CX survey design: a quick-reference handbook

CX survey design is choosing which of four surveys fits your objective, then building it from the right elements in the right order, under rules on scale, sequence and frequency.

0. Introduction

This is a working handbook for CX survey design. It says which survey to run for which objective, what each one is made of, and the rules that govern each.

There is no such thing as "a CX survey". There are four: the conversational survey, the transactional survey, the relationship survey, and the panel benchmark study. They answer different questions and they are built differently. Most of the damage in real programmes comes from applying one survey's rules to another.

All four produce drivers, but the drivers are not equally usable. The panel study cannot reach anyone, by design. The relationship survey is the home of the recommendation question, and it tells you which touchpoint carries weight but never why. Only the transactional survey ties a reason to one event and one identified customer.

The six kinds of question

Every survey in this handbook is assembled from six kinds of element.

---

1. Pick the survey from the objective

This table decides which survey you run. Three of the four you build yourself. The panel benchmark study is built for you by a market research company.

---

2. Choosing your hero question

A hero question is a rating question on a scale. Each of the three below belongs to a different job.

The full comparison of what each scale buys and costs is in the knowledge centre's scale-length article. It is not repeated here.

One rule that does belong here: hold one scale length across a questionnaire. A routing question, which is a yes/no, is the exception, because it is not a rating question.

---

3. The conversational survey

The conversational survey is for getting to granular reasons and root causes when you do not have a ready-made questionnaire. It is one rating question followed by a conversation that the model runs.

The exchange

The hero question, a rating question, is the recommendation question on 0 to 10. Everything after it is the conversation.

Survey: How likely are you to recommend us to a friend or colleague? (0 to 10) Customer: 4 Survey: Thank you. What made it a 4? Customer: I have been waiting three weeks for my account to be verified. Survey: Sorry to hear it has taken that long. What happened when you followed up? Customer: I called twice. Each time they said it was with the back office. Survey: Did anyone give you a date, or tell you what was outstanding? Customer: No. That is the frustrating part.

What the exchange shows:

• Only the first survey turn is written in advance.
• Every later survey turn follows from what the customer has just said. The model does not stop at "waiting three weeks".
• None of the customer's answers came from a list.
• The exchange ends with the customer still volunteering more, so it does not show the exit. The exit is a property of the instrument, set out in the table below.

How it runs

Why it beats a drill-down

A battery of drill-down options tops out at about 12. Past 12 the question stops working. So a drill-down returns only reasons someone thought of in advance, and only about 12 of them. A chat has no cap and no list, so the customer can raise anything, including what nobody at the company considered. The 12 is a practical design limit.

The limit

The approach is new. It is slowly being adopted, and it is not common practice yet.

Why companies hold back:

• It depends on an LLM.
• Changing the instrument changes the scoring and every dashboard built on it.
• Movement is the measure. Holding the method constant is the whole game, and switching breaks your own trend.
• A company that resists is protecting something worth protecting.

Numr has this capability.

---

4. The transactional survey

The transactional survey goes out right after an event to find out what went wrong at that one interaction, so that someone can act on it. It runs on a good rating too, and then asks what made it good. It is the only survey that ties a reason to one event and one identified customer.

The order

Frame the event first, before any question. Then the hero question, then the open end, then the drill-downs.

Why the open end comes before the drill-downs. The customer gets a chance to think about why they gave that rating in their own words, before you show them a list of reasons. Put the list first and you have told them what counts as a reason.

The example: a transactional survey

A transactional survey about an account verification. After the framing line come four elements, in order: a rating question, a text box, a multi-select, then a conditional multi-select.

You called us on 14 March about your account verification. This survey is about that call. Q1, the hero question, a rating question. How easy was it to get your account verified? (1 to 7, every point anchored, Very difficult to Very easy) Q2, the open end, a free-text box. In your own words, what made it that way? Q3, the level one drill-down, a multiple-choice question, multi-select. Which of these affected your experience? Select all that apply. - How long it took - The documents we asked for - The information we gave you - How we kept you updated - Something else (free text) Q4, the level two drill-down, a conditional multiple-choice question. Shown because they selected "How long it took". Which part took the longest? - Reaching someone who could help - The conversation itself - Waiting after that conversation for something to happen

Rules for the transactional survey

• The framing line names the interaction and the date. It asks nothing.
• Level one runs on a positive rating too, not only a negative one. A high rating gets the same question about what made it good.
• Level one reason options are randomised and rotated, with an "other" option, so position in the list does not decide what gets chosen. In the example, "Something else" is the "other" option.
• Keep the list to about 12 options at most. Past that the question stops working.
• Level two opens for each option selected at level one. Three selected means three follow-ups.
• Comment analysis runs on the open end and reaches granular reasons in the customer's own words.
• The limit: a customer can only choose from the options you put in front of them. Your list is the ceiling.
• No recommendation question here. Someone might recommend you after a good support call, but that is not what the question measures.

Timing

---

5. The relationship survey

The relationship survey goes to the whole customer base, sampled. There are two reasons to run one, and the first is the more important.

1. It is the home of the recommendation question. A recommendation is a judgement on the relationship, and the relationship is what this survey asks about, so this is where that question belongs.
2. It establishes what impact each touchpoint has.

The example: a relationship survey

Three elements, in order: a rating question, a text box, then a battery of rating questions.

Q1, the hero question, a rating question. How likely are you to recommend us to a friend or colleague? (0 to 10) Q2, the open end, a free-text box. What is the main reason for your score? Q3 onward, the touchpoint battery, a series of rating questions on one scale. How would you rate your experience with each of the following? (same scale throughout) - Opening your account - Everyday transactions - Getting help when something went wrong - Renewing or upgrading

Rules for the relationship survey

• The hero question goes first, before the battery. Once someone walks through their touchpoints one by one you have switched them into deliberate, rational thinking. The answer you want is the immediate one.
• The battery is a Likert, agree-disagree or satisfaction scale, held constant across every touchpoint. It is not the recommendation scale.
• The battery is not randomised. Its order follows the customer journey, from opening the account to renewing or upgrading, and that sequence is the point.
• No reason question. No drill-down. That belongs in the transactional survey.

An observation

Amitayu Basu has seen across programmes that the immediate answer is the one that best tracks financial outcomes.

What it tells you

---

6. The panel benchmark study

The panel benchmark study is the survey that gives you a comparable score against competitors for the board. You do not build this one. A market research company does.

• Why it is the board number and the competitive benchmark: the double blind.
• Why it is that long: panel members are asked about several companies and several situations in one sitting. It is a market research study that happens to contain your brand.
• It is a good home for the recommendation question, and it is where a competitive comparison comes from. But it answers a different need from the relationship survey, which is where that question belongs first.
• It is expensive, and the cost governs what it can answer. Every respondent costs money, because you are buying external people rather than surveying your own customers. So it cannot be cut finely. Each additional cut needs its own sample. A reading for the whole of the USA is a different thing from separate readings for New York and Chicago, and each city needs a sample of its own. A panel study is therefore run at a holistic, broad level. That is what it is for rather than a shortcoming.
• Two things it cannot do. You cannot close the loop from it. And it cannot be used in linkage, because the respondent is not your customer and there is no transactional record to link them to.

How it differs from the three you build

---

7. How often you may ask

These limits govern how many times one survey may be sent and how often one customer may be invited.

Worked example: the 30-day rule

The example below shows what goes wrong when the last rule is broken.

1. A customer calls about a problem. The ticket is closed. A survey goes out.
2. They call again about the same problem. A new ticket opens. A second survey goes out.
3. Two invitations now sit in their inbox about one unresolved issue.
4. Answer the first and they are rating a ticket that closed while the problem did not.

The measurement is wrong, not merely annoying.

Second reason for the limits

People who answer surveys tend to keep answering. Survey the same people repeatedly and you are hearing the same voices.

Two exceptions

---

8. Quick rules for CX survey design

The rules for building a CX survey, as a checklist.

• Pick the survey from the objective, not from habit.
• One scale length per questionnaire. A yes/no routing question is the exception.
• Hero question, then the open end, then anything else.
• Never let a drill-down come before the open end.
• Recommendation question belongs in the relationship survey, and goes first there. The panel study is also a good home for it. Never on a transactional survey.
• Effort question on transactional surveys.
• Randomise a list when its order is arbitrary. Keep the order when it carries meaning.
• Frame the event on a transactional survey. Never on a relationship survey.
• A routing question is a yes/no. It may report whether something happened. It is never a hero question, never a KPI, and never on an attribute.
• Nothing is compulsory. Mandatory input depresses completion.
• Close the loop from transactional and relationship surveys. You cannot from the panel study.

---

9. Frequently asked questions about CX survey design

Which survey should I run?

Pick it from the objective. Granular reasons and root causes with no ready-made questionnaire: the conversational survey. What went wrong at one interaction, so someone can act on it: the transactional survey. The recommendation question, or which touchpoint carries weight: the relationship survey. A comparable score against competitors for the board: the panel benchmark study.

Where does the recommendation question belong?

In the relationship survey, and it goes first there, before the touchpoint battery. The panel benchmark study is also a good home for it, and the conversational survey uses it as the opener. It never goes on a transactional survey.

Why does the open end come before the drill-downs?

So the drill-down options do not influence it. The customer gets a chance to think about why they gave that rating in their own words before they see a list of reasons. Put the list first and you have told them what counts as a reason.

How many drill-down options can I have?

About 12 at most. Past that the question stops working. At level one, randomise and rotate the options and include an "other" free-text option, and remember that the list is the ceiling: a customer can only choose from what you put in front of them.

How often can I survey the same person?

No more than 3 surveys in 3 months. Anyone who has completed a survey gets nothing for 3 months, and anyone who has been sent one gets nothing for 30 days, whether or not they answered. One survey may go out 3 times at most: the invitation and no more than 2 reminders. Two exceptions: a B2B business with few customers can break the two rules about how many surveys in 3 months but not the 3-sends cap on a single invitation, and premium and high-net-worth segments drop to one invitation every 3 months.

Can I use a yes/no question to measure something?

No. A yes/no is a routing question, not a rating question. It may report whether something happened, but it is never a hero question, never a KPI, and never on an attribute.

What is the difference between a relationship survey and a panel benchmark study?

The relationship survey goes to your own customers, sampled, and you build it. You can close the loop from it and link it to transactional surveys. The panel benchmark study is built by a market research company, answered by paid panel members who are not your customers, and double blind, which is what makes it the board number and the competitive benchmark. You cannot close the loop from it or use it in linkage.

Can I close the loop from any survey?

From the transactional and relationship surveys, yes. Not from the panel benchmark study.