What Is Chatbot Analytics? Where the Numbers Come From, and Why Two Dashboards Disagree (2026)

Quick answer: Chatbot analytics is the layer that records what happened in a conversation and adds it up. It is not one system. Your platform's runtime logs events it can see, the channel logs deliveries it handled, and your web analytics logs a widget that opened on a page, and all three are counting the same customer with different rules. So a single month can surface as, say, 2,400 conversations in the platform, 3,100 messages at the channel, and 1,900 widget sessions in your web analytics, with nobody in the room lying. This entry covers where each number comes from, what ships without configuration, where reporting sits on vendor price lists, and what none of it will tell you. Which metrics to actually watch is a separate question, answered in the chatbot metrics guide.

The three sources, and what each can see

Analytics disputes almost always turn out to be source disputes. A dashboard can only report what its own system observed, and the three systems observe different things.

The platform runtime is the only one of the three with any opinion about meaning. It knows a handoff fired because it fired the handoff. The channel knows a message was delivered at 14:02 and nothing about why. Your web analytics knows the visitor came from a paid ad and has no idea what they typed.

flowchart LR
 V[Visitor] --> W[Widget or channel app]
 W --> CH[Channel<br/>delivery + cost data]
 CH --> P[Platform runtime<br/>sessions, intents, handoffs]
 V -.page context.-> WA[Your web analytics<br/>referrer, device, funnel]
 P --> D[Analytics tab]
 CH --> R[Delivery reports]
 WA --> G[GA or product analytics]
 style P fill:#ffe8cc
 style D fill:#ffe8cc

Figure 1. Three collectors, three reports. The platform runtime, highlighted, is the only one that sees the conversation as a conversation, which is why it is the source most of this entry is about.

Why two dashboards disagree about the same bot

Six mechanisms account for nearly every mismatch. None of them is a bug, and all six are worth settling before anyone presents a number to anyone else.

The session window. A conversation has no natural end. The customer stops typing, and some period later the system decides that was the end. Each tool picks that period for itself, and the choice is usually documented well away from the report it distorts. Google Analytics 4 is the clearest published example: a session ends after 30 minutes of inactivity by default, the timeout is adjustable in the data stream's tag settings, and there is no limit on how long a single session can run. A thirty-minute window and a twenty-four-hour window split the same customer's morning and afternoon messages into either two conversations or one. When two tools report different conversation counts, compare this first.

Who counts as a conversation at all. A widget that opens and closes with nothing typed is a session in some systems and nothing in others, and that difference lands on every rate using conversations as a denominator, including deflection and completion. We have not measured the share of bounced opens ourselves and would not trust a single figure across industries anyway.

Estimated counts versus exact ones. Even one vendor's own two surfaces can disagree, and Google documents this openly. Since October 2021 GA4 has estimated distinct session counts using the HyperLogLog++ algorithm in reports, Explorations, and Looker Studio, while a BigQuery export computes them precisely. Google's published example shows 1,463 sessions in reports against 1,501 in BigQuery for the same week. The directional change is identical; the absolute numbers are not. Expect the same category of gap between any dashboard and any raw export, and use the export when the absolute number is what matters.

Bot turns versus human turns. In a platform with a shared inbox, one thread can contain both bot messages and agent messages. Whether that thread counts as bot-handled, human-handled, or both determines your containment rate more than any tuning you do. Ask the vendor how a mixed thread is classified before you set a target on it.

Repeat contacts. The same person asking the same question three times in a week is three conversations to a per-session counter and one unresolved issue to a customer. Deduplication by contact is rare in native dashboards and it is the specific reason a bot can post a rising resolution number while the same people keep coming back. Our containment rate calculator prints the rate under four competing definitions precisely because this choice changes the answer.

Attribution reach. Some tools credit an outcome only if it happened inside the conversation. Others follow the contact forward, and how far forward is a setting. GA4's session attribution runs on a non-direct last-click model with a key-event lookback window that defaults to 90 days, which is why a session can be credited to a channel the visitor last touched two months ago. Chatbot platforms differ from that and from each other: Landbot's vendor documentation confirms last-touch only, with no multi-touch attribution, a limit we noted rather than tested. Ask any vendor whether an outcome is credited only when it lands inside the conversation or whether the contact is followed forward, because a bot that looks unprofitable under one model can look essential under the other on identical traffic.

What ships native, and what you have to build

Reporting depth varies more between platforms than almost any other feature, and it is rarely what buyers evaluate on. The table is drawn from our own review work, and the basis column separates what we watched from what we read, because the difference matters when you are planning around it.

Chatbotscape earns affiliate commission on some of the platforms named below. It does not change what we publish. See our affiliate disclosure.

Two things in that table are worth more than they look. Botpress surfacing per-call model cost as a first-class widget is the fastest route to an honest chatbot ROI number, because it removes the guesswork from the variable side of the bill, and most competitors hide it. SendPulse putting a workspace-level aggregate above its per-bot view matters enormously to anyone running several bots and not at all to anyone running one. Neither capability appears on a feature comparison grid.

Note also what the Tidio row does and does not say. The Lyro dashboards report on the AI agent's conversations specifically. Broader deflection and satisfaction measurement across all your traffic is still something you instrument, which is the point our metrics guide makes about conversation-marketing platforms generally.

Where reporting sits on the price list

Analytics is a common paywall feature, and the gate usually sits between seeing a number and slicing it.

Botpress is the instructive case here, and the row above would mislead without one more sentence. We captured that full 11-widget analytics dashboard on a free-tier workspace, so whatever the Team line adds, it is not the dashboard: read it as governance around analytics rather than access to it. That is a genuinely good deal for a small team and the opposite of what a tier table implies at a glance.

SendPulse and Manychat are absent from that table because we did not verify a reporting-specific tier gate for either. Read the rest against the plan you will actually run. All prices here are monthly-billed rates rather than the annual-commitment numbers vendors usually display first, which is the convention explained in our methodology. A reporting requirement discovered in month three is a tier jump, not a setting, and on Chatbase the specific jump from adequate to sliceable is $150 a month to $500.

Intercom is priced on a different axis and does not belong in that table. Its tiers are per seat, at $29, $85, and $132 per seat per month as verified on the vendor pricing page on 26 May 2026, with Fin billed separately at $0.99 per resolved outcome. Reporting depth there is a function of which seats you buy rather than a single analytics gate.

Export and retention, the part nobody checks first

Two questions decide whether your analytics survive contact with a real business, and neither appears in a demo.

Can you get the raw rows out? A chart is a conclusion someone else reached. Export is what lets you ask a question the vendor did not anticipate, and it is also the only way to get exact counts rather than estimated ones when the absolute number matters. In our review set the control is common but not universal, and it usually sits on the conversations table rather than in the analytics view, which is why buyers miss it. Landbot puts a Raw Data Export button directly on that table; our Chatbase scenario run recorded CSV export working; Manychat and Tidio are both reported as shipping it, though in those two cases we saw the capability described rather than exercised end to end.

How long is the data kept? Retention windows are set by the vendor and are frequently shorter than the period you will want to compare against. Custom retention is an enterprise-tier concept: on Botpress it appears only at Enterprise, alongside data residency. If your first honest year-over-year comparison is twelve months away, export monthly from day one. A rolling window that quietly drops your launch period cannot be recovered later at any price.

There is a third question specific to regulated work. Analytics data is conversation data, which means it can contain whatever a customer typed, including things they should not have typed. Our security and PII guide covers scoping that before it becomes an export problem.

What analytics will not tell you

The reporting layer counts events. It is very good at that and structurally unable to do the other thing you need.

It will tell you the fallback rate climbed to, say, 18% and not what people were asking when it did. It will tell you completion dropped on one flow and not that the third question is ambiguous. It will hand you a CSAT score and not the sentence that explains it. Every one of those answers is in the transcripts, and reading transcripts is a different discipline with a different cadence, which is the subject of our conversational analytics guide.

The practical split: analytics tells you where to look, transcripts tell you what to change. Teams that only do the first tune numbers. Teams that only do the second fix vivid, rare problems while common ones persist. The metrics guide sets the watch list, and the KPI translation guide covers what to do with it once someone senior asks.

How to read a vendor's analytics claim

Four questions, asked while looking at the screen rather than the feature page.

• Which tier does this view turn on? Not whether analytics exists. Which plan, at what monthly-billed price, shows the slice you just saw.
• How do you define a conversation, and what is the session window? A vendor who answers immediately has thought about it. A vendor who has to check is fine. A vendor who says it depends should be asked to show you the setting.
• Can I export the raw rows, and from where? Ask them to click it. The answer is a button or it is a roadmap.
• How long is history kept on this plan? Then ask what happens to it if you downgrade.

One caution about using our own rankings as a shortcut here. Analytics and reporting is one of the seventeen dimensions in our rubric and it is deliberately weighted low. It carries 4% in the per-dimension scoring table published inside each review, against AI and NLU at 15% and pricing at 12% on our methodology page. If reporting is genuinely your deciding factor, read the per-dimension analytics score inside each review rather than the composite position on the ranked best AI chatbot platforms list, because a 4% weight cannot move a ranking far.

Related terms

• Chatbot deflection rate — the headline effectiveness number, and the one most sensitive to how a conversation is counted.
• Chatbot containment rate — what a mixed bot-and-human thread does to your self-service figure.
• Deflection vs containment — why the two most-quoted numbers on any dashboard are not the same number.
• Chatbot CSAT — the satisfaction signal that keeps the volume metrics honest.
• Chatbot fallback rate — the earliest warning any analytics surface gives you.
• Chatbot ROI — where every number in the reporting layer is eventually headed.

FAQ

What is chatbot analytics in simple terms?

It is the part of your chatbot platform that keeps score. Every time a conversation starts, a flow step runs, an intent matches, or a handoff fires, the runtime writes it down, and the analytics view adds those records up into charts and rates. The important thing to hold on to is that it can only count what it saw. Anything that happened on the channel before the message arrived, or on your website before the widget opened, belongs to a different system with a different total.

Why does my chatbot dashboard not match Google Analytics?

Because they are counting different objects with different rules. Google Analytics ends a session after 30 minutes of inactivity by default and ties it to a visitor identity managed with cookies; your platform uses its own window and its own contact record, which usually survives across devices in a way a browser cookie does not. Google also estimates distinct session counts in its reporting surfaces rather than computing them exactly, and publishes an example where its own reports and its own BigQuery export differ for the same week. Add ad blockers, which suppress web analytics but not server-side platform logging, and a meaningful gap is unremarkable. Pick one system of record per question instead of trying to reconcile them.

Which chatbot analytics metrics actually matter?

A short list, reviewed weekly: completion rate and fallback rate at the engagement layer, deflection paired with a satisfaction signal at the effectiveness layer, and cost saved plus ROI at the outcome layer. Six numbers you act on beat forty you scroll past. The reasoning, and the traps in each, are in the chatbot metrics guide.

Is there a free chatbot analytics tool?

The realistic free path is the free tier of a platform that reports natively, plus your existing web analytics for the page-side context. Two platforms in our review set stood out on a free workspace: SendPulse surfaced both a workspace-level view and a per-bot statistics tab with a date-range picker, and Botpress surfaced its full 11-widget dashboard including per-call model cost. Dedicated third-party chatbot analytics products exist, and the ones we have looked at are built and priced for contact-center volumes; below a few thousand conversations a month they answer questions you can answer with an export and a spreadsheet.

How do I measure chatbot ROI from analytics data?

Two inputs and a subtraction. Take deflected conversations from the analytics layer, multiply by your loaded cost per ticket to get savings, then subtract the true annual cost, which is subscription plus model usage plus the operator hours nobody books. The variable middle term is where estimates go wrong, which is why a per-call model cost widget like the one Botpress surfaces is worth more than it looks. The ROI quick math has the five-minute version.

Can I export chatbot analytics data to my own BI tool?

Usually yes on paid plans, and the route is often a raw-data export from the conversations table rather than an API. Intercom documents export for BI integration, and Landbot puts a Raw Data Export control directly on its conversations view. Ask two things before you build a pipeline on it: whether the export includes the fields you need or only the visible columns, and what the retention window is on your tier, since a pipeline cannot backfill data the vendor has already dropped.

Sources

• Google. About Analytics sessions — 30-minute default inactivity timeout, no limit on session length, non-direct last-click session attribution with a 90-day default key-event lookback, and HyperLogLog++ estimation of session counts in reports versus exact counts in BigQuery, including the published 1,463-versus-1,501 example. support.google.com/analytics/answer/9191807 (read 4 August 2026).
• Google. Configure your Google tag settings — where the session timeout and the engaged-session threshold are adjusted on a web data stream. support.google.com/analytics/answer/12131703 (read 4 August 2026).
• Chatbotscape platform reviews, with the basis for each claim given inline in the tables above: Botpress (11-widget dashboard observed on a free-tier workspace 29 May 2026; Team-tier and Enterprise retention rows from the vendor pricing page re-verified 9 June 2026), Landbot (Conversations Data table and Raw Data Export control observed in a Pro-trial session 29 May 2026; last-touch attribution and funnel gating vendor-documented, not verified), SendPulse (workspace-level and per-bot analytics observed on a free-tier workspace 27 May 2026), Chatbase (dashboard contents from Scenario F on a Standard-tier account 24-26 May 2026; tier prices from the vendor pricing page 26 May 2026), Tidio (Analytics Overview observed on a business-trial workspace 29 May 2026, escalation-rate and fallback-trigger metrics vendor-documented; tier pricing from the vendor pricing page 26 May 2026), Intercom (Reports dashboard observed on a trial workspace 28 May 2026, remaining capabilities from vendor documentation 26 May 2026 with paid-account testing outstanding; per-seat and Fin outcome pricing from the vendor pricing page 26 May 2026), Manychat (analytics line from the methodology-anchored scenario assessment of 24-25 May 2026).
• Chatbotscape. Chatbot metrics that matter — the SMB KPI guide. /academy/chatbot-metrics-guide — the metric-selection layer this entry deliberately does not repeat.
• Chatbotscape. Containment rate calculator — the same period of sessions scored under four competing definitions. /tools/containment-rate-calculator.
• Chatbotscape evaluation methodology, including the monthly-billed pricing convention used above. /methodology publishes the cluster-level rubric weights and the 15% and 12% figures; the 4% analytics weight is published in the per-dimension scoring table of each Tier 1 review (continuously updated).