Conversation analytics
The dashboard your agent earns every time someone chats with it.
Sign in to your dashboard and open AI Agents → Analytics. You'll see the last 30 days of activity across all of your agents in one page — no filters, no setup, no spreadsheets.
This chapter walks you through every box on the page so you know what it means, what to do when a number looks wrong, and which numbers tell you the agent is paying for itself.
What this page is for
Three questions every owner asks once the agent is live:
- Is it actually being used? → look at the Conversations stat and the Daily volume chart.
- Is it making me money? → look at Leads captured, Conversion %, and the Lead score distribution.
- Where are visitors getting stuck? → look at Escalations, Knowledge gaps logged, and the page heatmap at the bottom.
If you only have two minutes, scan the four big numbers at the top and the green-on-green table at the bottom. Everything else is for digging deeper.
Finding the page
Sign in to your dashboard. In the left menu under AI Agents you'll find Analytics (icon: a small bar chart). The shortcut URL is:
/dashboard/{your-tenant-id}/conversation-analytics
The menu item only appears once you have at least one agent created. If you don't see it, head back to AI Agents → Agents and add an agent first.
The four big numbers (top of the page)
| Stat | What it counts | Why it matters |
|---|---|---|
| Conversations | Distinct chat sessions in the last 30 days. Each new visitor starts one; if they come back a few hours later it becomes a new conversation. | Your top-of-funnel signal. If this is zero, the embed snippet isn't live on the right pages — check the Embed install chapter. |
| Leads captured | Conversations where the visitor gave you contact details. The percentage beneath ("XX% conversion") is leads ÷ conversations. | The number that pays for the platform. A healthy site lands between 8% and 25% depending on industry. |
| Escalations | Times the agent (or a visitor) flagged the conversation for a human. Coloured amber when above zero so you spot them. | Each escalation is a fresh customer who needs you in the next few hours. See Leads & hand-offs for the workflow. |
| Knowledge gaps logged | Questions the agent couldn't confidently answer in the last 30 days. Coloured amber when above zero. | Each gap is a 30-second fix: open AI Agents → Knowledge gaps, read the question, paste an answer into your knowledge base. |
Sandbox conversations don't count. When you preview your agent from the dashboard or run a test from the agent editor, those chats are flagged as sandbox traffic and excluded from every number on this page. You're only ever looking at real visitors.
Daily volume chart
A simple bar-per-day chart of the last 30 days.
- Blue bar = total conversations that day.
- Green bar = leads captured that day (only shown when there's at least one).
Hover any bar for a tooltip with the exact date, conversation count, and lead count. Look for:
- A flat-line week — embed may have dropped off (CMS pushed a stale template).
- Big spikes on weekdays only — you have a B2B audience; consider showing different hours-of-operation copy on weekends.
- Spikes on a single day — usually an email blast or social post landed; correlate with your marketing calendar.
When there's no activity in the last 30 days, the chart is replaced with a calm "No activity in the last 30 days" message instead of an empty graph.
Top intents
Each conversation is tagged with an intent as it happens — words like pricing, support, booking, quote, comparison. The chart shows the top 8 intents and their share of all conversations.
- A long pricing bar means visitors are ready to buy — make sure your pricing page is up-to-date and that you've configured a Quote intake capability if you sell custom work.
- A long support bar on a marketing site means visitors are landing on the wrong page — check your ads.
- An unknown bar means the agent couldn't classify those conversations. A small percentage is normal; if it's over half, your knowledge base may be too generic. Open AI Agents → Knowledge and add 3–5 industry-specific FAQs.
Intents are tagged by the agent's own language model — you don't configure them. If you want to see exactly how a conversation got its intent label, open it from AI Agents → Conversations and look at the Intent badge on the transcript.
Lead score distribution
Every conversation that captures contact information gets a lead score from 0 to 100 — how likely the visitor is to convert into a customer. The chart shows how many leads landed in each of five buckets:
| Bucket | What it usually means |
|---|---|
| 0 | No contact info captured. Pure browsing or bounced. |
| 1–25 | Curious / early research. Don't chase yet — let your nurture emails do the work. |
| 26–50 | Comparing options. A short, friendly follow-up email within 24 hours converts well. |
| 51–75 | Active buyer. Call them today; they're talking to your competitor next. |
| 76–100 | Hot. They've shared budget, timeline, and contact info. Drop everything. |
Use this distribution to size your follow-up effort. If 80% of your leads land in 1–25, you need higher-intent traffic, not a faster sales process. If 30% land in 76–100, get on the phone.
When no conversations have been scored yet, this section shows "No conversations scored yet."
The page heatmap — "Which pages are working best?"
The most actionable widget on the entire page. One row per page (capped at the top 10 by conversation volume), with five columns shaded by intensity:
| Column | What it counts | Colour |
|---|---|---|
| Conversations | Distinct chat sessions started on this page. | Green — darker = more activity |
| Leads | Of those conversations, how many captured contact info. | Green — darker = more captured |
| Conversion % | Leads ÷ Conversations × 100. |
Green — darker = higher rate |
| Avg lead score | Average lead score (0–100) of conversations from this page. | Green — darker = higher intent |
| Escalations | Conversations from this page that got escalated to a human. | Amber — darker = more visitors needed help |
Read it like a chess board:
- All-green row with high conversion % → that page is your best lead magnet. Send more traffic to it.
- Dark green Conversations + pale green Leads → high interest, weak ask. The embed prompt or page copy needs a clearer call to action.
- Dark amber Escalations + lots of conversations → visitors keep needing human help on that page. Either improve the page copy or add an FAQ to your knowledge base.
The tip at the bottom of the table — "Pages with high conversions but low conversion % may benefit from a clearer call to action" — is your one-line strategy guide for this widget.
A page URL truncates if it's long; hover it for the full path.
Numbers that look wrong — what to check first
| Symptom | Likely cause | Fix |
|---|---|---|
| Conversations = 0 | Embed snippet isn't installed, or it's installed but commented out. | Open one of your live pages, view source, search for agent.js. Reinstall from AI Agents → Embed install if missing. |
| Conversations > 0 but Leads = 0 | Lead-capture capability is disabled, or your agent's prompt never asks for contact info. | Open the agent → Capabilities → confirm Lead capture is enabled and its required fields make sense. |
| Conversion % stuck under 5% | Embed is on low-intent pages (homepage, blog index). | The heatmap will show this — most low-intent pages will be at the top. Move the embed to high-intent pages (pricing, services, contact). |
| Escalations very high (>20%) | Knowledge base is missing fundamentals OR the agent's tone is too restrictive. | Open Knowledge gaps, answer the top 5 in 10 minutes. Re-check the next day. |
| Knowledge gaps climbing daily | New product or service that nobody's documented. | Add a one-page FAQ to the knowledge base. The agent will pick it up within an hour. |
Top intent is unknown |
Knowledge base is too thin for the agent to anchor visitor questions. | Add 5 industry-specific FAQs. Intent classification improves with more grounding content. |
Privacy and what's stored
Every number on this page is computed on the fly from these models, all scoped to your tenant only:
agent_conversations— the chat session record, includingintent,lead_score,current_url,has_lead,has_escalation, andis_test.knowledge_gaps— questions logged as unanswered.
The page never reads message bodies for analytics — only the flags and counts above. The actual transcripts live behind AI Agents → Conversations and stay scoped to your tenant with per-row visibility checks.
The page also never sends data anywhere external. Everything is queried directly from your database when the page loads.
What this page doesn't have (yet)
Some things you might expect that aren't here today:
- Custom date ranges. The window is hard-locked to 30 days. If you want last-90-day or last-7-day views, use the REST API (Pro) and pull the underlying counts.
- Per-agent filter. Numbers are tenant-wide. If you run multiple agents (e.g. one per site), they're combined on this page. Per-agent breakdowns live on the individual Agent → Overview page.
- CSV / PDF export. The page renders fresh on every visit; copy the numbers you need or use the REST API for programmatic pulls.
- Comparison vs. previous period. Today's numbers stand alone — no "vs last 30 days" badge. We may add this once we have enough customers with 60+ days of history.
These are tracked alongside other UX gaps in gaps.md. If one of them is blocking you, email support@coffield.io and we'll bump it.
Want to read the actual chats behind these numbers? Open the Conversations inbox. It's the row-level audit log that pairs with this dashboard.
Want to pull these numbers into your own dashboard?
If you're on the Pro plan, every metric above is available via the REST API. Look for the /api/v1/analytics/* endpoints, or query the underlying agent_conversations model directly with ?include=intent,lead_score,current_url.
If you're on the Growth plan, the MCP servers shipped with Coffield.io include a query_conversations tool that you can call from Claude Desktop, n8n, Zapier, or Make.com — same data, no code required. See External Integrations & MCP tokens for how to issue a token.