Website Analytics
Conversation volume, top questions, knowledge gaps, and lead sources.
Analytics show you how your agent is actually performing — what visitors ask, where it helps, where it falls short, how many leads it produces, and what it's worth to your business. You'll find it in Dashboard → Analytics.
By default this screen is scoped to one website at a time and to a date range you pick at the top. Switch the website with the workspace switcher at the top of the left-hand navigation, under the logo; switch the period with the Period dropdown. Looking at one site at a time keeps a spike on one site from masking a problem on another, and you can compare this week to last, or this month to last, just by changing the period.
Period offers Last 7 days, Last 30 days, Last 90 days, Last 180 days, and Custom range — the last of which reveals From and To date fields. A custom range can reach back no further than the analytics retention period published in the Data Retention Policy, because session data older than that is deleted rather than kept.
Your time zone. Days and hours on this screen — where each day of the period starts and ends, the daily and hourly charts, the time a conversation started — are in the time zone saved on your account (Dashboard → Account → Time zone), and each chart names the zone under it. Last 7 days means today and the six days before it on that clock, and the figures on the tabs, the visitor-engagement card included, cover that one window. Two people in different zones looking at the same website can therefore see a late-evening conversation on different days. The one exception is the Plan Usage block, which counts your plan's allowance period rather than the one you picked and gives its billing dates in UTC — see Plan Usage.
Scope sits next to it, with an All websites switch. Turn it on and the per-site tabs are replaced by an account-wide view: your totals across your websites, plus a row per site ordered busiest first (by sessions). It's the quickest way to see which of your sites is carrying the traffic. Turn it off to go back to the single-site tabs described below.
The dashboard is organised into tabs across the top. Which tabs you see depends on your plan — the ones your current plan doesn't include show a small padlock and open an upgrade panel (with a Learn more link to this page) instead of the data.
Which tabs are on which plan
| Tab | Starter | Pro | Business |
|---|---|---|---|
| Overview | ✓ | ✓ | ✓ |
| Messages | — | — | ✓ |
| Geographic | — | ✓ | ✓ |
| ROI | — | — | ✓ |
| Export | — | ✓ | ✓ |
| Document Analysis | — | ✓ | ✓ |
| Conversations | — | ✓ | ✓ |
The table is generated from the plan catalogue our servers read. A locked tab still shows you what's inside it and which plan unlocks it.
Overview
The default tab — a one-glance snapshot of usage, business impact, and engagement for the selected period.
Plan usage
This block on Overview shows your plan's allowance: how many AI replies your agent has sent in the current period against your plan's monthly allowance, then one card for each other feature your plan meters. One AI reply from your agent counts as one message — see Billing & plans for what counts.
The period. The line under the heading names the period by its dates, for example Period Sep 1, 2026 – Oct 1, 2026 · resets in 10 days. The second date is the moment the count starts again. These are billing dates, so they follow your subscription's calendar in UTC rather than your own clock, and the caption under the line says so; near midnight they can therefore read a day apart from your local date. This countdown runs to the allowance reset, not to your next payment: the website card in Workspaces counts the days to your subscription's renewal, so on a yearly plan the two numbers differ, except in the last month before renewal, when both count to the same day.
How the period is set. Periods run from your billing date to the same day of the next month. The billing date is normally the day you subscribed, or the day your free trial ended. If it falls on the 29th, 30th or 31st, a shorter month changes over on its last day and the next month returns to the original day: a subscription that started on January 31 resets on February 28 (29 in a leap year), then on March 31 and April 30. On a yearly plan the allowance is still monthly: it renews every month on that day, while you are billed once a year. During a free trial the allowance renews monthly too, counted back from the day the trial ends. A 7-day trial is a single period, from the start of the trial to its end. A trial of 29 days or more can hold two periods, and the first can be as short as one day — a 29-day trial started on January 31 runs January 31 – February 1, then February 1 – March 1.
What the count covers. The replies your agent sent since the period shown began. If support restarts your count, the period starts at that moment, and the first date moves to it. The Period filter at the top of the page does not change this block: it shows the allowance period, not the period you picked.
The chip grades the count — Healthy Usage → Approaching Limit → Near Limit → Limit Exceeded. A plan with no cap shows the count with an Unlimited pill instead. If the usage could not be read at that moment, the block says so in one line instead of showing figures; reload the page shortly.
When a period has ended. A few minutes after the reset date the new period opens on its own, and its count starts from zero; a visitor who reached the limit can write again without reloading the page. Until then, the block keeps showing the period that ended, with its final numbers, marked Period ended and without a countdown. A workspace that shows Unpaid or Canceled has no Analytics to show: the page asks you to complete payment instead.
The block updates by itself while the page is open — after a reply, a reset or a plan change — and the countdown moves on at midnight UTC.
When the widget has dropped into contact-only mode, a banner says so in this block: the widget stays on your site and, when Show native lead form to visitors is ticked (as it is on the default form), collects contacts until the quota has room again. The banner reflects the server's own view of the widget's state rather than being read off the chip, so trust it as the answer to "is my widget still answering". See What happens when you run out.
Below the count, one card per other feature your plan meters — leads, bookings, Text-to-Speech characters, document uploads, and so on — counted over the same period. Each card is in one of six states:
- Active — a used / limit bar with the percentage consumed.
- Unlimited — the feature has no cap on your plan; only the used count is shown.
- No stated ceiling — your plan names no allowance for this row. Deliberately not shown as unlimited, because it isn't a promise of one.
- Needs setup — the feature is in your plan but the integration isn't connected yet, with a shortcut to Integrations to finish it.
- Coming soon — catalogued but not launched yet, shown greyed out.
- Locked — the feature isn't in your current plan, with the plan that unlocks it.
New features you gain on an upgrade appear here automatically. There is no card for voice input: speech-to-text runs in your visitor's browser and is not metered.
Who sees it. Only the account owner. Teammates with access to Analytics see the rest of the Overview, including the chart below, but not your plan's count — see Team & Operators.
Visitor messages per day
A bar chart of the messages your visitors wrote each day, over the period you picked at the top — useful for spotting a traffic spike. It counts what visitors wrote, not the replies counted against your allowance, and it follows the Period filter, so its total rarely matches the plan usage above. Its days are in your own time zone, named under the chart.
Your business impact
Four headline outcomes, measured over the period you picked:
- Customers Helped — unique visitors who talked to your agent.
- Leads Captured — visitors who filled out a lead form during a chat (see Capturing Leads).
- Questions Answered — the messages your visitors wrote in the period you picked: the same count as the chart above. It is not your plan's count: it follows the Period filter, and a visitor message can be counted here without a reply being counted against your allowance, so this figure can be higher.
- Support Cost Saved — an estimate, worked out from that same count of answered questions at a flat per-question rate.
So switching the period changes these four figures.
Visitor engagement
How many times visitors actually opened the chat in the period, reconciled across three distinct sources:
- Visitor-initiated — the visitor clicked the chat launcher.
- Auto-show after delay — your website's auto-open setting opened the chat after its configured delay.
- Proactive — a proactive-message rule prompted the open; eligible plans can also see the trigger breakdown.
The total is the sum of those three sources; timer-based auto-show has its own row and is not counted as a proactive open. This card appears once there's at least one open. It's deliberately shown on its own — it's an engagement signal, not a conversion ratio, so it is not set against your lead count.
Engagement metrics
Four behavioural numbers:
- Bounce Rate — share of conversations that held no more than one message.
- Returning Users — visitors in the period who also had a session before it. A returning visitor is recognised only when your site declares the
audience-measurementsignal and the visitor has not declined it, so this number undercounts by design; see Visitor counting. - New Users — first-time visitors in the period.
- Avg Messages / Session — how deep a typical conversation goes.
Lead form quality
For each field on your lead form, the fill rate — how often visitors actually complete it. A low fill rate is a hint that a field may be confusing or should be optional. See Capturing Leads.
Insights
Automatic, plain-language callouts about the period — each with a short recommendation (for example, "response quality dipped on topic X — add content"). Severity is colour-coded (info / warning / critical).
Still warming up? On a site with very little traffic yet, the Overview replaces its Insights with a "Collecting your first questions" note, and the ROI tab shows a matching one, rather than publishing numbers drawn from a handful of chats. Both of those notices appear only on the 180-day window, and the period selector opens on 30 days — so if you want to know whether your figures are still settling, switch to 180 days and look. On shorter windows the numbers are shown without that caveat, however thin the traffic behind them.
Messages
(Which plans include it: see the table above)
The messages themselves — volume, quality, what people ask, and where answers land or miss.
Summary
Four cards: Total Messages (with average response time in ms), User Messages (with the matching assistant-message count), Satisfaction (average visitor rating, with chats with a reply beside it), and P95 Response (the response time 95% of replies came in under — a truer "worst-case" than the average).
Quality metrics (satisfaction, answered rate, sentiment) need about 10 messages to be reliable; below that the tab shows a "still warming up" note.
Message Volume by Hour
Your peak and quiet hours plus an hourly bar chart — useful for knowing when your audience is active and when to have a human operator online.
Message Topics
The themes visitors raise most, with a count and share for each. This is your demand map — what your audience actually cares about.
The topic names are taken from your own site, not invented by the AI: they come from the headings and page types of the content each conversation drew on. So they read like your site's own vocabulary, and a subject your site doesn't cover will not appear here however often visitors ask about it — for that, read Questions without an answer instead.
Quality Metrics & User Sentiment
Customer satisfaction (out of 5) and chats with a reply side by side, plus a sentiment split from the thumbs ratings visitors leave on individual answers.
The split is bucketed as positive (4–5★), neutral (3★) and negative (1–2★), but the widget only offers a thumb up or a thumb down — which are recorded as 5 and 1. Nothing a visitor can press lands in the middle, so the neutral row is always empty. Read the split as thumbs up against thumbs down.
Chats with a reply is a percentage, not a count: the share of conversations in the period that ran to at least two messages. Two messages means the visitor said something and got something back, which is why it stands in for "the agent replied" — but it is measuring shape, not success. It is not a measure of whether the visitor's problem got solved; for that, read How these chats ended on the Conversations tab.
Questions without an answer
The most useful list on this dashboard. These are the questions your agent found no matching content for — each one is a knowledge gap, and a page to write or a file to upload. Fix the content and the next visitor who asks can get a confident, cited answer. See "Turn gaps into content" below.
Two kinds of turn are deliberately not counted here, and knowing that is what makes the list trustworthy:
- Answers given with a caveat. If the agent found evidence and answered, hedging where it was unsure, that is an answer — not a gap.
- Booking slots, contact details, confirmations. When a visitor types "Thursday 9am" at a booking prompt or leaves a phone number, the turn is not a question and there is no content for the knowledge base to find.
So a short list here is a good sign, not a sign the report is broken — as long as the questions were kept. A site whose storage is set not to store chat text keeps no question wording, and questions older than the site's retention window have been cleared, so in either case the list is empty.
Conversation performance
- Chats with two or more messages — share of conversations that ran to at least two messages, the same measure as Chats with a reply. It does not check whether a person was needed; for that, read How these chats ended on the Conversations tab.
- Avg Handle Time — the average active time of the conversations that ran to at least two messages: the gaps between one message and the next, added up, with pauses longer than 10 minutes counted as 10 minutes. A chat left open in a browser tab and answered hours later therefore adds 10 minutes, not hours. Messages that a visitor exchanges with one of your operators in Live Chat are not part of this figure.
- First Reply Generation — how long the AI took to write its first reply, not counting the preparation before it.
Conversation depth
A breakdown of conversations by how many messages they ran to — Bounce (1 msg), Quick (2-4 msgs), Standard (5-10 msgs) and Deep (10+ msgs) — so you can see whether people get a quick answer or dig in. A conversation of ten messages counts as Standard; Deep starts at eleven.
A fifth row labelled unknown appears when some sessions have no message count recorded against them. It isn't a depth band; it's those sessions, and they're excluded from the shape of the other four.
Question lists
Four ranked lists that turn raw chats into a to-do list:
- Top Questions Asked — your most frequent questions; surface the recurring ones as a visible FAQ.
- Topics Without an Answer — topics drawn from chats where at least one question went unanswered. It's a roll-up by topic, so it complements the Questions Without an Answer card lower down the tab, which lists the individual questions.
- Questions Users Rated Low — answers visitors gave a low rating, with the average rating shown.
- Topic Satisfaction — average rating grouped by topic; the red topics are the most actionable — that's where answers aren't landing.
Geographic
(Which plans include it: see the table above)
Where your visitors are and what languages they prefer.
- Summary — counts of Countries, Regions, Languages, and Total Sessions in the period.
- Top Countries — sessions and unique users per country, ranked, with a bar for relative size.
- Languages — the preferred languages of your visitors with their share. Translation of your welcome message is automatic, so this isn't a to-do list — it's how you check that the languages you kept match the visitors you actually get, and which ones you could safely prune (see Chat configuration). If a language here is getting real traffic, it is worth opening its welcome tab once to confirm the translation actually landed.
- Engagement by Region — average active time per session by country (pauses longer than 10 minutes count as 10), so you can see where visitors engage most deeply.
Geographic and language data is derived from the visitor's connection at first contact — it appears as people start using your assistant.
ROI
(Which plans include it: see the table above)
The financial picture — an estimate of what the assistant may be worth against what you pay for it, based on your conversation volume, modelled human-support cost, and plan cost. These figures are directional planning aids, not measured revenue, audited savings, or a guarantee of future results.
- ROI Forecast — a chart projecting cumulative savings against cumulative subscription cost month over month, so you can see when the assistant pays for itself.
- ROI overview — Total Savings, Return on Investment (%), Payback Period (months to break even), and Hours Saved versus manual support.
- Value Generated — outcomes of the period: Leads Captured, Chats with a reply (conversations that ran to at least two messages: the visitor wrote and the assistant replied), Avg Handle Time, First Reply Generation.
- Cost Breakdown — estimated human support cost versus AI cost per month, and the monthly saving.
- Operational Efficiency — your subscription cost split per outcome: Cost Per Conversation, Cost Per Lead, and Cost Per Hour Saved.
Who sees the costs. Only the account owner. The ROI Forecast, the ROI overview, the Cost Breakdown and Operational Efficiency are built on your subscription price, so they stay with you. A teammate with access to Analytics sees a note that these figures are the owner's in their place, and the rest of the tab as you see it — see Team & Operators.
ROI projections need traffic to be meaningful — at least 10 questions answered in the period. Below that the tab says savings and ROI are not available for this period, and on the 180-day window it shows a "Your widget is just warming up" note with tips to drive more traffic. Even after warm-up, ROI remains an estimate.
Document Analysis
(Which plans include it: see the table above)
Analytics for the Document Analysis feature — when your visitors upload files (invoices, resumes, spreadsheets) in the chat and the agent parses them. This tab is where you watch how well that's working. It's empty until visitors start uploading documents.
Volume & Mix
Documents Analyzed, Avg Pages / Document, Avg Words / Document, an Uploads Over Time chart, and a breakdown of uploads By File Type, By Template, and By Parser Tier.
Reliability
How dependable the parsing was: Success Rate (share that produced a usable result), Needs Review (low-certainty results), and Rejections (couldn't be processed), plus Rejections by Reason, a Confidence Distribution, and Top Quality Flags — the most common reasons a document scored lower (fix the source documents or expect more manual review for those).
By Template
Metrics specific to each analysis template you run, one card per kind — Translation Quote (quotes, average total, the share that came out at your minimum quote as Min floor, currencies, top language pairs), Resume Screening (count, average score, verdicts, top missing skills), Document Summary (count, average bullets, tones), Document Checklist (count, complete %, average items found, top missing items), and Custom (count, empty-extraction rate). Print Quote, Editing Quote, Certified Translation Quote and Accounting Estimate have no card of their own; their uploads are counted under By Template in Volume & Mix.
Feedback on document answers
Thumbs up / down that visitors gave specifically to document answers (separate from the overall message sentiment on Messages), and a Feedback by confidence breakdown — the thumbs-down rate per confidence band. If low-confidence answers really do draw more thumbs-down, the confidence signal is working and you can trust it to flag documents that need a human look.
Recent Uploads
A paginated table of individual uploads — date, document, template, parser, confidence, result, feedback, status — with a Needs review only toggle so you can work through just the low-certainty ones as a review queue.
Conversations
(Which plans include it: see the table above)
The individual conversations behind the other numbers. The other tabs show conversations only in aggregate; this is the one where you open a single session and read what was actually said.
How these chats ended
The strip above the list answers the question the other tabs don't: did the visitor get what they came for? A conversation in the period in which the visitor wrote something is counted once, under the first of these that applied to it; a conversation without a visitor message is left out:
- Passed to a person — the visitor asked for one of your operators: they wrote to your team in live chat, or asked for a person in a messaging channel. It counts whether or not an operator then joined.
- Left with a question unanswered — at least one question your agent found no content for.
- No answer was sent — the visitor wrote and your assistant sent no reply.
- Contact details left — no question went unanswered, the assistant replied, and the visitor left their details.
- Answered, nothing further — no question went unanswered and the assistant replied; the visitor left no details and didn't ask for a person.
The headline figure is Handled by the assistant — the last two rows together. Read in order, the list is deliberately unflattering first: a captured lead does not hide an unanswered question, because the negative outcomes are counted before the positive ones.
What this number is not. It does not tell you whether the visitor was happy. A chat log cannot show that — someone who leaves satisfied and someone who leaves annoyed both simply stop typing. Treat it as "did the agent finish the job on its own", and use the thumbs ratings on Messages for how it felt.
Two more things the strip keeps honest. Conversations still in progress are left out until they've been quiet for half an hour — otherwise a chat happening right now would be filed as "Answered, nothing further" while the visitor is still typing. And on a quiet week, when there are too few conversations for percentages to be meaningful, the strip shows counts without them rather than a confident-looking figure built on four chats.
The breakdown describes the full period you picked; the filters below it narrow the list, not these numbers.
The session list
One row per conversation in the selected period:
- When it started and how long it was active — the same measure as Avg Handle Time.
- How many messages were exchanged, split into visitor and assistant.
- Average response time for that session.
- How many questions in that chat went without an answer — shown as a badge with the count, and absent when no question went unanswered. Plus a marker when a lead was captured during the conversation.
The count is worth chasing rather than dreading. It's a fact, not a verdict: a twenty-turn conversation where one question missed is not the same thing as a conversation that failed, and the row says so. These are the same gaps you see counted on Messages, except here you can read the conversation around one and see what the visitor was really after. The lead marker points at a conversation that turned into business — the pattern you want more of.
Two switches above the list let you narrow it to chats with unanswered questions or chats that captured a lead.
Reading one conversation
Open a session to see it turn by turn, in order, with each reply's grounding status, confidence, and response time. Very long conversations are shortened for display — when that happens the transcript says so and tells you the true message count.
How much text you'll see
This is the part to understand before you judge the tab. The wording of a conversation is only there if your website was set to store it, so a transcript may legitimately show a turn with no text. The tab tells you which reason applies, or that the record does not say why, because they mean completely different things:
- The site stores no chat text. Storage is currently set not to store the bot's chat text, so no wording is kept for new turns.
- The agent answered confidently. Under the setting that keeps only the turns the bot did not answer confidently — the default — a well-grounded answer stores no text. Normal, not a fault: you get the text where something went wrong.
- The text expired. It was stored, then cleared when the retention window ran out.
Older sessions disappear from the list altogether: the analytics record itself is removed at the end of its retention period, taking its text and its metrics with it. Within that, retained text is cleared earlier, after the window you set for your site. The default, and the range you may choose within, are in the Data Retention Policy, which states the periods our systems enforce; this page names no number of its own.
Conversation storage and the retention window are per-website settings and are available on every plan — see Chat configuration.
Storage follows your setting on your current plan; reading it back needs the Conversations tab (which plans have it). That asymmetry is deliberate. A site on a plan without the tab still accumulates conversations under its storage setting, so upgrading later shows you the history you already have. Text that was not stored is not there to recover later, so if you think you'll want full transcripts, turn the setting up now, not at upgrade time.
Export
(Which plans include it: see the table above)
Download your analytics as CSV to share with your team or dig into a spreadsheet. Choose one section; each download is one CSV file for the section you chose, so download twice to get both:
- Analytics Overview — the sessions, users, messages and leads for the period you picked, and your plan's message count for the current billing period.
- Leads — the leads captured in the period. You can select specific leads (or all of them) before exporting.
Leads can be exported on every plan. The plan gate on this tab is on the analytics half. If you only want your captured leads, Dashboard → Leads → Export downloads them for the period you're viewing — see Lead forms.
The overview downloads as overview-<website>-<period>.csv and the leads as leads-<website>-<period>.csv. Each file opens on its header row and is saved as UTF-8 with a byte-order mark, so Excel shows accented and non-Latin text as typed. The leads file has one column per form field, in the order your form asks for them; a field that is no longer on your form gets a column after them, headed by its key.
The plan rows follow your billing period, not the Period filter. Plan Messages Used and Plan Messages Limit are the same figures as Plan usage on Overview, for the billing period that Plan Period Start (UTC) and Plan Period Resets (UTC) name in the same file. Plan Messages Limit reads Unlimited when your plan has no monthly limit and Not stated when no limit is on record, and Plan Period Status reads Ended once that period has passed and the count has not started again. A count for the dates you picked (for example the last 90 days) does not exist: the allowance is counted per billing period only. Total Messages is a different count: it adds the visitors' and the assistant's messages of the period you picked, so it is larger than Plan Messages Used, which counts one for each visitor question your agent takes. The file is in English, like the leads file.
When the file is ready, a notice names the section and the period and offers Save file. Your browser may block repeated automatic downloads; Save file is your own click, and if it is blocked too, allow downloads for this site in your browser settings.
Turn gaps into content
The Questions Without an Answer and Topics Without an Answer cards (on Messages) are a to-do list for your agent. When you see a recurring question it found no content for, add the answer to your site — we'll pick it up on the next crawl if the page is within your plan's page and depth limits — or upload a document in Knowledge Base (see Crawler & content). The next visitor who asks can get a confident, cited answer.
Live-chat analytics are separate
Conversations your human operators handle in Dashboard → Live Chat have their own analytics — conversation volume, response times, a per-operator breakdown, and CSAT — under Live Chat → Analytics. They're kept separate from AI analytics and don't count against your AI message quota. See Live Agent.
Privacy
Most of this dashboard is aggregated — counts, rates, and trends. How much conversation wording is kept is your decision, per website, through the conversation-storage setting described under Conversations above: no text, only the turns your agent did not answer confidently (the default), or the full transcript. Whatever is kept is cleared automatically after your retention window, and the analytics record is removed at the end of a window of its own — which is also the ceiling your text window can be set to. Both are in the Data Retention Policy. Captured leads are cleared on their own window, published in the same place. See the privacy policy for details.