finding.
Rewrite the why findings as things that happened
The design guide says surveys do well on recent behaviour and badly on motives that people have no way of knowing about themselves, which is why the interviews reconstruct episodes instead of asking for predictions. A survey cannot reconstruct an episode, but it can ask whether a specific thing happened. Take each why finding and find the event, condition or circumstance inside it: not "why did you delay the purchase" but "before you chose, did any of these happen". The stated and revealed preference framework sets out why answers about intentions and predictions differ from what people then do. What cannot be turned into something that happened, such as how a person weighed two options, stays with the interviews, and the log records that it was kept there on purpose.
Choose what to measure and set the rest aside
Every question in the survey has to feed an output, so go through the findings that raise a measurable question and keep those where the decision depends on the answer. A finding the decision would ignore, however vivid, is set aside with a note of the reason. A finding from one participant gets a decision of its own. The reporting guide treats a theme from a single participant as a possibility to check rather than a pattern, which is a reason to give it one option in the survey if the decision would change should it prove common, and a reason to set it aside if it would not. Keep the list of what was set aside and why. A reader who asks later why the survey did not cover something deserves an answer from the record.
Build a word bank from the transcripts
For each finding that will be measured, pull the participants' own phrases word for word into a small table with the participant code beside each. Look for three things as you go. Several participants may use different words for the same thing, and because the survey can offer only one wording it has to choose the plainest. One word may cover two different things for different participants, a trap that only shows when the phrases sit side by side. And participants may split what the team had lumped together, in which case the survey should split it as well. Leave out the team's internal product names and labels, since the design guide's advice is to use customers' own words. The word bank is also where a vivid phrase can be kept when it cannot become an option, if the participants agreed to that use too.
Write answer options in customers' words
Turn each measured finding into options a person can tick. Write each as one thing that happened or one thing the customer did, so no option carries two ideas and no option suggests the answer a team hopes for; the leading questionA survey or interview question whose wording contains or suggests its expected answer, producing measurement bias rather than evidence. rules apply to options as much as to questions. Keep two options where participants kept two ideas apart, and merge phrases only when participants treated them as the same. Where the interviews found two sides to a theme, such as customers who valued it and customers who resented it, give each side its own option instead of an average. Then add what the interviews cannot supply, the options the design guide's table of question types and its note on "don't know" call for, and mark each one as added by the survey. The list should let a respondent say that nothing in it applied, which is the way the survey can find that a theme was rare.
Choose the form of each question from the output it feeds
Take the form of each question from the design guide's table of question types, choosing by the output the finding feeds, and record the form beside the question in the trace table.
Order the questions so the options do not prime the answers
An option list teaches respondents what the team thinks matters. The design guide's order rules apply to questions built from interviews, including its rule for the order of options in a list. The design guide points to Nielsen Norman Group's articles on survey design for many of these wording and order effects. As general practice, which applies the design guide's rule of unprompted before prompted, put an open question about what happened first, so that people answer in their own words before the options appear, and the list of options after it.
Trace every question to a finding and an output
Build the trace table described below: each question beside the finding it came from, the participants who raised it, the word-bank phrase behind each option and the output it feeds. A question with no finding is the team's assumption and either goes or is labelled as one. A finding chosen for measurement with no question is a gap. An option with no participant behind it is allowed, because a survey may add what the interviews missed, but it is marked as added. The trace table is what lets a second reader see the hand-off was done, and it is what the next person needs when a result surprises everyone.
Draw the survey's own sample and leave room for what the interviews missed
Interviewees were chosen to cover the cases that matter, not to stand for the population, so the survey needs a frame and a draw of its own, as the tutorial on drawing a sample from a customer list describes, with its own check of respondents against the frame. Write the population in a sentence before the draw. The diagnosis for interview rounds that give no insight makes the matching point: a survey that follows interviews gets its own frame and its own checks, and the diagnosis for unrepresentative results is where respondents are read against that frame afterwards. Keep an open question and the "other" answers in the survey, since the interviews were a handful of people and will have missed something. Those answers are coded against a codeframeThe defined set of codes open-text answers are tagged with, built from a sample, applied consistently, checked by a second coder and counted against a base. as the open-text coding tutorial describes and counted against a base.
Pretest with people who were not interviewed, then plan the explaining step
The words that worked in conversation may mean something different to someone who was not in it. Run the pretesting tutorial with people from the audience who were not interviewed, and use its meaning probes on the options you wrote from the word bank: in your own words, what does this option say happened? Read the "other" answers from the soft launchReleasing a survey or test to a small share of its audience first, then pausing to read the data for problems before the full launch. for options the word bank missed. Finally, plan the explaining stage the comparison describes: a short round of conversations aimed at a surprise, such as an option nobody expected or a group that answers against the others. As general practice, agree before fieldwork which results would send you back to conversations.
What a finding can become in a survey
Not every finding survives the hand-off in the same form. The table shows what a survey can do with each kind of finding and what stays with the interviews, so that the sorting in the second and fourth steps goes quickly.
| What the interviews gave you | What the survey can do with it | What stays with the interviews |
|---|---|---|
| A trigger or circumstance before the decision | Ask whether it happened and measure how common it was, by group | Why it mattered to that person |
| A barrier or worry | Offer it as an option in the customers' words and count who met it | How the person weighed it against the alternative |
| A word or phrase customers use | Supply the wording of options | How its meaning shifts between customers, which the pretest probes |
| A job or situation customers described | Size the group that was in it and compare it across the groups the plan names | The detail of the situation that made it a job |
| An opinion or a prediction of what someone would do | Little: stated intention is weak evidence of behaviour, so ask about the last time instead | Leave it out or use it as a lead |
| A theme from one participant | Check whether it is shared, at the cost of one option | Nothing more until the survey says it is shared |
| Participants who disagreed with each other | Offer each side as its own option and see how the sides split | Why the two sides see it differently |
A survey that carries only the middle column is a short, answerable questionnaire. A survey that tries to carry the last column as well is the second round of interviews described in the first step.
A trace table that shows where each option came from
The trace table stays short when it is built as the questions are written. The table below lists its columns and a closing row, with what each one holds; the other rows follow the order of the questionnaire.
| Column or closing row | What to write | Why it matters |
|---|---|---|
| Question and form | The question number, its wording and its form, named as in the design guide's table of question types | Lets a reader find the item in the questionnaire |
| Finding | The finding from the evidence table that it came from, or "added by the survey" | Shows what was carried across and what was assumed |
| Participants | The codes of the people who raised it | Lets a sceptic see how thin or broad the source was |
| Word-bank phrase | The participants' words behind each option, beside the option as written | Shows how far the options moved from customers' language |
| Output | The sentence in the analysis plan this question feeds | Proves the question earns its place |
| Not measured | A closing row that lists the findings set aside, with the reason for each | Answers the question of why something was left out |
Store the table beside the evidence table and the questionnaire. When the results are written up, the reporting guide asks for the answer first, the evidence with its strength and the limits where a reader will see them; the trace table is where a reader checks how the survey's questions were derived.
A curtain retailer turns its interviews into a survey (illustrative)
Glebask Curtains carries its interviews into a survey
Glebask Curtains, an invented online made-to-measure curtain retailer, wants to know why first-time buyers do not come back for a second pair. Its product-page budget will fund either a guide to measuring a window and its rail or a service that posts fabric swatches, and the choice depends on how many customers hit each problem. The marketing lead interviews 10 first-time buyers who did not reorder. The evidence table holds six findings, each with its participants: three about curtains that were the wrong size for the window, would not hang on the rail or needed an alteration that felt like a hassle, one about the fabric looking different from the photo, one about the lead time for a made-to-measure order, and one about price, which only P7 raised.
The lead marks five of the six findings as questions of how many. The remaining one, why asking for an alteration felt like a hassle, is a question of why. The event inside it, that an alteration was needed, is already counted by the wrong-size option, so only the hassle stays with the interviews, and the log notes that it was kept there deliberately. The lead time is set aside, because no count of it would change which of the two product-page options is funded, and the log says so. The word bank shows that "fit" meant two things: P2, P4 and P9 used it for a curtain that was the wrong size for the window, while P5 used it for a curtain that would not hang on her pole. The survey gets two options, "The curtains were the wrong size for my window" and, in the team's first draft, "The heading did not suit my track", instead of one option that would count both. P3's words that the fabric "looked greyer than in the photo" become "The colour was greyer than it looked in the photo", and P3's other remark, that it looked like a hospital corridor, stays in the word bank, with its participant code.
The options follow an open question, "Tell us what happened when you hung the curtains", and appear in random order with "Other, please specify", "None of these" and "Don't know". Price stays as one option, because one voice cannot be judged without a count and a common answer would reopen the choice between the two product-page options. The size, rail and colour findings stay because a high count on any of them would change which option is funded first. The trace table gives each of the 4 options from the interviews a participant code and a named output, such as the proportion of first-time buyers who ordered the wrong size, and marks Other, None of these and Don't know as added by the survey. The survey goes to 2,000 first-time buyers drawn from the past year's first orders of customers who have not reordered, not from the interview list. A pretest with 5 buyers who were not interviewed shows that 2 do not know what "heading" means, so the rail option, which the team had written in trade words, goes back to P5's own phrase, "The curtains would not hang on my pole".
Of the 310 who respond, 43% choose the wrong size, 12% the colour and 8% the rail, and the price option is chosen by 6%. Of 44 "other" answers, 15 mention creases that would not drop out, which no interviewee raised, so the team books 4 follow-up conversations to explain it and carries the option into the next wave. The measuring guide is funded first and the swatch service is planned after it.
Common mistakes
- Writing the questionnaire from the team's labels. Options in the business's words confuse or steer respondents, as the design guide says of unfamiliar terms, and do not describe what happened to them.
- Turning every finding into a question. The survey grows, and every extra question costs completions.
- Asking why instead of what happened. A rating of reasons asks customers to explain their motives, which a survey measures badly; an event they can confirm is measurable.
- Merging phrases customers kept apart, or splitting what they merged. An option that holds two meanings counts both as one, and the result cannot be read.
- Dropping the single-voice finding without a word. Either give it one option or record why it was set aside; a silent drop leaves no way to tell a considered cut from an oversight.
- Showing the options before the open question. The options tell respondents what the team thinks matters, and their own words come late or only in the other answers.
- Quoting interview counts as survey results. A tally of interviewees describes the people spoken to, as the comparison of the two methods and the reporting guide warn; the survey's figures describe the population its sample stands for.
- Using the interviewees as the sample. They were chosen to cover cases, not to represent a population, and a survey that leans on them inherits the choice.
- Leaving no room for what the interviews missed. A handful of conversations cannot have found everything, so a survey without an open question or an "other" cannot find what was left out.
- Skipping the trace table. It becomes hard to say later where an option came from, why a finding was left out or which words were the customers' own.
