Conversations

A conversation is one chat thread. Every conversation is pinned to the exact agent version that produced it and carries a conversation source telling you where it came from (see Core concepts).

This is the full, filterable conversation list. It has three tabs: Log, Conversation topics and Key figures. The AI judge's review queue moved to the Quality report (see Needs review). Reach it from the Conversations item in the sidebar's Reports section (it opens this grid directly), from a row of the Overview report's Conversation types card (it opens this grid filtered on that type, with the same filter bar carried along), or go straight to /dashboard/sites/{siteId}/conversations. For the aggregate view across all of a site's conversations (volume, tool usage, per-agent breakdowns), see the Activity report instead.

Conversation sourcesCopy Anchor Link

Review conversationsCopy Anchor Link

The Conversations table lists this site's conversations, freshest first by default. Each row shows its title, the agent it ran through, its source, its Resolution (see below), its Panel language, its Device (Desktop, Mobile, Tablet or Unknown, see The Device filter), its Duration (the time from its first message to its last), the message count, a Feedback count, and when it was created. The title is derived from the first user message, and trimmed to 40 characters with an ellipsis if it's longer (hover it to see the full title); a conversation with no first message yet shows as "New conversation".

The Feedback column shows how many of the conversation's messages were rated, as separate badges: a red thumbs-down badge with the negative count, a green thumbs-up badge with the positive count, or both together if the conversation has messages rated each way (a conversation's messages are rated independently; see Rate a message). A conversation with no rated messages shows a dash.

Above the table, this list uses the same filter bar as every report in the section: the date range (defaulting to the last 30 days, see the callout above) plus all eleven filters:

• Source (a checklist, not a single-select) narrows the list to any combination of Web, AI Storefront, and Playground; at least one source always stays checked. It defaults to every source except Playground. Source is the one filter shown alongside the date range by default; the other ten live behind Add a filter.
• Agents: narrows to one or more agents (behind Add a filter).
• Agent versions: narrows to one or more configuration versions of your agents (behind Add a filter). Each option is one saved state of a version and reads Agent name · V3.1 (see version labels). It combines with Agents: a conversation has to match both. Every conversation is pinned to one version, so this is the way to read only what a specific version said. See Agent versions for the details.
• Customer feedback: narrows to conversations rated Positive feedback, Negative feedback, or No feedback; check any combination (behind Add a filter).
• Panel language: narrows to one or more of the site's enabled languages, or to Unknown for a conversation with none recorded (behind Add a filter).
• Subject or keyword: replaces the old always-visible search box: add one or more terms (behind Add a filter) and a conversation must match every term, against its title, to stay in.
• Device (behind Add a filter): narrows to conversations started on Desktop, Mobile, Tablet or Unknown; check any combination. Unknown includes every AI Storefront and Playground conversation, and any Web conversation with no recorded start, so it is often the biggest bucket. iPads running Safari count as Tablet; iPad conversations started before that detection shipped (see the changelog) stay under Desktop. The Device filter has the full rule.
• Resolution (behind Add a filter): narrows to conversations the AI judge labeled Resolved, Partly resolved, Unresolved, No real need or Visitor left. Check any combination.
• Issues (behind Add a filter): narrows to conversations the judge found a given problem in. See Filter by label.
• Type (behind Add a filter, Beta): narrows to conversations the judge gave one of the nine conversation types. See Filter by label.
• Objection (behind Add a filter, Beta): narrows to conversations where the shopper raised one of the nine purchase objections. See Filter by label.

Filters combine: a conversation has to satisfy every filter you've set, and within one multi-select it needs to match at least one of the values you checked. Every filter you add, and the values you set for it, follows you to every other Analytics surface for your current browser session. See Filters for exactly what's remembered and for how long. The two yes/no switches are the exception: they're never remembered. There's no separate mobile collapse for this bar; on a narrow viewport the pills simply wrap onto a second line below the title.

Two yes/no filters, Only one shopper message and Opened from a starter, sit in the Add a filter menu as a switch row each. They're offered on the Log tab, and on the queue when it's opened from a link that still carries view=review. They don't apply to Key figures or Conversation topics, which don't list rows. Turning one on adds a chip to the filter bar, and removing the chip turns it off. Like the other filters, a switch lives in the page address, so a filtered view can be shared or bookmarked. A switch is not remembered between visits or carried to another report.

• Only one shopper message keeps the conversations where the shopper wrote a single message. Every row then shows 1 in the Shopper messages column.
• Opened from a starter keeps the conversations the shopper opened from a starter question.

Turn both switches on to combine them: a conversation has to match both.

A notice appears above the table when the Source selection leaves out Playground, a Playground conversation matches every other active filter (the date range, Agents, Agent versions, Customer feedback, Panel language, Subject or keyword, Device, Resolution, Issues, Type, Objection, and the two switches), and no Web or AI Storefront conversation matches those same filters. The notice reads: "Test conversations from Playground are hidden. Select it in the Source filter to see them." It only explains an empty list. If any Web or AI Storefront conversation matches, the table shows those rows and no notice appears. The check looks at Web and AI Storefront conversations whether or not you've checked them in the Source filter, so when the only matches sit on a source you've unchecked, the list is empty and no notice appears. The notice isn't something you dismiss: it stops showing once you add Playground to the Source filter, or once a Web or AI Storefront conversation matches the filters. The notice follows your Agent versions selection too: if you pick a draft or candidate version you've tried in the Playground (the only place those run), the list is empty and the notice tells you why.

A result count above the table shows how many conversations match the current period and filters. If nothing matches, the table is replaced by one of two messages: if the site genuinely has no conversations yet, a plain "No conversations yet" message; if your date range or filters are the reason nothing shows, a distinct "No conversations match your filters" message suggests widening the period or clearing a filter. The message itself carries no button. Use the filter bar's own Reset control (see Filters) to actually clear back to the defaults.

Click Agent, Source, Messages, Feedback, Created, or Duration in the column header to sort by that column, ascending or descending. Title, Resolution, Panel language and Device aren't sortable. Sorting by Feedback ranks by negative-count alone, not the combined positive-plus-negative total: it's meant as a triage signal for conversations that need attention, so a conversation with 5 compliments doesn't outrank one with 1 complaint. To find your most-praised conversations instead, add the Customer feedback filter, check Positive feedback, and sort by Created.

The table shows 20 conversations per page, with Previous / Next controls underneath when there's more than one page.

Your filters, sort, and current page all live in the page's URL: copy or bookmark the link and it reopens the same view.

Opening a conversation puts the transcript front and centre, with a rail of metadata beside it. The page fills the height of the screen: the transcript and the rail each scroll on their own, so a long conversation scrolls in place rather than pushing the page down, and the transcript opens on its first message. On a wide screen the rail sits to the right of the transcript; on a narrow screen it opens from a Details button in the header, in a panel that slides in from the side, with the same tabs.

• Transcript: a read-only view of the messages, always rendered with the full technical detail: every tool call and the assistant's reasoning, visible in full. This is the same detail as the playground's Technical view, regardless of which display (Technical or User view) was active while the conversation happened live, and regardless of source: an AI Storefront transcript shows the same raw detail a shopper never saw live, for auditing what the agent actually did. Any file a shopper attached to a message (see File upload) shows in the transcript too (an image as a preview, any other file as a named tile), so you see exactly what was sent. When the assistant offered quick replies, the row shows under the assistant's last sentence, where the shopper saw it, and its chips are inert: you can't tap them. A shopper message also carries an Asked from line underneath it, naming the page it was sent from and linking to it, whenever the tag read that page (see What the assistant reads from the page). Messages with no page recorded, including every AI Storefront message and any Playground message sent with no Current page, show no such line.
• Rail: the conversation's metadata beside the transcript, 384 px wide on a wide screen. It has three tabs, Details, Analysis and Result, and opens on Details each time you load the page, whichever tab you had open before. If you can delete conversations, a Delete conversation button sits at the bottom of the Details tab, under the details. The other two tabs don't show it. See Delete a conversation below.
  • Details lists the agent and the exact saved state of the version it ran through, such as V2.1 (each linking back to the agent), the Source, the Panel language (or "Unknown" for a conversation with none recorded), when it was created, the Entry widget it opened from (name, format, and version, e.g. "Homepage bubble · Chat bubble · v2", or "Not attributed" for a merchant-triggered open, an iadvize('open') call, a hosted or Playground conversation, or one recorded before this field existed), and the conversation id with a copy button.
  • Analysis (Beta) shows the labels an AI judge recorded about the conversation: its resolution and Quality score, the facts it noted (conversation type, objections, content gaps, order issue), a To review first list of the failed criteria and alert signals, and every label in Service, Trust and Signals groups that fold on their header. Every name opens a plain-words definition. When there are no labels yet it says why. A Label now icon button, for members who can edit conversations, labels this conversation straight away. See The Analysis tab.
  • Result starts with Shopper feedback: this conversation's message ratings as a thumbs-up and thumbs-down tally (the same counts as the table's Feedback column), always shown, so a conversation with nothing rated reads 0 positive and 0 negative. Next come this conversation's own impression and click counts, recorded from the product cards it rendered (the same events listed in the Event check table): 0 impressions and 0 clicks when none happened yet, or a dash if the counts couldn't be loaded. Last, Sale attribution lists the orders linked to this conversation. See Orders linked to a conversation.

Orders linked to a conversationCopy Anchor Link

The Sale attribution part of the Result tab lists every order that carries this conversation's join key, oldest first. Each order shows:

• its order reference, as your store reported it, with a status badge;
• its net amount: after the refunds recorded for the order, and 0 when it was cancelled. Refund recorded follows the amount when a refund came in;
• when it was placed, and how long after the shopper's last message before the order: in minutes under an hour, in hours under two days, in days beyond that. The delay is left out when the order's time is at or before that message: the window tolerates up to 10 minutes of clock difference between the two;
• for any status other than Credited, one line saying why the order doesn't count.

When no order carries the key, the section says No order is linked to this conversation. A dash means the orders couldn't be loaded; it doesn't mean there were none.

Some orders are never listed:

• test orders your store flagged as such;
• every order of a conversation that does not qualify for attribution, such as a Playground conversation. For such a conversation the section reads the same as one with no order.

The statuses follow the same rules as the Orders report. How an order gets the key in the first place is in How sales are attributed.

Filter by labelCopy Anchor Link

The Resolution, Issues, Type and Objection filters read the labels the AI judge recorded (see Conversation labeling). They use only a conversation's current label. A conversation with no current label (still waiting, not eligible, or only a failed attempt or an earlier, replaced label) never matches an active label filter. So turning one on also hides every conversation that isn't labeled yet. All four filter panels are marked Beta.

Issues offers eight options. A conversation matches when any checked option holds:

Type offers the nine conversation types: Policy or FAQ, Product question, Product discovery, Comparison, After purchase, Promotion, Asked for a person, Technical support and Off topic. A conversation can hold several types. It matches when it holds any checked type.

Objection offers the nine purchase objections: Price, Size or fit, Suitability, Delivery, Returns, Stock, Authenticity, Compatibility and Competitor. It matches when the shopper raised any checked objection. The Issues option Raised an objection matches any of the nine.

Within one filter, a conversation needs to match at least one checked value. Across filters, it has to match all of them: Type Comparison and Objection Price together keep the conversations that hold both.

These four filters are offered on this list's Log tab only. The Conversation topics tab and the other reports don't offer them, and the Sales report's Purchase Objection tab links into this list with an Objection already set. If a label filter is remembered from here, a report's filter bar says it doesn't apply there. Like the other filters, they are remembered for your browser session and live in the URL, as resolution, issues, types and objections.

The Resolution columnCopy Anchor Link

Resolution is a read-only column marked Beta. It shows the resolution in the conversation's current label: green for Resolved, amber for Partly resolved, red for Unresolved, plain text for the other two. A conversation with no current label shows Not labeled.

The column shows the current label even if the conversation continued after it was written, so it can describe an earlier part of the conversation. The Analysis tab of the conversation says so when that's the case.

Needs reviewCopy Anchor Link

The Needs review queue is on the Quality report, as its own page in the tab row, after Conversations. It lists the conversations the AI judge thinks are worth a person's look. The queue used to be a tab on this list. It is no longer one.

A link that still carries view=review opens the queue in this table, with the same columns and order described on the Quality report. The count line above the table counts the queue on that view, and the freshness line shows above it too.

Conversation topicsCopy Anchor Link

The Conversation topics tab ranks the nine conversation types by how many labeled conversations hold each one. It's marked Beta, it sits beside Log and Key figures, and it's in the URL as view=topic. The ranking replaces the list while the tab is open.

Exclude conversations opened from a starter is off by default. It is the one yes/no filter this tab offers, and the list's two switches don't apply here. Turn it on to leave out the conversations the shopper opened by clicking a starter question. The counts, the shares and the coverage line above the ranking then count the rest. The switch is kept in the page's URL only (excludeOpenedFromStarter=true), so it doesn't carry over to the other reports.

• Count and share. The count is the number of labeled conversations that hold the type. The share is that count divided by all labeled conversations in your period and filters. One conversation can hold several types, so the shares add up to more than 100%.
• Order. Most frequent first. On a tie, the order of the list above. A type with zero conversations stays in the list.
• Bar. Sized against the most frequent row, not against 100%.
• Each row opens the list. The row links to the Log tab with your period and filters carried over and the Type filter set to that type.

The ranking reads only a conversation's current label, like the label filters, so a conversation with no current label isn't counted. A line above the ranking says how many conversations the figures rest on and how many aren't labeled yet. Until 30 conversations are labeled in the scope, the ranking shows "Not enough labeled conversations yet" and the number so far instead. If nothing in the period is labeled, the tab shows an empty state. The tab reads at most the 10,000 newest labeled conversations in your scope.

The tab offers the Source, Agents, Agent versions, Panel language, Subject or keyword and Device filters. It doesn't offer Customer feedback, Resolution, Issues, Type or Objection, and it has no Compare to previous period option.

Key figuresCopy Anchor Link

The Key figures tab replaces the table with three figures. They cover the same date range and filters as the list, and the filter bar works as on the other tabs.

• Avg. messages / conversation: every message in the conversations, shopper and assistant together, divided by the number of conversations.
• Avg. conversation length: the average time from a conversation's first message to its last. A conversation with a single message counts as zero.
• Conversations with one shopper message: the share of the conversations in scope where the shopper wrote exactly one message. A conversation opened from a starter question counts the same as any other. The tile shows that share as a whole-number percentage of every conversation in scope.

The tile doesn't read the Only one shopper message and Opened from a starter switches: its share is always taken over every conversation in scope.

The tab offers no comparison with the previous period. A range with no conversation shows the list's empty state, not a zero average. The tab is in the URL as view=averages.

Step through conversationsCopy Anchor Link

Open a conversation by clicking a row on this table and the header gains Previous and Next arrows next to the title, with a count (e.g. "4 of 128") showing where it sits in the list you were looking at: same filters, same sort, same date range. On the Needs review tab it follows the queue's order instead. Stepping past the last row of a page carries on to the first row of the next page, so you can sweep through a whole filtered list without going back to the table in between. The back arrow in the top-left corner now returns you to that same list too (same filters, tab, sort, and page) instead of resetting to the unfiltered table.

The left and right arrow keys (← / →) do the same thing as clicking the buttons. Hover either button to see its shortcut. The shortcut doesn't fire while you're typing in a text field, while a dialog, a menu, or the mobile Details panel is open, or while you're holding a modifier key (so your browser's own back/forward shortcuts still work).

At the first conversation of the list the Previous arrow is disabled; at the last, Next is disabled. Both stay visible rather than disappearing, so the header doesn't shift as you step through. On a narrower screen the arrows still show; only the "N of total" count is hidden to save space.

A conversation opened from an older link to the Overview report's former recent-conversations list still shows the same arrows, count and shortcut. The Overview no longer has that list, so new links don't carry this. There, Previous and Next follow that list's own rows, with the Overview's date range and filters, and the back arrow returns you to the Overview, scrolled to the list. Its tooltip reads Overview instead of Conversations.

Delete a conversationCopy Anchor Link

Deleting a conversation is permanent: it removes the conversation and every one of its messages, and there's no undo. Two people can do it, each in their own place:

• You, from here in the dashboard: if your role grants the conversations:edit permission. The built-in Admin role grants it and the built-in Member role doesn't, and it's available as Edit conversations when building a custom role. The Delete conversation button sits at the bottom of the rail's Details tab on a conversation's detail page; there's no bulk-delete from the table. Clicking it asks you to confirm before anything happens. Deleting sends you back to the conversations list, since there's nothing left here to show.
• The shopper themselves, from the chat panel: a Delete this conversation action in the panel's header menu (next to "New conversation"), on both the AI Storefront and the web tag. It only works while their browser still holds the cookie that keeps their conversation going: the same one covered in Embedding the assistant's cookie section. Once that cookie is gone (a different browser, a cleared cache, or it having expired), there's no way for them (or you) to single out that one past conversation for deletion on request; it's only ever removed on its normal retention schedule.

Either path removes the same data the same way: deleting isn't a soft flag, it's gone from both surfaces once it happens.

Rate a messageCopy Anchor Link

Every assistant message (here in a transcript, in the playground, and on a site's AI Storefront) has a thumbs-up and thumbs-down control. The one exception is in the live chat, in the playground or on the AI Storefront: a message that ends on a clarifying question still waiting for an answer has no thumbs until the shopper answers, skips or closes it, or moves on by typing a message or tapping something else in the chat, because there's no reply yet to rate. A transcript here always shows them, including on a conversation that ended on a question nobody answered.

Clicking either thumb doesn't submit right away; it opens a dialog first:

• Thumbs-down opens with an optional issue category (UI bug, Didn't understand my request, Refused to help without good reason, Incorrect product information, Incomplete response, Irrelevant recommendation, Should have searched the catalog, Problem with cart or checkout, Inappropriate tone or content, or Other) and an optional free-text comment.
• Thumbs-up opens with only the free-text comment; there's no category to pick.
• Cancel closes the dialog without saving anything. Submit saves the rating on that message.

A rating belongs to the message, not to the surface it was given on: rate a message live in the playground or on the AI Storefront, and reopening that same conversation here shows the same rating, category, and comment. Clicking a thumb again on an already-rated message reopens the dialog pre-filled with what was submitted, so submitting again revises it. There's no separate way to clear a rating once given.

This summary block only appears when you're reviewing a conversation here in the dashboard. Live in the playground or on the AI Storefront, a rated message shows only the highlighted thumb: you just filled in the dialog yourself, so there's nothing more to surface in the moment. Reopen that same conversation in the dashboard afterward and the block appears there.

LabelingCopy Anchor Link

Web, AI Storefront and Playground conversations are labeled in the background by an AI judge once they've gone quiet: whether the need was met, whether the replies were accurate and on brand, what was missing. A conversation whose shopper objected to measurement is never labeled. The labels are in the Analysis tab of the conversation's rail, in Beta. On this list, the Resolution column, the label filters and the Needs review tab read them too. The REST API returns them as well and can filter conversations by label. To label one conversation without waiting for the hourly pass, use Label now in its Analysis tab. See Conversation labeling.

Deleting a conversation deletes its labels with it.

Where feedback, exposures and events wentCopy Anchor Link

This page used to have three sibling tabs: Feedback, Events, and Exposures. Each moved to the report it belongs to:

• Feedback (every rating, one row per rated message) is now the Quality report's ratings list, and its own full-table view.
• Exposures (every tag-load and conversation-open signal) is now the Engagement report's Exposures tab.
• Events (the raw impression/click debug table) moved to the Event check card of Site settings → Installation. It's developer output, not something you'd read while reviewing a report.