The supplier's dashboard showed three figures, all of them green: 1,240 conversations this month, 92% resolved without human intervention, average rating 4.6 out of 5. The owner of the heating firm looked at it for a while and asked how many surveys had come out of it.
Nobody knew. The dashboard did not measure that.
This is not unusual. It is the default conversation once someone has run a website chat or a WhatsApp bot for a few months and has to decide whether to renew: plenty of numbers, no answer.
Why dashboards measure what they measure
A supplier can only count what happens inside its own system — messages sent, sessions opened, conversations closed without escalation. The outcome you care about, the job in the diary, happens elsewhere: in your calendar or your CRM. Measuring it means joining two places up, and almost no dashboard does that on its own.
So internal traffic gets measured and called performance. The effect is that the biggest number in the report is usually the one that tells you least.
Four numbers do tell you something.
1. Outcomes per hundred conversations
Before measuring anything, define what counts as an outcome. For a private dental practice it is a booked appointment with a slot assigned. For an MCS-certified heat pump installer, a survey booked with an address and a date. For a solicitor, perhaps just a phone number with permission to ring. Write it in one sentence and do not change it mid-quarter: move the definition and you move the result.
The denominator matters too. Do not count widget opens or sessions — count inbound conversations from real people, minus the spam and your own team's tests. Ten bookings out of a hundred conversations and ten out of four hundred are different businesses.
The other three numbers explain why this one goes up or down.
2. Median time until a person replies (and the 90th percentile)
The bot's response time is not a metric. It is always two seconds; measuring it is like measuring whether the lights are on. What does vary, a lot, is how long your team takes once the bot escalates, or when the conversation arrives with no bot in front of it.
Two warnings. First, use the median, not the mean: one enquiry answered the following Monday is enough to inflate the average for the whole week. In the audit Harvard Business Review published in March 2011 covering 2,241 US companies, 23% never replied to a web lead at all, and among those that did, the average response time was 42 hours — an average that describes almost nobody. Second, look at the 90th percentile too, because the customer waiting longer than 90% of the others is precisely the one ringing your competitor.
The same HBR work — 1.25 million leads across 29 B2C and 13 B2B firms — found that contacting a lead within the first hour made you nearly seven times more likely to qualify it than waiting one more hour, and more than sixty times more likely than waiting twenty-four. It is a 2011 study, and American: treat it as a direction, not a promise. The direction has not reversed in fifteen years.
3. Which step people drop out at
This one is not a figure, it is a short list. Count how many people reach each step:
- open the chat,
- write the first message,
- answer the first qualifying question,
- give a contact detail,
- accept a slot or a date.
The step with the biggest drop is your to-do list. Fall-off at 3 means too many questions, or the wrong order. Fall-off at 4 means you are asking for a phone number before giving anything back. Fall-off at 5 means the chat is not the problem: you have no slots to offer, or the price never appeared in time.
None of the other three tells you what to fix. This one does.
4. Cost per outcome
Add up everything that channel costs you in a month: the licence, whatever Meta bills, and your team's hours at their real cost. Divide by the outcomes from point 1.
Meta's share is a genuine line item, not a footnote. Since 1 July 2025 the WhatsApp Business Platform bills per delivered template message rather than per 24-hour conversation, and service messages inside the customer service window are not charged (Meta's pricing documentation). Meta has also announced that from 1 October 2026 more markets move to their own country rate cards. If your supplier will not show you that figure broken out, your cost per outcome is an optimistic estimate. We go into it in what a WhatsApp chatbot really costs.
The number only means something next to another one: compare it with what a booking from Google Ads costs you. If you have never worked that out, the calculator is a reasonable place to start.
The numbers that do not matter
Conversation volume. It goes up when your traffic goes up. It measures your SEO, not your chat.
Containment or deflection rate. The share of conversations that end without reaching a human. The flaw is structural: it counts the customer whose problem you solved and the customer who gave up and closed the window identically. An abandonment is logged as a success. If your supplier leads with this figure, ask how it distinguishes the two.
The thumbs-up. Hardly anyone answers those surveys, and those who do tend to have had an extreme experience. We know of no sound public figure for response rates on UK small-business chats; the ones circulating come from supplier blogs with no published method. Useful for spotting a fire, not for knowing whether the system works.
"Percentage of correct answers." Correct according to whom. If nobody has read a sample by hand, that number was produced by the model marking its own homework.
Getting all four without buying anything
A spreadsheet, one row per conversation, six columns: date and time, channel, gist of the first message, whether it escalated, when a person replied, and what the outcome was. For a business with forty enquiries a month that is ten minutes a week, and it beats any dashboard. The steps in point 3 can be tracked with events in Google Analytics 4 or whatever you already have. Keep the minimum: if you retain conversations to measure them, UK GDPR still applies.
What this does not solve
The four numbers tell you whether the chat works. They do not tell you whether it answers well. We have found no shortcut there: you have to read conversations, a small sample, every week.
Nor does the benchmark you would like exist. Ranges circulate — "3% to 18% conversion", "70% to 90% containment" — from supplier blogs with no sample, no dates and no method. Compare that with the UK Customer Satisfaction Index, which the Institute of Customer Service put at 78.3 out of 100 in July 2026 from 59,500 responses, with its sampling and timing published. That is what a benchmark looks like. There is no equivalent for chat conversion, so your reference point is your own previous quarter.
And beware small numbers. At forty conversations a month, going from eight bookings to twelve may be noise. Let three months run before believing an improvement, and write down what you changed and when: change four things at once and you will not know which one worked.
Zatio is one option if what you want is for these four numbers to fall out on their own: the agent replies immediately on WhatsApp and on your website chat, qualifies, and hands the conversation over with context when it should. We do not promise a number. We do think you should know yours, rather than a green dashboard that never answers the only question that matters.