How to design CSAT surveys for website chat, WhatsApp and Messenger: timing, wording, scales, response bias and how to act on the scores you collect.
Most chat CSAT surveys fail quietly. They pop up at the wrong moment, ask a vague question, and collect scores from only the happiest and angriest customers. This article shows how to design a short, honest CSAT survey for website chat, WhatsApp and Messenger, and how to read the results without fooling yourself.
What CSAT measures in chat, and what it does not
CSAT (customer satisfaction score) answers one narrow question: how satisfied was the customer with this specific interaction? It is a snapshot of a single conversation, not of your brand, your product or the customer's long-term loyalty.
That narrowness is useful. It also means CSAT should not be asked to do the job of other metrics:
- NPS measures willingness to recommend, which is a relationship question.
- CES (customer effort score) measures how hard it was to get something done.
- Resolution rate and first response time are operational measures, not opinions.
CSAT sits best next to these, not instead of them. If you only track one number, you will optimise for that number, sometimes in ways that hurt customers.
Decide what you are scoring: the bot, the human, or the whole journey
In a chat setup with an AI agent and human handoff, "the conversation" can mean three different things. A customer who is answered by the AI, escalated to a person and resolved in the second half may give one score that blends both experiences.
Decide up front which of these you want:
- Whole-conversation CSAT. One question at the end, regardless of who answered. Simple, and closest to what the customer actually felt.
- Segmented CSAT. Tag each rating by whether the chat was resolved by AI only, or involved a human. Same question, different slices.
- Handoff-moment CSAT. A separate check right after escalation, asking whether the transition felt smooth.
For most small teams, option 2 is the sweet spot: one survey, with the "who handled it" segment applied at analysis time. Comparing AI-only and human-assisted scores tells you where to invest, without making customers answer two surveys.
Choose the question and the scale
Keep the question specific
"How was your experience?" invites a verdict on your pricing, your shipping and the weather. Anchor the question to the conversation:
- "How satisfied were you with the help you received in this chat?"
- "Did this chat solve what you came here for?"
Pick a scale that fits the channel
| Scale | Best for | Trade-off |
|---|---|---|
| Thumbs up / down | Messaging apps, mobile, low-friction | Very high completion, little nuance |
| 1 to 5 stars or numbers | Website chat widgets | Standard CSAT maths, a bit more effort |
| Emoji faces (3 to 5) | International audiences | Crosses language barriers, but faces can be read differently across cultures |
| 1 to 10 | Rarely worth it for CSAT | More choices without more insight |
If your customers span several languages, as many Asia-Pacific and international teams do, a visual scale or thumbs reduces translation risk. A numeric scale with the endpoints labelled ("Very dissatisfied" to "Very satisfied") is the safest text-based option.
Add one optional follow-up
A single open text box, shown only after the rating and clearly optional, is where the useful detail lives. A good prompt is: "What could we have done better?" for low scores and "Anything we got right?" for high ones. Do not make it mandatory; a required comment field is one of the fastest ways to lower completion.
Timing and delivery by channel
When and how you ask matters as much as what you ask.
Website chat
Ask when the conversation ends, not mid-flow. Good triggers are the customer saying thanks, a conversation going idle after the last answer, or the customer closing the widget. Avoid asking before the answer has had time to land.
WhatsApp and Messenger
These are asynchronous: customers drift in and out over hours. Sending a survey seconds after the last message can feel abrupt, while sending it a day later means the details have faded. A short delay after the last message or resolution, with a single-tap response, tends to fit the medium. Keep the prompt to one line and one action.
Be mindful that on WhatsApp, the 24-hour messaging window limits when you may send follow-up messages to a customer. Design your survey so it is delivered inside the live conversation or the window, not as a later outbound message.
- 01Conversation resolved
- 02One-tap rating
- 03Optional comment
- 04Thank-you and tag for review
Frequency
Do not survey every conversation if your volume is high and customers return often. Sampling, or suppressing the survey for a customer who has already rated recently, protects goodwill.
Writing the survey so it does not bias the answer
Small wording choices shift scores. A few rules:
- Neutral phrasing. "Please rate our excellent service" nudges the answer. "How satisfied were you?" does not.
- No incentives tied to a high score. Offering a discount for a five-star rating corrupts the data.
- Do not coach agents on scores. If staff plead for fives, you are measuring their persuasion.
- Keep it to one screen. Every extra step loses respondents.
- Say why you ask. "This helps us improve our support" is enough.
Reading the results honestly
Calculating CSAT
The common method counts the share of respondents who chose the top two ratings on a five-point scale.
For example, if 80 customers respond and 60 of them choose 4 or 5, CSAT is 60 / 80 = 75%. That figure is a hypothetical illustration, not a benchmark.
Watch for response bias
Customers who respond are rarely a perfect sample. Very happy and very unhappy people answer more often, while those with a neutral, forgettable experience skip it. Always report CSAT alongside the number of responses and the response rate, and be cautious about drawing conclusions from small samples.
For example, if a shop receives 300 chats a month and 30 customers answer, a swing of a few ratings can move the percentage visibly. Treat month-to-month wobble with scepticism and look for sustained trends.
Segment before you conclude
An overall number hides more than it reveals. Useful cuts include:
- AI-only versus human-assisted conversations
- Topic (billing, shipping, technical issues)
- Channel (website chat, WhatsApp, Messenger)
- Language
- Time of day or day of week
A flat overall score can hide one topic that is failing badly and another that is excellent.
Turning scores into action
A survey that no one reads is theatre. Build a lightweight loop:
- Review every low rating. Read the transcript within a few days.
- Classify the cause. Was it a missing answer, a wrong answer, a slow handoff, or an unreasonable expectation?
- Fix the source. Add or correct knowledge, adjust the agent's instructions, change the handoff rules or retrain staff.
- Close the loop where it makes sense. For serious complaints, follow up personally.
- Re-check the segment a few weeks later to see whether it moved.
Doing this in AI SNS CHAT
AI SNS CHAT gives you the pieces to run a practical feedback loop, with some limits worth knowing. The facts below describe what the product does today.
Ratings. Conversations in the Chat logs area (Activity → Chat logs) carry thumbs up and thumbs down ratings. You can filter, bulk select, flag, escalate, resolve or export conversations, which makes it straightforward to pull low-rated chats for review or hand a batch to a colleague. The Chat logs are where conversations from every channel other than email appear, including website chat widget, WhatsApp and Messenger. Email conversations live in the Inbox instead.
Analytics. Under Analytics you will find Chats, Topics, Question history and Sentiment. Topics and Sentiment are useful companions to a rating: if low ratings cluster around one topic, that is where to look first. Advanced analytics require the Pro plan.
Fixing what the ratings reveal.
- If a low-rated chat shows the agent lacked an answer, add it through Data sources (Files, Text snippets, Website, Q&A and others) or one of the connectors under Channels.
- If a human replied and gave the ideal answer, you can edit a reply into the ideal answer to train the agent, in both the Inbox and Chat logs.
- Use the Playground to test the corrected behaviour before it reaches customers, and Playground A/B model compare if you want to compare models side by side.
- Adjust persona and guardrails under the agent settings if tone or boundaries were the problem.
Handoff quality. Helpdesk → Scheduling (Standard or higher) lets you set shifts per person and timezone, so handoffs are auto-assigned to whoever is on shift. On handoff, the customer message includes a "what happens next" line based on the rota, and the product never invents a wait time. If low scores correlate with slow handoffs, the rota is the first place to check.
Asking for more detail. If you want richer feedback than a thumb, the Custom form and Collect leads actions can capture structured answers, and the Suggested messages action can offer quick reply options. Note that this is a form-based approach you configure yourself, not a dedicated built-in CSAT survey with automatic 1 to 5 scoring or a CSAT percentage report; if you need that calculation, export the logs and compute it in your own spreadsheet. Actions require Hobby or higher, with limits per agent that vary by plan.
Plans and pricing are listed at /pricing, and the full feature set at /features.
Common mistakes to avoid
- Asking before the problem is actually solved.
- Making comments mandatory.
- Comparing your score to an industry number without knowing how it was measured.
- Averaging AI-only and human-assisted chats and never separating them.
- Celebrating a rise that comes from fewer, happier respondents.
- Collecting scores and never reading a single transcript.
Conclusion
A good chat CSAT survey is short, specific and well timed, and it is judged by what you change because of it. Pick one clear question, use a scale that travels across languages, ask at a natural end point on each channel, segment the results, and read the low scores yourself. The number matters less than the loop that follows it.
Frequently asked questions
What is a good CSAT response rate for chat?
There is no universal figure, and published benchmarks vary widely in how they are measured. Track your own response rate over time instead, and aim to improve it by shortening the survey, using one-tap ratings and asking at a natural end point.
Should the AI agent and human agents be scored separately?
Ask one question per conversation and segment at analysis time. Comparing AI-only chats with human-assisted ones shows where to improve without burdening customers with two surveys.
Is thumbs up/down enough, or do I need a 1 to 5 scale?
Thumbs are enough to find problem conversations and work well on messaging apps. A five-point scale gives more nuance and the standard CSAT calculation, at the cost of slightly lower completion.
Does AI SNS CHAT calculate a CSAT percentage for me?
Not as a dedicated report. Conversations carry thumbs up and down ratings in Activity → Chat logs, and you can filter and export them. If you want a CSAT percentage, calculate it from the exported data.


