Most chatbot dashboards lead with conversation volume, which is close to useless on its own. Volume rises when the bot is confusing just as reliably as when it is popular, and a busy bot and a productive one look identical until you measure the right things.
This guide covers the handful of numbers that do track to leads, revenue and workload, the supporting numbers worth a glance, how to set a baseline, and what to stop reporting.
Decide what the bot is for
Before you measure anything, name the bot's main job. A bot built to take repeat questions off your team succeeds when it resolves them. One built for lead generation succeeds when it captures contacts worth following up. One on product and pricing pages succeeds when chats are followed by purchases.
Chasing the wrong metric is worse than chasing none, because it pushes you to tune the bot in the wrong direction. Pick the primary job, then treat everything else as supporting evidence.
The four numbers that matter
1. Qualified enquiries produced
The number that connects the tool to revenue. Not conversations, not captured emails: enquiries with enough detail and intent that someone can act on them.
Sort captured leads into two piles. Qualified ones asked about pricing, described a specific project, requested a quote or gave a timeline. Unqualified ones wanted support, checked a policy or grabbed a coupon. Ten of the first kind are worth more than a hundred of the second. In SpideyChat each lead on the Leads page sits next to the conversation that produced it, so you can sort them by reading what was asked.
Three rates sit around this number:
| Rate | How to work it out | What it tells you |
|---|---|---|
| Conversation-to-lead rate | Leads captured divided by conversations | How well the bot turns a question into a contact |
| Qualified lead rate | Qualified leads divided by leads captured | Whether you are collecting buyers or browsers |
| Lead-to-customer rate | Customers divided by qualified leads | Whether follow-up turns leads into revenue |
Read them together. Plenty of leads with a low qualified rate means the bot is collecting the wrong people. A strong qualified rate with weak lead-to-customer conversion means deals are leaking in your follow-up, not in the chat. Divide the bot's monthly cost by leads captured for a cost per lead you can compare with your other channels. SpideyChat's dashboard also shows a funnel from conversations started to engaged, resolved and lead, which points to the step where interested visitors drop away. The ROI guide shows how to turn these into money.
2. Unanswered question rate
The share of conversations where the bot had nothing relevant to say.
This is the most actionable number you have, because every one is a specific content gap. Read the questions, add the missing content and watch the rate fall. SpideyChat's Answer gaps page collects these for you: questions with no match in your content, weak matches, replies that admitted not knowing, and answers a visitor rated with a thumbs down. Answer one there and the bot uses it from then on.
3. Handover rate, and why
The share of conversations that reached a person, split by reason.
Handovers are not failures. A bot that never hands over is either brilliant or stubborn, and stubborn is far more common. What matters is why: "asked for a person" is healthy, "could not answer" is a content gap, and "got stuck in a loop" is a bug. The reason is rarely in the count itself, so read the transcripts behind it.
Look at the quality of the exit too. A good handover captures the question and contact details and reaches your team with the conversation attached. A bad one leaves the visitor with "I didn't understand" and sends them to a competitor. Exits like that are a design problem, covered in designing chatbot conversations that don't frustrate users.
4. Out-of-hours share
What proportion of conversations happen outside your working hours.
This tells you whether the tool is doing the job it was bought for. If nearly all activity falls in hours you were available anyway, you have added convenience, not coverage. SpideyChat's dashboard charts message activity by hour of day over the last four weeks, so this one takes a minute to read. Each out-of-hours enquiry you capture is one that used to wait until morning, which is the cost described in what missed enquiries cost your business.
Two supporting numbers worth a glance
Resolution rate. The share of conversations the bot finished without a person. It is the clearest sign of work removed, but be careful what "resolved" means. Many tools count any conversation that didn't escalate as a success, which puts genuine answers and visitors who gave up under the same label. Pair it with thumbs-down ratings and repeat contacts so it measures answers, not silence. Measuring customer satisfaction with an AI chatbot covers the ratings side.
Conversation-to-conversion rate. Compare the purchase or signup rate of visitors who chatted with those who didn't. People who start a chat are often more interested to begin with, so part of the gap is self-selection. You don't need to prove cause to use it; watch whether the gap widens as you improve the bot, and check it page by page. A bot that converts well on the pricing page and poorly on a product page tells you exactly where to work.
Set a baseline before you change anything
The most common mistake is tuning the bot before you know where you started. Spend the first two to four weeks watching, and write down:
- Unanswered question rate and resolution rate before any tuning
- The ten questions people ask most
- How often the bot hands over, and the usual reasons
- Leads captured, how many qualified, and how many became customers
- Any failure patterns that show up in transcripts
You will see published benchmark figures for "good" resolution or lead rates. Treat them as loose reference points at most. A bot answering opening-hours questions should resolve far more than one handling technical software problems, and a pricing page will produce more leads than a blog post. The comparison that counts is you this month against you last month.
Read the numbers together when you do. Suppose the baseline shows a high resolution rate but poor ratings, and the transcripts reveal the bot giving vague reassurance on questions that need a person. Routing those to your team lowers the resolution rate and raises satisfaction. On paper that looks like a step back; in practice the bot has stopped bluffing, which is the improvement you wanted.
Then change one thing at a time. If you rewrite the greeting, add qualifying questions and move the widget in the same week, you won't know which move helped. Change one thing, wait for enough conversations to judge it, then decide the next.
A worked example
A hypothetical example, with invented figures to show the method. Ledgerline Consulting, an imaginary B2B services firm, logs 320 conversations in its first month and captures 48 leads, a conversation-to-lead rate of 15%. That looks fine until they sort the leads: only 12 fit their target profile, and 3 became clients.
The transcripts explain it. Many contacts were students and job seekers, not buyers. So they add two qualifying questions, about company size and budget, before the bot asks for contact details. The next month, leads captured fall but the qualified rate roughly doubles, and the leads that come through are far more likely to close. The headline count got smaller and the business got better, which is the trade you want. For how to choose questions like these, see how to qualify leads automatically with AI, and for reading the numbers behind them, how to read your chatbot's conversation analytics.
What to stop reporting
Total messages. A vanity number and a billing input, not a performance measure. It goes up when the bot answers badly, too.
Average conversation length. Longer is not better or worse without knowing the outcome. A quick answer and a frustrating loop can look the same.
Raw conversation count. Useful as the bottom half of other rates, not as a headline.
Resolution rate on its own. It counts the visitor who gave up in exactly the same way as the one who got an answer.
The guard is simple: pair every activity number with an outcome number. Conversations next to qualified enquiries, resolutions next to thumbs-down ratings, leads captured next to leads that became customers.
A one-line monthly report
Qualified enquiries: X (last month Y). Unanswered rate: Z%. Handovers: N, mostly [reason]. Out-of-hours share: P%.
Four numbers, one line, and every one is actionable. Review it monthly, read a sample of transcripts weekly, and you will know what to fix next. To see what these numbers look like on a live bot, try the demo.