Engagement report
How many shoppers were exposed to your assistant, saw an engagement widget, opened a conversation, and actually engaged, per site, per widget, per device and per page type.
The Engagement report shows how many shoppers your assistant reaches on the site, and how many of them actually engage with it. Open Reports in the sidebar and choose Engagement, or go to /dashboard/sites/{siteId}/reports/engagement. It's the fourth item in the Reports section, after Overview, Conversations, and Sales, before Quality.
Its filter bar carries the date range only
Unlike Overview, Insights, and Quality, this report's filter bar offers no Channel, Agents, Customer feedback, Panel language, or Subject or keyword: just the date range, with no Add a filter button. It reads a separate event ledger that carries none of those columns, and neither the Playground nor the Hosted Page write any engagement events to it, so a control here would have nothing real to narrow. See the population statement below for exactly what these figures do cover.
Just below the filter bar, a population statement states in plain language what's counted: every session exposed to the embedded tag on your site, with the Playground and the Hosted Page writing no engagement events into it. It's shared verbatim with the Exposure events full table below, so the two never describe the population differently.
Exposed stays accurate behind an ad-blocker
A shopper's browser can run an ad-/tracker-blocker (Brave Shields, uBlock Origin, Firefox tracking protection) that silently drops the page-side signal the tag sends when it loads. To keep the Exposed count honest regardless, a session counts as exposed the moment it's seen under any of the funnel signals below, including opening or starting a conversation. So a blocker can occasionally under-count how many times a widget was shown, but it never affects whether a session counted as exposed, or the conversations-opened and conversations-started counts. That's an accepted limitation of the widget-views figure, not a bug.
If there's no exposure data at all in the selected range, the whole report shows a single empty state instead of zeroed tiles and an empty chart.
The engagement rate now means "engaged", not "opened the panel"
Engagement rate is conversations started divided by exposed, not conversations opened. Opening the chat panel only proves it rendered; starting a conversation proves a shopper actually sent a message. The two are tracked as separate tiles (Conversations opened and Conversations started), and the headline rate is built on the stricter one.
This is a redefinition of a number that used to mean opened / exposed. It changed while the platform has no live merchants, so there's no historical rate being silently reinterpreted for a real account. But if you're comparing today's rate against a screenshot or an export from before this change, it's not the same metric.
Started isn't tracked before this change shipped
Conversations started and Engagement rate show a dash (—) instead of a number for any selected range that ends before this capability started recording. A range straddling that date (like "Last 30 days" right after ship) still shows a real, if partial, figure; only a range that falls entirely before it shows the dash. This is intentionally not rendered as a zero, which would read as "shipped, but nobody engaged."
What it shows
Everything in this section is live, scoped to the current site and the selected date range.
KPIs
- Exposed: sessions where the tag was active on a page. Counted from any of the funnel signals below (see the callout above), so it never undercounts relative to the later stages.
- Widget views: sessions where an engagement widget was visible, across every format: chat bubble, question bar, button, or Starters. Can undercount when a blocker interferes, or when a shopper opens a conversation through another surface (a merchant-added trigger, say) that shows no widget at all.
- Conversations opened: sessions where the chat panel rendered.
- Conversations started: sessions where the shopper sent a first message. This is the tile the engagement rate is built on.
- Engagement rate: conversations started divided by exposed. Shown as a percentage; 0% when there's exposure but nothing started, a dash when the whole range predates tracking (see above).
All five tiles are unambiguously "higher is good". Unlike the Quality report's negative-ratings tiles, there's no metric here where a rise is bad news. When comparison is on, every tile gets a delta badge: a relative change (e.g. "+12%") for the count tiles, and a change in points (e.g. "+3 pts") for Engagement rate. The badge is omitted for a tile whose previous period had zero, or whenever the current tile itself shows a dash.
Engagement over time
A chart with one point per day for the selected range, plotting Exposed, Widget views, Opened, and Started as four separate lines rather than stacking them: they're successive funnel stages, not parts that add up to a whole, so overlaying them side by side is the right read. A day with no events of any kind still appears on the chart at zero rather than being skipped. Like the other reports' charts, this always reflects the current period only. The comparison toggle affects the KPI deltas above, not this chart.
Widgets: per-widget start rate
A table listing every engagement widget on the site, sorted by start rate (best first): its name, format, views in the selected range, and its own start rate (started ÷ views for that widget alone); a dash when it had no views, or Not configured in the format column for a widget with no version at all. Each row links to that widget's own Performance tab for the deeper per-widget funnel. A widget's figures roll up every version that was live during the range, not just its current one, so republishing mid-range doesn't split its history across two rows.
By device
A small table (one row per device: Desktop, Mobile, Tablet, or Unknown, whichever recorded any exposure) showing Exposed, Started, and the rate between them for that device. It's a table rather than a chart because three figures per device don't fit one chart axis cleanly, and the rate is the one that actually answers "does device matter here."
Conversations opened by page type
A bar showing how opens split across the page types you declare on the page (Home, Category, Product, Checkout), plus a Not declared bucket for opens with no page type attached: an older tag bundle, or a page you haven't classified.
A distribution, not a rate
This is a count of opens, not a per-page-type conversion rate. There's no per-page-type exposed or started figure yet, so there's nothing to divide by: a page type with a tall bar had more opens, not necessarily a better rate. Read it as "where opens are coming from," not "which page type converts best."
Widgets that failed to mount
Appears only when at least one widget recorded a mount failure in range: an inline button or Starters widget whose configured mount point wasn't found on the page. This is a configuration signal, not a performance one: a widget with zero views and at least one mount failure reads as "Never appeared", distinct from a widget that mounted fine and simply had nobody click it. A widget with some views and some mount failures elsewhere reads as "Mounts on some pages and not others": partially broken, not fully. Each row links to the widget's own page. See Place an inline button on your pages to fix a mismatched selector.
By starter question
Below the chart, a What opens a conversation section carries a table: one row per starter question that was shown at least once in the selected range, its click-through rate on each surface it rendered on (welcome screen, question bar, chat bubble, and the inline Starters card, labelled Starter questions) and combined across all of them. Cells are shaded from pale to deep green by rate, so the questions worth a second look stand out without reading every number. Each cell also carries the raw impressions and clicks behind its rate, and a surface a question never rendered on reads Not shown here, never a 0% rate.
Rows are ordered most-shown first, not by rate: a question with a lot of traffic and an average rate outranks one shown a handful of times at a high rate, because the first is where a point of improvement is worth real clicks. Each row also carries the same performance verdict the questions grid shows.
A question shown to nobody in the range doesn't get a row of dashes. It's left out, and the section states the count instead ("3 questions in the library were shown to nobody in this period"). This section has no date range or filter of its own: it always follows whatever range and comparison you've set at the top of the page, unlike the questions grid and the question's own page, which are both fixed to the last 14 days.
Only shown when the site's library has at least one question. On a site with no starter questions, the section doesn't appear.
Exposure events
Below the questions section (or right after the funnel, on a site with no starter questions), an Exposure events list shows the raw funnel events the period's Exposed/Widget views/Conversations opened/Conversations started figures are computed from, one row per event, not aggregated into any rate. A See all link opens the full table at /dashboard/sites/{siteId}/reports/engagement/exposures, carrying the same date range. Its own filter bar carries the date range alone, same as the report above it (see the callout at the top of this page). There's no content filter, search box, or sorting on this table; it's always newest-first, and no row is clickable: the Conversation column shows the raw id as text, not a link to the transcript. A result count above the table shows how many events match the period. Empty for the period shows an explicit message rather than falling back to older events. The same population statement described at the top of this page appears here too.
Each row shows:
- Type: one of Tag loaded, Chat bubble seen, Widget seen, Widget mount not found, Starter question shown, or Conversation opened.
- Session: the session id the event was recorded under, or a dash for none.
- Conversation: the id of the conversation the event is tied to; a dash when none exists yet (a Tag loaded row, for instance).
- Device: the shopper's device:
desktop,mobile,tablet, orunknown. - Occurred at: when the event was recorded.
This is unaffected by the KPI empty state above: the report can show "no exposure data" for its headline tiles while this list still has rows from other funnel signals, since it reads its own underlying events rather than the exposed count.
What this report doesn't cover (yet)
The report always shows a note at the bottom about what's still missing, even when the KPIs and charts above are showing real data:
- No revenue or commerce outcomes. Add-to-cart, order, and revenue figures have no writer in the events ledger yet, so they're left out entirely rather than shown as an empty chart.
Choosing a range and comparing
The Engagement report uses the same header as the other reports: the same date range picker and compare-to-previous-period checkbox.
Sales report
Attributed revenue and the commerce funnel from conversation to purchase, for the selected period. Coming soon.
Quality report
Answer quality from the thumbs-up/thumbs-down ratings on the assistant's replies (satisfaction, negatives by category, satisfaction by agent, and the most recent complaints), for a date range you choose.