Security notice: The Life Inside platform is not affected by the PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619. No action is required for your use of Life Inside. Assessed 23 October 2026. Questions: legal@lifeinside.io

Chatbot analytics: which metrics matter and how to act on them

Chatbot analytics tells you whether the chatbot or AI agent on your website is doing its job: answering visitors, showing you where your content falls short and turning conversations into leads. Most dashboards show plenty of numbers. This guide covers the eight that matter, how to read them without fooling yourself, and what to change when they move.

It is written for teams that run a chatbot or AI agent on a website, in marketing, sales, recruitment or customer service. At the end we show how AgentLoop, the analytics layer in Life Inside, handles each one.

Chatbot metrics: the short list

Conversation count is the number everyone looks at first, and it tells you the least. A busy chatbot can be failing most of its visitors. These eight metrics tell you whether it works.

  • Answer rate: the share of questions the bot answered from its knowledge, rather than falling back to "I don't know".
  • Resolution rate: the share of engaged conversations that ended with the visitor's need met.
  • Unanswered questions and knowledge gaps: what visitors asked that the bot could not answer, grouped by topic.
  • Containment and handover: how many conversations the bot handled on its own, and how many it passed to a person, a form or a booking link.
  • Conversation-to-lead rate: the share of conversations that produced a lead, a booking or an application.
  • Drop-off: where visitors leave a conversation or a video before reaching a next step.
  • Topics and sentiment: what people talk about, and how they feel while doing it.
  • CSAT: what visitors say when you ask them to rate the conversation.

Answer rate and resolution rate

Answer rate is the simplest health check. Microsoft defines a close cousin, the generated answer rate, as the share of user questions that got a generative answer instead of falling through to a fallback or unanswered state. If it drops, either visitors are asking new things or your content changed.

Resolution rate goes a step further: did the conversation end with the need met? In Microsoft Copilot Studio, resolution rate is the share of engaged sessions that ended resolved, either confirmed by the user or implied when the user stopped responding and the flow judged it a success.

Two cautions. First, an answer is not a resolution. A bot can answer every question and still leave visitors without what they came for. Second, implied resolutions flatter you. A visitor who stops replying may be satisfied, or may have given up. Read a sample of implied resolutions every month before you trust the rate.

Check the denominator too. Some tools divide by all sessions, others by engaged sessions only. The same traffic can produce very different percentages, so compare your own trend over time rather than another vendor's headline number.

Unanswered questions and knowledge gaps

This is the most useful report in chatbot analytics, because it tells you exactly what to fix. Every question your bot could not answer is a visitor telling you what your website is missing.

An AI agent should say so when it does not know, rather than guess. That honesty is what makes gaps visible. A bot that guesses hides its gaps inside wrong answers, and you only find them when a customer complains.

  1. Pull the unanswered questions for the last week or month.
  2. Group them by topic. Ten wordings of "do you offer remote roles?" are one gap, not ten.
  3. Rank the groups by volume and by value. A rare question from a buyer about pricing can matter more than a common one about opening hours.
  4. Fix the source: add or update the page, document or FAQ the bot learns from. Do not patch the bot with a one-off answer that drifts from your website.
  5. Check the same topic a week later. If questions still go unanswered, the fix did not land.

Containment vs handover

Containment is the share of conversations the bot handled without a human. IBM describes containment rate as how often the AI solves customer help requests without intervention from human agents. Its mirror image is escalation or handover rate.

For a support team that counts cost per ticket, high containment is the goal. On a website the picture is different. When a qualified buyer asks to talk to sales, or a candidate wants to speak to a recruiter, handing over quickly to the right contact or booking link is a success, not a failure.

So split handovers by reason. Wanted handovers (book a meeting, talk to sales, apply) are wins to grow. Forced handovers (the bot could not answer, the visitor got frustrated) are gaps to close. One number that mixes both tells you very little.

Conversation-to-lead rate and drop-off

Conversation-to-lead rate is the share of conversations that end in a lead, a booked meeting, a form or an application. It is the metric that links a chatbot to revenue or hiring, and the one your leadership will ask about.

Track it by entry page and by topic. A pricing page conversation and a careers page conversation should not share one target. Then look at drop-off: the point in a conversation, or in a video, where visitors leave. If many visitors ask about price and then leave, your answer to the price question is the place to work.

Journeys help here. Seeing the path a visitor took, from landing page to question to video to booking, shows where the chatbot moves people forward and where it stalls them.

Topics, sentiment and CSAT

Topics show demand in visitors' own words. They are often the best market research you have: what buyers ask before they buy, what candidates worry about before they apply. Share the top topics with marketing, sales and HR, not only with whoever runs the bot.

Sentiment adds tone. Microsoft describes its sentiment metric as the share of sessions with negative user sentiment, which helps you find friction even in resolved conversations. Watch the trend and read the negative conversations. They usually point to one confusing policy, price or process.

CSAT is the classic satisfaction score, typically a 1 to 5 rating at the end of a conversation. It is useful, but few visitors answer, and those who do are often the very pleased or the very annoyed. Treat it as a signal to read alongside sentiment and resolution, not as the verdict.

How to measure chatbot performance: a weekly routine

Metrics only help if someone acts on them. Set aside half an hour a week and work through the same list.

  1. Look at the trend in answer rate, resolution rate and conversation-to-lead rate. Note anything that moved.
  2. Open the unanswered questions. Pick the top three gaps and assign each one to the owner of that content.
  3. Read five conversations with negative sentiment and five that ended in a lead. Note what made the difference.
  4. Check handovers by reason. Did wanted handovers reach the right person or booking link?
  5. Check drop-off on your most important page or video. Change one thing and see whether it moves next week.
  6. Share the top topics with the teams that own the answers.

Chatbot analytics tools: what to look for

  • Knowledge gaps, grouped and ranked, with the original questions one click away.
  • Lead and outcome tracking tied to your CRM, so conversations show up where sales and recruitment work.
  • Full conversations, not just counts. You need to read what was said.
  • Topics and sentiment across all conversations, without manual tagging.
  • Video and chat in one view, if you use video, so you see the whole journey.
  • Alerts and digests, so the important things reach you without opening a dashboard.
  • Data you can export, through native integrations, webhooks or an API.

How AgentLoop does it

AgentLoop is the analytics layer in Life Inside. It runs over every conversation with the AI Chat Agent and every interaction with Interactive Video, in one place.

  • What visitors asked, grouped into topics, with sentiment and the journey each visitor took.
  • Knowledge gaps with suggested fixes. The AI Chat Agent answers only from content you approve. When it does not know, it says so, and the question shows up in AgentLoop.
  • Drop-offs across conversations and video, including click-through, branch choices and heatmaps for interactive video.
  • Lead scoring and routing, with leads and conversation data pushed to HubSpot and Salesforce, alerts to Slack, events to GA4 and anything else through webhooks or the REST API.
  • A real-time dashboard, weekly insight digests and conversation recordings, so you can see the trend and read the conversations behind it.

Start measuring what visitors actually ask

If your current chatbot only tells you how many conversations it had, you are missing the part that improves it. Book a demo to see AgentLoop on real conversations, read more on the AgentLoop page, or see how the HubSpot integration sends conversation data to your CRM. Comparing tools? Our chatbot alternatives page can help.

Frequently asked questions

Chatbot analytics is the measurement of how a chatbot or AI agent performs: what visitors ask, how often it answers and resolves their need, where it fails, where visitors drop off and how many conversations turn into leads. Good chatbot analytics tells you what to fix, not only how busy the bot was.

For a website chatbot, start with answer rate, resolution rate, unanswered questions, conversation-to-lead rate and drop-off. Add containment and handover by reason, topics, sentiment and CSAT once those are in place. Conversation count on its own says little.

Pick a small set of metrics with fixed definitions, track their trend weekly, and act on them. Each week, review answer rate, resolution rate and conversation-to-lead rate, fix the top three knowledge gaps, read a sample of negative and successful conversations, and check where visitors drop off.

There is no reliable universal benchmark, because tools define resolution differently and divide by different totals. Some count implied resolutions when a visitor simply stops replying. Compare your own rate over time using the same definition, and read a sample of conversations to check it.

Containment counts conversations the bot handled without a human. Resolution counts conversations where the visitor got what they needed. A conversation can be contained but unresolved, when the visitor gives up, and a handover to the right person can be the best possible outcome on a sales or careers page.

Use an AI agent that says when it does not know instead of guessing, then review the unanswered questions regularly. Group them by topic, rank them by volume and value, and fix the source content on your website or in your documents rather than patching the bot.

Yes, if the tool supports it. Life Inside AgentLoop pushes leads and conversation data to HubSpot and Salesforce, sends alerts to Slack and events to GA4, and connects to other systems through webhooks or a REST API.

It can be, but it depends on the vendor. Check where data is hosted, whether there is a data processing agreement, how data is encrypted and whether visitors are told they are talking to AI. Life Inside hosts data in the EU, signs a data processing agreement, has a dedicated Data Protection Officer and its website widget sets no cookies.

Sources

Updated 2026-10-27