19 min read

Chatbot Survey Design

Five Questions, One Element per Channel, and a Reason to Answer

Quick answer: A survey that a chatbot runs is a conversation, and it fails the way conversations fail: the other person stops answering. So the design starts from the ceiling, not the wish list. Our editorial rule is five questions at most, the one you would keep if you could keep only one asked first, and every question rendered with the element the channel actually has: on WhatsApp three reply buttons of 20 characters or a list of ten rows of 24, on Messenger up to thirteen quick replies of 20 characters, on a website whatever the survey tool draws (Intercom's, read 13 September 2026, caps a survey at "12 questions max" and shows "1 question per step" in its small format). A typed answer is stored and tagged, not parsed. On WhatsApp the survey goes out inside the 24-hour window the customer's own message opened, or not at all unless it is an approved template, and Meta's guidelines say a feedback survey is only a utility template if it names the specific order or interaction it is about. Then the survey is tested on ten people before it is sent to a thousand, because the question that reads well in a builder is the one that gets "?" typed back at it.

The conversational survey entry owns the term and the per-channel limits as facts; this guide owns what to do with them. The single post-chat rating (one tap, at resolution) is the CSAT guide; the 0-to-10 flow and closing the loop is the NPS guide; a form that has to validate what is typed is the multi-turn form guide.

What a chat survey can carry, and what it cannot

Three jobs suit a survey in a chat. Feedback on an interaction that just happened, where the customer is present and the bot has the context (the order number, the agent, the topic). Research with a small number of closed questions: which of three features, which day of the week, which of four reasons. And profile enrichment: one or two facts about the customer that make the next conversation better (company size, role, preferred channel), stored as attributes the way Intercom's help describes ("storing responses as a user attribute").

Three jobs do not suit it. A matrix question (rate each of six items on the same scale) has no rendering on any messaging channel and becomes six questions. Anything that needs the customer to believe the answer is anonymous, because a WhatsApp or Messenger reply is attached to their number or profile by construction, and saying otherwise is untrue. And anything that needs more than a handful of answers from the same person: our ceiling is five questions, because on a messaging channel every question is a message the customer has to want to answer, and a survey that outlasts the customer's patience produces a half-finished record that is worse than none. Intercom's own cap is twelve, and its small format, the one it recommends for "a quick pulse on customer sentiment," shows one question per step; we set the ceiling lower because a widget that stays on the page is not a phone that buzzes five times.

The five questions and their order

The order is a rule of loss: assume every question after the first loses a share of the people who answered it, and put the questions in the order you would grieve losing them.

The one number. The rating, the recommend question, the effort question, or the single choice the survey exists to collect. It is the first message, with the scale on the buttons, and no preamble longer than one line.
The reason. "What is the main reason for that?" as buttons where the reasons are known (three on WhatsApp, more on Messenger or the web), with a typed answer as the last option. The NPS guide makes the same point for its two-question shape: the number says where the customer stands, the reason says why.
The one thing you would change. Free text, optional, one message. This is where the answers that change the product come from, and it is asked third rather than second because a customer who has just tapped a reason will type more than one who has been asked to type cold.
A profile fact, if the survey is for research: role, size, frequency of use, as buttons.
Permission. "Can we follow up on this?" with two buttons. It is last because it is the most skippable and because the answer decides whether the retention playbook's later moments have a permission to spend.

A survey for a single purpose stops at two or three. Five is the ceiling, not the target.

One element per question, per channel

A survey tool on a website draws whatever control it wants. On a messaging channel the survey is limited to the interactive elements that channel's API exposes, and every question type has to be mapped onto one of them before the copy is written, because the element decides how long the labels can be and how many there can be.

Question type	WhatsApp	Messenger	Website (Intercom Surveys)	Website (Tidio)
Yes/no, or three choices	Reply buttons: "up to three predefined replies," label "maximum 20 characters"	Quick replies, "20 character limit"	Multiple choice (large format only) or dropdown	A Flow Decision action after the Agent marks conversation as solved trigger, the reply tagging the contact
1-to-5 scale	Five does not fit three buttons: a list message (ten rows, title "maximum 24 characters," opened by one button of 20) or a typed digit	Five quick replies in one row	Star rating (five, "customizable labels"), 3- or 5-emoji rating, numeric scale with custom start and end	The built-in Customer satisfaction rating: "a scale comprised of five options (emojis)" then a comment field; widget conversations only
0-to-10 (NPS)	Eleven fits neither element; a typed digit with a range check, or a WhatsApp Flow, which Meta's service-message list says can "collect customer feedback"	Eleven quick replies fit under the cap of thirteen, but eleven chips on a phone is a clumsy row; the NPS guide prefers a typed digit	NPS question type, "a 0 to 10 numeric rating scale"	Not a built-in element (the feedback article describes only the five-emoji rating and the Decision action); a Flow with typed input is our inference, not a documented pattern
Pick one of many	List message, "up to 10 rows for all sections combined"	Up to thirteen quick replies	Dropdown / list, options "using commas to separate values"	A Flow with a Decision action, several if the options outrun one message (our inference from the article's template)
Pick several	Not an element; one question per option, or a Flow	Not an element; one question per option	Multiple choice with "Multi-select" on	Not described on the cited page; one Decision per option is our inference
Short free text	A typed reply	A typed reply	Short text entry "up to 255 characters"	The comment field of the rating; a typed reply in a Flow is our inference
Long free text	A typed reply	A typed reply	Long text entry "up to 2,000 characters"	as above

Two design consequences. On WhatsApp, a question with four or five options costs the customer three taps (open the list, pick, send) instead of one, so a scale is worth compressing to three buttons ("Good / Okay / Bad") when the survey does not have to produce a named metric, and worth keeping at its full width when it does; the CSAT entry explains why a compressed scale is still a satisfaction score but cannot be compared with anything measured on another. And a button label of 20 characters holds a word or two, not a sentence: "Price" fits, "The price was higher than I expected" is a question body, not a label. The quick reply entry carries the full per-channel comparison and the buttons vs free text guide the rule for when a typed answer beats a menu.

Writing the question

One idea per question, because a button row cannot answer two. "How was the delivery and the packaging?" has to become two questions or one about the order as a whole. A scale carries its labels on the ends ("Very easy" to "Very hard"), and the same direction on every scale in the survey, or the customer who tapped the left button meaning "good" on question one taps it meaning "bad" on question three. A choice list always has an escape ("Something else") that leads to a typed answer, because a survey with no escape records the least-wrong option as the customer's view. And the question does not lead: "How much did you enjoy the new checkout?" is a marketing sentence with a scale on it.

A typed answer is stored, tagged and read, not parsed. A bot that tries to interpret "it was fine I guess" as a score will be wrong often enough to poison the average, and the sentiment analysis guide sets out when an inferred signal is safe to act on and when it is not. Tidio's feedback flow does the simple thing: the Decision action tags the contact with the reply, and the article suggests sending the tags on to a sheet through Zapier; Intercom's does the equivalent by "storing responses as a user attribute" for every question type "except long text entry." Free text goes to a human reader, in batches, and its value is the sentences, not a number.

When the survey may be sent

On a website the survey appears when the conversation ends or when a page rule fires, and the only design question is whether it interrupts. Intercom's help draws the line for its two formats: the small one "shouldn't disrupt customers from their task" and stays "as a floating banner fixed to the bottom of the page"; the large one is "a large modal fixed to the center of the page" and customers "won't be able to interact with other elements on the page until they complete the survey or dismiss it." On the mobile SDK the distinction disappears: both formats "are displayed full-screen" and "will need to be completed or dismissed." A five-question research survey is a large-format interruption and should be targeted, not broadcast.

On WhatsApp the survey either goes inside the 24-hour customer service window that the customer's own message opened, as ordinary interactive messages, or it has to open with an approved template. Meta's categorization page, read 13 September 2026, lists Feedback Surveys among the utility objectives and attaches a condition that decides the design of the opener: "Specificity of the order or interaction to which these relate is necessary. A general/generic survey or request for feedback will not be approved as utility." Its example names the interaction ("You chatted with us recently about order . How was your experience?"). So the template body carries the order number or the conversation date as a variable, and it carries no offer, because the same page says "a feedback survey with promotional content" is categorized as marketing: billed at the marketing rate rather than the utility rate (the WhatsApp channel guide carries the per-market rates), rationed per recipient, and, as the win-back entry sets out from Meta's per-user limits page, not delivered to US phone numbers (Meta's page does not say whether an open service window changes that). A template categorized as marketing by mistake is not a rejection; since April 2025 Meta approves it "as MARKETING," and a business that keeps filing surveys as utility risks the escalating restrictions the page describes. The reply to the template opens a window, and the remaining four questions go inside it as reply buttons and lists.

On Messenger the survey goes inside the window or not at all for most businesses, since the tags that used to carry follow-ups cannot carry promotion and three of them stopped working in 2026 (the Messenger tutorial has the dates). The end of a support conversation is therefore the cheapest place to ask on every messaging channel. The retention playbook lists the survey as one of its five moments and one of the two where the business has to speak first, which is exactly why this guide tells you to attach it to the moment before, while the customer's own message still holds the window open.

One vendor caveat belongs here because it changes the design. Tidio's built-in rating is "only available for chats that took place inside the Tidio widget; the survey will not work for 3rd-party integrations like Messenger, Instagram, or WhatsApp," which our Tidio review saw in the inbox as "We are not collecting satisfaction ratings on Messenger." A business running Tidio on Messenger builds the survey as a Flow rather than a setting; whether the solved-conversation trigger fires on a Messenger thread is not stated on the cited page and should be tested before it is relied on.

The three biases a chat survey adds

A survey in a chat inherits the usual survey biases (the very happy and the very angry answer more) and adds three of its own.

The window bias. On WhatsApp and Messenger the survey can only reach customers who wrote recently, so it samples the engaged. A research survey sent inside windows measures your active customers, and should say so.

The escalation bias. A conversation that the bot handed to a human ends with a person, and the survey that follows scores the person. The CSAT guide's rule applies to every survey, not just the rating: survey bot-resolved conversations for the bot, agent-resolved ones for the agent, and never pool them.

The button bias. A three-button scale records no middle; a list with an "Other" row records what the customer could not find. The survey's own design decides what it can see, so the closed options are drawn from the free-text answers of the last version, not from the product roadmap.

Report the response rate with every result, and the sample with every score; the chatbot CSAT calculator returns a score with its margin of error from the two counts, which is the shape a survey result should be reported in whatever the question was. A result from fewer than a few hundred answers is a signal to read the free text, not a number to put on a slide.

Where the answers go

Every answer is personal data attached to a phone number or profile, and a survey that asks for a reason can receive health details, complaints about named staff or a card number in the free-text field, because the customer types whatever they think will get them helped. Store answers where the rest of the customer record lives (the attribute, the tag), keep the free text out of any field that is exported by default, and set a retention period; the data privacy guide has the rules by jurisdiction. Question five, the permission to follow up, is what makes a later message lawful and welcome, which is why it is in the survey at all.

The ten-person test

Before the survey goes to a segment, it goes to ten people who have not seen the builder, on the channel it will run on, on a phone. The test is not whether the flow runs; it is what comes back. A "?" typed at question two means the question is ambiguous. A free-text answer at a button question means the options were wrong. A drop after question three means the ceiling is three for this audience. Then the survey is edited and sent to a hundred, and the response rate from that hundred is the number the full send is judged against. The QA testing protocol has the general method; for a survey, the only test that matters is the one with real customers, because the failure mode is silence.

What our fifteen reviews record

The conversational survey entry prints the full tally (searched 13 September 2026 across the fifteen files matched by sample-reviews/*-review.md: survey in 6, csat in 4). What matters for design is what the reviews saw as a starting point: the Landbot review captured a Survey Template among four starters, the BotPenguin review captured Surveys and feedback as a wizard purpose tile and records WhatsApp Flows on the King tier "used for forms, surveys, and product browsing," the Blip review records "post-conversation evaluation surveys" in the default template, the Intercom review captured an "NPS survey" category in the templates library, and the BotPenguin and Chatbase reviews record a post-conversation CSAT prompt on the dashboard. Every one of those is a one-question rating or an empty template; none of the fifteen protocols built a multi-question survey and measured its completion, which is why this guide publishes no completion or response rates of its own. Corrections to editorial@chatbotscape.com.

FAQ

How many questions should a chatbot survey have?

Our editorial rule is five at most, with the one number the survey exists to collect asked first and the reason second. A survey for a single purpose stops at two or three. Intercom's tool caps a survey at twelve; we set the ceiling lower for messaging channels because every question is a message the customer has to want to answer.

How do I show a 1-to-5 or 0-to-10 scale on WhatsApp?

WhatsApp gives a survey three reply buttons (labels up to 20 characters) or a list of ten rows (titles up to 24 characters), read 13 September 2026. A 1-to-5 scale needs a list or a typed digit; a 0-to-10 scale fits neither and needs a typed digit with a range check or a WhatsApp Flow. Compress to three buttons only when the survey does not have to produce a named metric.

Can I send a survey on WhatsApp after the conversation has ended?

Only as an approved template. Meta's guidelines treat a feedback survey as a utility template when it names the specific order or interaction it is about; a generic request for feedback "will not be approved as utility," and a survey with promotional content is categorized as marketing. The reply to the template opens a 24-hour window in which the rest of the survey runs as ordinary messages.

What should the bot do with a typed answer?

Store it, tag the contact, and have a person read it in batches. Do not have the bot turn it into a score; an inferred sentiment is wrong often enough to poison an average. Intercom stores responses as user attributes for every question type except long text; Tidio's feedback flow tags the contact from a Decision action.

Why is my chat survey's response rate so low?

Because it is a chat: the customer has to want to answer each message, the survey only reaches people inside a messaging window, and a three-button scale with no escape gives some customers nothing to tap. Report the rate with every result, test the survey on ten real customers before a full send, and read the free text before trusting a number from fewer than a few hundred answers.

Conversational survey (glossary) — the term, its two senses, and the per-channel limits as facts.
How to improve chatbot CSAT — the single post-chat rating: one tap, at resolution, response rate.
Chatbot NPS guide — the 0-to-10 flow, timing, and closing the loop.
Multi-turn form design — when what you are collecting is data that has to validate, not opinion.
Buttons vs free text — the per-question rule this guide's rendering table depends on.
Chatbot retention playbook — the survey as one of five moments, and the permission it collects.
Sentiment analysis for support — why a typed answer is read, not scored.
Chatbot data privacy guide — what to do with the answers.

Sources

Meta for Developers, WhatsApp Business Platform, Template categorization (developers.facebook.com/documentation/business-messaging/whatsapp/templates/template-categorization, undated on the page; latest dated notice 16 April 2025), read 13 September 2026: the Feedback Surveys utility objective, "Specificity of the order or interaction to which these relate is necessary. A general/generic survey or request for feedback will not be approved as utility"; the example "You chatted with us recently about order . How was your experience? Click below to fill out a short survey"; "a feedback survey with promotional content" among mixed-content templates "also considered marketing"; "Effective April 9, 2025, If you selected UTILITY as the template's category and WhatsApp determined it should be MARKETING, the template is approved as MARKETING"; the Warning, Rate limiting, Utility restriction and Business portfolio level restriction ladder for "consistently misclassifying marketing templates as utility."
Meta for Developers, WhatsApp Business Platform, Interactive reply buttons messages (developers.facebook.com/documentation/business-messaging/whatsapp/messages/interactive-reply-buttons-messages, undated), read 13 September 2026: "up to three predefined replies"; label "Maximum 20 characters."
Meta for Developers, WhatsApp Business Platform, Interactive list messages (developers.facebook.com/documentation/business-messaging/whatsapp/messages/interactive-list-messages, updated 2 July 2026), read 13 September 2026: "up to 10 sections, with up to 10 rows for all sections combined"; row title "Maximum 24 characters"; button "Maximum 20 characters."
Meta for Developers, WhatsApp Business Platform, Service messages (developers.facebook.com/documentation/business-messaging/whatsapp/messages/send-messages, undated), read 13 September 2026: the 24-hour customer service window; Interactive Flow messages to "collect customer feedback."
Meta for Developers, Messenger Platform, Quick Replies Reference (developers.facebook.com/docs/messenger-platform/reference/buttons/quick-replies, undated), read 13 September 2026: "A maximum of 13 quick replies are supported"; title "20 character limit."
Intercom Help, Intercom's survey types (intercom.com/help/en/articles/5973834-intercom-s-survey-types, dated 9 March 2026), read 13 September 2026: small format for questions that "shouldn't disrupt customers from their task," "a floating banner fixed to the bottom of the page"; large format "a large modal fixed to the center of the page," customers "won't be able to interact with other elements on the page until they complete the survey or dismiss it"; "12 questions max"; "1 question per step" and "3 questions per step"; the eight question types and their limits (star rating "customizable labels," emoji "3-emoji or 5-emoji," dropdown options "using commas to separate values," short text "up to 255 characters," long text "up to 2,000 characters," multiple choice "Multi-select" and "only available for large format surveys"); "storing responses as a user attribute" for "all question types, except long text entry."
Tidio Help Center, Collect customer feedback (help.tidio.com/hc/en-us/articles/5473041023772-Collect-customer-feedback, stamped 24 June 2026), read 13 September 2026: "a rating on a scale comprised of five options (emojis)"; "only available for chats that took place inside the Tidio widget; the survey will not work for 3rd-party integrations like Messenger, Instagram, or WhatsApp"; the Agent marks conversation as solved trigger, the Decision action whose reply results in the contact "being tagged," and sending tags "via Zapier to the app of your choice (e.g. Google Sheets)."
Chatbotscape review corpus (the fifteen platform reviews at /reviews), searched 13 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE '\bsurvey' sample-reviews/*-review.md returns 6 (aisensy, blip, botpenguin, intercom, landbot, tidio); grep -liE '\bcsat\b' sample-reviews/*-review.md returns 4 (botpenguin, chatbase, intercom, tidio). Passages cited: landbot-review.md lines 648 to 662; botpenguin-review.md lines 178, 449 and 619; blip-review.md line 574; intercom-review.md line 633; chatbase-review.md line 193; tidio-review.md lines 635 and 739.
Ahrefs Keywords Explorer, US overview, queried 13 September 2026: the figures in this guide's keyword note.
Chatbotscape evaluation methodology. /methodology (continuously updated).

About this guide

Chatbotscape launched in 2026 as an independent review site for chatbot platforms. This guide is part of our SMB chatbot Academy and is written for the owner or marketing lead of a small business who wants to ask customers something through the chatbot they already run, and who needs to know what the channel will let them ask before they write the questions. It reads the survey primitives of three vendors (Meta's WhatsApp and Messenger platforms, Intercom, Tidio) as worked examples and carries observations from six of our own reviews; it does not rank them. Where a platform has a Chatbotscape review, the link goes to it and the review carries the affiliate disclosure.

Methodology

The Meta, Intercom and Tidio pages named in Sources were read in full on 13 September 2026, and every limit and setting name carries its source and date; the review observations come from our own dated review sessions and are cited by line. The corpus counts are ours and the commands are printed. The editorial judgments on this page, and they are judgments rather than measurements: the five-question ceiling, the order of loss, the reading of free text rather than scoring it, the three chat-specific biases and the ten-person test. See /methodology.

Last updated

14 September 2026.