Skip to content
Documentation

Review conversations and dashboard stats

Read and filter chat transcripts, label them by topic and sentiment, see why the agent answered as it did, turn weak replies into Q&A pairs, export chat logs and follow outcomes on the dashboard.

Every saved chat with your agent appears on its Conversations tab, where you can read the full transcript. The Dashboard adds up conversations, messages and leads across all your agents, and shows what those conversations led to.

The Conversations tab

Open your agent and go to Conversations. Four tiles at the top count this agent's chats of all time: Total conversations, Active today, Active last 7 days and Avg. messages per chat.

Below them, Conversation History lists the chats, 10 per page and newest activity first. Each entry shows the visitor's first message, a preview of the last message, the number of messages and the time of the last activity. When they are known, a line below shows the visitor's country as a flag and two-letter code, and the page the chat started on. A red thumbs-down badge counts the replies in that chat that the visitor rated not helpful. A Handed off badge marks a chat the agent emailed to your team; in the transcript it sits next to the visitor's reply address, and hovering over it shows when it was sent. See Hand-off by email.

  • Search: type in Search all conversations... to find chats whose messages contain that text. The search covers every conversation, not only the page on screen, and ignores upper and lower case.
  • Date filter: All time, Today, Last 7 days or Last 30 days, based on each chat's last activity.
  • Country filter: pick a country to show only chats from there, or Unknown for chats without a country. Each entry shows how many chats it has. The menu appears once at least one chat has a country.
  • Rated not helpful: the thumbs-down button next to the filters shows only conversations with at least one reply rated not helpful. It works together with the search and the other filters.
  • Topic and Sentiment: pick a topic, or Positive, Neutral or Negative, to show only conversations with that label. They appear once topic analysis is on or a topic list exists. When the topic you picked has subtopics, a Subtopic filter appears next to it.
  • Channel: Widget, WhatsApp, Messenger, Instagram, Slack or Email, with how many chats each has. The menu appears once the agent has a conversation outside the widget, on WhatsApp, Messenger or Instagram, in Slack or by email; those rows carry a channel badge, with a pause icon while your team has paused the AI.
  • Spam: the shield button shows only conversations that spam protection paused and nobody has unpaused.
  • Low confidence: the question-mark button next to it shows only conversations with at least one reply whose confidence is below 0.5. Its tooltip reads "Only conversations with a low-confidence reply". Replies that weren't scored never match. See "Why this answer" below.
  • Read a transcript: click a conversation to see every message in order. Replies someone from your team wrote in a live chat show their first name and "· team, live chat" above them, as plain text exactly as typed; see "Live chats in a transcript" below. Files a visitor attached appear in their message: click an image to open it in a new tab, or a PDF to download it. See Visitor attachments. When a visitor booked a meeting, the confirmation appears as a message from the agent; see Booking with Cal.com or Calendly. Copy conversation puts the whole transcript on your clipboard.
  • Delete conversation removes the conversation, all its messages and the files the visitor attached. This can't be undone. A lead from that conversation stays in your leads list and shows "Conversation deleted".
  • Deleted by the visitor: with Delete my chat on, visitors can delete their own conversation from the widget. It is removed the same way, with its messages and files, and no longer appears here or in the conversation counts. Its leads, form answers and bookings stay.
  • Refresh conversations loads chats that arrived while the tab was open.
  • Export CSV downloads the conversations that match the search and filters, as described under "Export chat logs" below.

A labelled conversation shows a coloured dot for the visitor's sentiment (green positive, grey neutral, red negative; hover to read it) and up to two topic chips, in the list and in the transcript header. A topic that got a subtopic reads "Billing › Refunds". See "Topics and sentiment" below.

Live chats in a transcript

When someone from your team took over a conversation in live chat, the transcript header shows it:

  • Live chat marks a conversation a person took over earlier. While it is live, the badge says Waiting for a person or Live now instead, with an Open in Live chat link to the Live chat inbox.
  • A star badge with the visitor's rating, for example 4/5, when they rated the live chat. Hover over it to read their comment.
  • Take over, for the account owner and admins: shown when live chat is on for the agent, the visitor wrote in the last 30 minutes, and the chat isn't already waiting or live. It takes the chat over and opens it in the Live chat inbox. See Take over a recent chat from Conversations.

WhatsApp, Messenger and Instagram conversations

A conversation on WhatsApp, Messenger or Instagram shows the channel in the transcript header. Below the transcript, the owner, admins and editors can Pause AI and Resume AI, and reply to the person through the channel within 24 hours of their last message. A reply pauses the AI, and it resumes by itself 24 hours after the last pause or reply. Replies show your first name and the channel above them, for example "· team, WhatsApp". Messages the agent can't read, such as voice notes, show as a placeholder in the transcript. See Human takeover.

Improve answer and Why this answer aren't shown under a team member's replies, because the agent didn't write them.

When a visitor reopens an earlier chat with Previous chats, they continue the same conversation: their new messages are added to it, and it moves up the list by its latest activity. No new conversation is started, unless that conversation was ended for being idle. See Appearance.

The Inbox in a transcript

The transcript header shows Inbox · and the status (Open, Pending, Snoozed or Solved) for a conversation in the Inbox; click it to open the conversation there. For any other conversation, people who can edit conversations see Move to inbox, which puts it in the Inbox as open.

Spam protection

With spam protection on, a conversation the check paused shows a Spam badge in the list. In the transcript header, the badge is followed by the reason the check gave and an Unpause button.

  • Unpause lets the visitor chat again from their next message. The conversation isn't checked again, and the header then reads "Unpaused after a spam flag" with the reason, as a record.
  • Unpausing is for the owner, admins and editors. It doesn't change the conversation's place in the list.

Ended conversations

With End idle conversations on (see Limits and access), a conversation with no message for the number of days you chose is ended once a day:

  • In the list it gets an Ended badge. Its transcript header reads, for example, "Ended after 7 idle days". It keeps its place in the list, because ending it isn't activity.
  • intoCHAT writes a short summary of it: what the visitor wanted, what the agent answered and any contact details they gave.
  • Until the visitor writes again, the widget still shows that conversation to them.
  • When they write again, a new conversation starts. Its transcript header shows Continues an earlier conversation, which opens the ended one, and Summary the agent started from, which shows the summary. The agent gets that summary with the new conversation, so the visitor doesn't have to repeat themselves. The ended conversation then shows Continued later, which opens the new one.

Conversations waiting for or in a live chat aren't ended, and neither are conversations paused by spam protection. Chats through the REST API aren't ended either. The CSV export lists ended and continued conversations like any other.

Reading transcripts is the quickest way to find questions your agent answered badly. Fix them with Improve answer, described below, or add a Q&A pair yourself; see Text and Q&A. Questions to improve on the Knowledge tab collects these questions for you once a day, with repeats counted together; it is described on the same page.

Procedures in a transcript

When the agent followed a procedure in a conversation, the transcript header lists it with how far it got: "On step 2 of 6 (Ask the visitor)" while it is running, "Completed", or "Stopped" with the reason, for example "another procedure started". Click it to see the answers the visitor gave to its questions.

Where a chat started

intoCHAT saves two details once, when a conversation starts:

  • Country: the visitor's country, taken from their IP address by intoCHAT's hosting provider. Only the country is saved with the conversation. Chats where the provider sent no country, and chats from before this was added, have none.
  • Page: the page where the visitor started the chat in the chat bubble, as the address without its query or fragment, for example https://example.com/pricing. The part after ? or # never leaves your site, because it can hold personal data such as email addresses.

The page is recorded only for the chat bubble: the script tag, the WordPress plugin and the Shopify snippet. Chats in the inline iframe or through the direct link have no page, and neither do Playground chats. If the visitor moves on to other pages of your site, the conversation keeps the page where it started.

In the list, the page appears as its path, or as the domain for your home page. In a transcript, the header shows the country's name and the domain with the path. Click the address to open the page in a new tab.

Why this answer

Under each of the agent's replies in a transcript, Why this answer shows what the reply was built from. A badge next to it gives the reply's confidence: High, Medium or Low. Click it to open the details:

  • Confidence: the level and the number, for example "High (0.82)".
  • Knowledge used: the knowledge excerpts the agent was given for this reply, grouped by source. Each source shows its name, its type (Website, File, Text, Q&A, Notion or Google Drive), "priority" if you marked it as priority, its best match from 0 to 1, and the number of excerpts when there is more than one. A website page's name is a link that opens the page in a new tab. "keyword match" instead of a number means the excerpt was found by keyword search alone. When the agent searched and nothing matched, it says "No matching knowledge was found for this question."
  • Actions called: only when the reply used one. Each API action, client-side action (listed as "Website action"), custom button, chat form (listed as "Form"), the lead form, a hand-off by email, a web search and the open times of Booking with Cal.com or Calendly (listed as "Booking:" followed by the event type's title) is listed with Succeeded, Failed or Outcome unknown. Each step of a procedure is listed as "Procedure" with the procedure's name and what happened, such as "Started", "Step 2 done", "Confirmation asked" or "Completed". A web search also shows its query in quotes and links to the pages it cited, up to 5. Visitors see up to 3 of them under the reply, on a line starting with "From the web".

Replies saved before intoCHAT started recording these details have no Why this answer. Saved Playground chats have it too.

What confidence means

Confidence is a number from 0 to 1 that says how closely your knowledge matched the visitor's question. It doesn't check whether the reply is correct. An outdated page can match well and lead to a wrong reply, and the agent can still write a good reply from a weaker match.

  • The best-matching excerpt decides most of it. A second matching excerpt raises it a little and never lowers it.
  • A Q&A pair counts a little higher than another source with the same match.
  • An excerpt found by keyword search that contains every meaningful word of the question, or every product code in it, counts as at least 0.6.
  • When the agent searched and found nothing, confidence is 0.
LevelConfidence
High0.75 to 1
Medium0.50 to 0.74
Low0 to 0.49

Some replies are Not scored: they have no badge and no number, because they didn't rely on your knowledge. This happens when:

  • the agent has nothing trained yet, or the message had nothing to search for;
  • the visitor only said hello, thanks or goodbye, in one of the widget's five languages;
  • nothing matched, but the reply used an action, such as an API action, a button, the lead form, a hand-off, a web search or booking.

Replies that aren't scored never count as low confidence, neither in the filter nor in Questions to improve.

Improve an answer

Under each of the agent's replies in a transcript, Improve answer turns that reply into a Q&A pair.

  1. Open a conversation and click Improve answer under the reply.
  2. Question is filled with the visitor's last message before the reply, and Answer with the reply's text, without any lead form, buttons or product cards. If the visitor rated the reply not helpful, the dialog says so.
  3. Edit the question into the one you want to answer, up to 200 characters, and correct the answer.
  4. Click Save and train.

If a Q&A pair with the same question already exists, ignoring upper and lower case, the dialog warns you, and saving replaces that pair's answer and keeps its other phrasings. Otherwise a new pair is added. The counter under the answer shows how many of your plan's characters the agent will use after saving.

Training starts as soon as the pair is saved and continues if you close the dialog. If the dialog says the pair is waiting to train, finish it with Train agent on the Knowledge tab. The pair is listed there under Knowledge sources, where you can edit it and add other phrasings.

Reply ratings

Visitors can rate each of the agent's replies with a thumbs up or a thumbs down, shown under the reply once it has finished. Clicking the same thumb again takes the rating back, and the other thumb changes it. Their tooltips, Helpful and Not helpful in English, follow the widget's language, which you set under Widget language on the Appearance tab.

Ratings are available in the chat bubble, the inline iframe, the direct link and the Playground. In the Playground, you rate the replies of the saved test chat, and your ratings appear in its transcript like a visitor's. The welcome messages can't be rated, and neither can replies in a temporary chat.

In a transcript, a rated reply shows a green thumbs up or a red thumbs down next to its time. To work through the weak answers, turn on the thumbs-down filter, read what the visitor asked, then fix the source or click Improve answer.

Export chat logs

Export CSV, next to the filters, downloads every conversation that matches the current search and filters, not only the page on screen. The button is greyed out when no conversation matches. The file is named after the agent and the date, and has one row per message with these columns:

ColumnContent
Conversation IDThe same for every message of one conversation
Started at (UTC)When the conversation started
CountryTwo-letter country code, if known
PageThe page the chat started on, if known
RoleVisitor, Agent, or Team member for a reply someone from your team wrote in a live chat
Time (UTC)When the message was sent
MessageThe message text. Buttons, product cards, a lead form and a chat form that came with a reply are written out in square brackets, and so are booking times, as [Booking times: followed by the event type's title. Files a visitor attached are listed by name after their message, as [Attached: followed by the file names.
RatingHelpful or Not helpful, for rated replies
ConfidenceThe reply's confidence from 0 to 1, with two decimals. Empty for visitor messages and for replies that weren't scored.
TopicsThe conversation's topics, separated by ; . Empty until the conversation is labelled.
SentimentPositive, Neutral or Negative. Empty until the conversation is labelled.
AuthorThe first name of the team member who wrote the reply in a live chat. Empty for the visitor's messages and the agent's replies.
Verified user IDThe visitor's user ID on your site, when identity verification verified them. Empty otherwise.
SubtopicsEach topic's subtopic as "Topic › Subtopic", separated by ; . Empty when no subtopic was given.

Cells that start with =, +, - or @ get a leading apostrophe, so spreadsheet apps show them as text instead of running them as formulas. One export holds up to 50,000 messages. A larger export stops there and ends with a row giving the total; narrow the date range or filters to export the rest.

Topics and sentiment

Topic analysis labels each conversation with up to two topics from a list you keep, and with how the visitor came across: positive, neutral or negative. Use it to filter conversations and to see what people ask about most on the dashboard and in the weekly email. It is off by default, per agent.

  1. Open your agent and go to Conversations. The Topics card is below Conversation History.
  2. Switch on Topic analysis. It saves at once.
  3. Add up to 15 topics of up to 40 characters each, such as "Shipping" or "Pricing", and click Save topics. Or click Suggest topics: intoCHAT reads up to 50 recent conversations and proposes 6 to 10 topics with subtopics, in place of your list, which you can edit before saving. Nothing is saved until you click Save topics.

If the list is still empty when the first nightly run comes, intoCHAT suggests a list from your recent conversations, saves it and uses it. You can edit it any time.

Subtopics

Each topic can have up to 8 subtopics that split it further, such as "Refunds" and "Invoices" under "Billing". Add them under each topic in the Topics card, or click Suggest subtopics: intoCHAT reads up to 50 recent conversations and proposes subtopics for the topics on your list, keeping the topics as they are. Suggested subtopics are added to the ones you have, and nothing is saved until you click Save topics. Subtopics follow the same rules as topics: up to 40 characters, no duplicates under one topic, and not "Other". The same subtopic name may sit under two topics.

When a topic has subtopics, the model also picks the one that fits, or none when none fits well. A conversation labelled with "Billing" may then also carry "Billing › Refunds". Subtopics are optional: a topic without subtopics works as before, and conversations labelled before you added subtopics keep their topics without one.

How labelling works:

  • Once a night, around 01:00 UTC, an AI model reads conversations that haven't been labelled yet. Transcripts are processed by OpenAI, as chats are.
  • A conversation is labelled once it has been quiet for 30 minutes, if it has at least one visitor message and had activity in the last 30 days. Each is labelled once; messages added later don't change its labels.
  • The model may only pick topics from your list, and subtopics from the list of their own topic. A conversation that fits none of the topics is labelled Other, which is added automatically and has no subtopics.
  • Up to 500 conversations per agent are labelled a night, newest first. The rest follow the next night.
  • It doesn't use your plan's messages and costs you nothing extra.
  • Changing the list affects only conversations labelled from then on. Switching it off stops labelling; existing labels stay.

Labels are a model's reading of the chat and can be wrong now and then. Temporary chats are never saved, so they are never labelled.

Temporary chats

Both the Playground and the chat widget on your site have a button in the header to start a temporary chat, marked "Temporary chat — messages aren't saved". Nothing from it is stored: it doesn't appear on the Conversations tab, the dashboard or the daily summary, and the lead form is never shown in it. Each message still counts toward your monthly message allowance; see Plans and limits.

Daily summary email

The daily summary is one email a day about all your agents. It is off by default. To turn it on, go to Settings, Account and switch on Daily summary email under Notifications. It goes to your account email address.

intoCHAT sends it once a day, around 07:00 UTC. It covers the 24 hours before it is sent and states when that period ended, in UTC. For each agent with activity, it lists:

  • New conversations: chats that started in the period.
  • Visitor messages: messages visitors sent. Your agent's replies are not counted.
  • New leads.
  • Replies rated not helpful, counted by when the rating was given, with up to three of the visitor questions behind them. Each has a See the conversation link.

Agents with no activity are left out, and the email says how many. When none of your agents had any activity, no email is sent. Saved Playground chats and their ratings are counted too. To stop the emails, switch the setting off; the link at the bottom of each email takes you there. Usage alerts are set in the same place; see Plans and limits.

Weekly insights email

The weekly insights email compares last week with the week before, for each of your agents. It is off by default. Switch on Weekly insights email under Notifications in Settings, Account.

It is sent once a week, usually on Monday morning (UTC), and covers the last full week, Monday to Sunday in UTC. If you switch it on mid-week, you get last week's email the next morning. For each agent with activity it shows, each with the change from the week before:

  • New conversations, Visitor messages, New leads, Form submissions and Hand-offs.
  • Replies rated Helpful and Not helpful.
  • The share of scored replies with low confidence, with the change in percentage points.
  • Top topics and the sentiment split, when topic analysis has labelled last week's conversations.
  • How many Questions to improve are open now, with links to the agent's Conversations and Knowledge tabs.

A change from zero shows as "new" instead of a percentage. Agents with no activity last week are left out. After a quiet week for all your agents, no email is sent.

The dashboard

The Dashboard covers all agents in your account. Pick a range at the top: 7 days, 30 days (the default) or 90 days. Days are counted in UTC.

  • Tiles: Conversations, Visitor messages, Leads captured and Messages per chat. Each shows the change compared with the previous period of the same length. Visitor messages count only what visitors sent; your agents' replies are not included.
  • Chart: conversations, messages or leads per day across the selected range.
  • Outcomes: what the conversations led to, described below.
  • Agents by activity: each agent's chats, messages and leads in the range, plus the time of its last chat.
  • Needs attention: problems worth acting on, such as an agent whose training failed, an agent with no knowledge yet, a message limit that is close or reached, or a live agent with no chats in the range, which often means the embed code is missing from your site.
  • Plan usage: Messages this month and Website crawls this month against your plan's limits.

Outcomes

Outcomes shows what conversations led to, for all agents or for one: pick it under Agent. It uses the range at the top of the page. Each card shows the change from the previous period of the same length; shares change in points (pts).

CardWhat it counts
ConversationsChats visitors started in the period, with the number of visitor messages.
ConvertedThe share of those conversations that produced a lead, a form submission or a hand-off. A returning contact counts toward their latest conversation.
New leadsContacts saved for the first time in the period. A returning contact isn't counted again.
Form submissionsChat forms visitors filled in and sent.
Hand-offsConversations passed to your team by email. One handed off twice counts once, on the latest day.
Button clicksClicks on your custom buttons in the chat bubble, the inline iframe and the direct link. Clicks in the Playground, in transcripts and in temporary chats aren't counted.
Rated helpfulOf the replies sent in the period that visitors rated, the share rated helpful.
Low-confidence repliesOf the replies sent in the period with a confidence score, the share with low confidence: a sign something is missing from your knowledge.

The Live chat card appears once an agent in scope has live chat on, or had live chats in the period. It shows Live chats, the conversations someone from your team took over in the period, counted by when they took over; Average rating, the mean of the ratings from 1 to 5 that visitors gave those chats; and Ratings, how many of them were rated. Each is compared with the previous period. Without live chats in the period, it links to the Live chat inbox instead.

Below the cards, a chart shows these per day, with a table view. Top buttons lists the most-clicked buttons, and Topics and sentiment shows a treemap of topics and the sentiment split of labelled conversations.

In the treemap, each topic is a tile sized by the labelled conversations from the selected period (7, 30 or 90 days) that carry it, and shaded by how many of them ended negative: the darker the red, the larger the negative share. Click a topic to see its subtopics the same way, with a "No subtopic" tile for its conversations without one. Click a subtopic to open that agent's Conversations with the Topic and Subtopic filters set; Open all … conversations does the same for the whole topic. When all agents are shown and a topic's conversations come from more than one agent, choose an agent in the selector first. Every tile can be reached with the keyboard and reads out its name, conversation count and negative share, and Show as a table lists the same numbers. A conversation can have two topics, so the tiles add up to more than the labelled conversations. Without topic analysis, that card explains how to switch it on. How these are counted repeats these definitions on the page.

What isn't tracked

Outcomes count what happened inside the chat: leads, forms, hand-offs and button clicks. They don't follow a visitor after they leave the chat, so a sale made later on your site isn't linked to the conversation. A button click means the link was opened, not that the visitor did anything on the page.

Ratings exist only for replies someone chose to rate, confidence measures how well your knowledge matched, not whether the reply was right, and topics and sentiment are a model's reading. Reading transcripts remains the way to judge answer quality.

View as Markdown