Conversations
Review every conversation, read every step behind an answer, find the questions ARCY could not answer, see what last night's check flagged, and read how your chat is doing.
Conversations is what the chat agent said, what it could not answer, which days it sent to a person, and what visitors wanted.
Before you start
Conversations and the content gap questions need Store conversations on for the Product (Product); off, the table reads Question not stored and the counts still show.
What it changes
Conversations lists every conversation, searchable by name or message, with a link to the person on Users. Each row names the day it is gone, from your plan's window (Data retention). Intent narrows the list to one intent (below); Every intent clears it. A row carries its intent beside the time. A Conversation is one visitor's whole day, read after that day closes at midnight UTC, so a row still inside its day reads Read once the day closes. On Development no row carries one: intents are read on Production only. The Lifecycle stage menu narrows it to threads answered in one stage (Every stage clears it); a row carries the stage of its last answer, and Stuck when any answer went to a visitor who was stuck (Persona). Everyone, Signed in and Anonymous narrow it to who was asking: a row's name is a link to that person either way. A row whose last answer ran with no Procedure carries No Procedure: three such conversations asking for the same thing in a week have ARCY draft one (Procedures). A thread ARCY's nightly check flagged carries a chip for each problem, such as Asked the same question twice, and its transcript opens on Found by last night's check; a program reads them, never a model. Flagged last night, from Home, narrows the list as a chip you can remove. A link from a Procedure narrows the list to that Procedure's conversations, or to the ones that asked for it, as a chip you can remove.
What the agent did is one line over an answer with the number of steps behind it, and opens on a press. Two steps under the same words draw as one. Inside, every step reads in order with its outcome, and a field an Endpoint answered as a list or a record reads as its count and opens into its keys and values. A failed step says why: Did not answer, Refused, Took too long, Needs a sign-in, Blocked or Not found on the page; a failed API call adds its HTTP status and your API's message. A step that needed a yes carries The visitor said yes or The visitor said no; one they cancelled reads The visitor cancelled here. Anything the visitor typed, and any value the agent filled into your site, reads Masked: it never reached ARCY. A Response field whose Visible to is AI only or Neither reads Masked too by name, while the agent goes on using its value (Endpoints). A conversation handed to a person ends on Sent to a person, linking the ticket where your help desk gives one (Integrations).
Under an answer, a chip names what the Supervisor stopped, a Restriction, an unbacked promise, a Persona repeat or a slow check, and what came of it: rewritten or offered a person. The stopped words and the matched value are never stored. The same row shows the answer's Lifecycle stage and when its first word arrived. A document the visitor attached opens by its type while it is kept.
A finished flow run is replayed as the visitor saw it, under a summary such as "Rent a stroller · Completed · 4 of 5 steps". Its word comes from the steps, never from the run ending: Completed only where every step succeeded, otherwise Cancelled, Failed, Declined, or Partly done where some landed. The agent's own run leaves the same summary. With Store conversations off, only the summary stays.
Content gap lists the questions ARCY could not answer from your content, most asked first. Label topics has ARCY name the groups. A row filters the table above to its conversations; the chip clears it.
Each night ARCY does one thing with each of the most asked questions. When a page it already read answers it, ARCY adds a text source citing that page on Knowledge. When visitors got the wrong treatment, it proposes a change to Persona. Otherwise it asks your team one plain question in a chat named Questions from ARCY under Recents, shared by everyone in the workspace. Your reply there is saved to Knowledge as you wrote it. A question is asked once, and again only when many more people ask it.
Analytics closes the page, three numbers and four charts over one range, Last 30 days to start; the picker, shared with Home, offers 7, 14, 30 and 90 days and a custom range. Visitors, Conversations and Resolution rate lead it. Visitors counts every browser that loaded a page in the environment in the address, whatever it answered the banner.
Chat open rate draws how many of the people shown the chat opened it, each point read over that day and the six days before it. A day nobody was shown the chat leaves a gap in the line. Under 30 people shown, the line draws the people who opened it each day and no rate. These are day figures, never unique across a range.
The three charts under it count Conversations, the day unit your plan counts (Billing); on Development they are counted from the conversations.
Intent places each Production Conversation, once its day has closed, in one of seven: How to, Find, Do, Problem, Account and billing, Pre-purchase and Other, which a greeting alone is. All seven draw, at zero where nothing was asked.
How Conversations ended puts every Conversation in exactly one of three: Resolved by ARCY, which stayed with ARCY and had no answer rated down; Sent to a person; and Rated down, which stayed with ARCY but had an answer rated down. Resolution rate is the Resolved by ARCY share of the range. Nobody is asked whether it worked, so an unrated Conversation counts as resolved. A person was reached by a ticket opened, the summary given to the visitor, or your team answering from WhatsApp; a declined offer does not count. Why the handover happened lists the five causes. Counts show even with Store conversations off.
A share stays hidden until its population reaches 30; counts show either way.
Chat speed reads the first sign of work and the first word at p50 and p95, against the targets of 1 s, 2 s and 4 s, then Whole turn (p50) and ARCY's own time (p50), which is the whole turn with the time a visitor spent reading taken out. Where the time goes lists the parts of a turn under them, slowest first, the Visitor reading row is time nobody is waiting on ARCY for. Answers by Lifecycle stage counts answers per stage and how many went to a stuck visitor. Supervisor counts answers stopped, rewritten and offered a person.
When it goes wrong
The content gap table says the Product has no training content. Add a source on Knowledge.
See also
Knowledge
Read the two Sources ARCY drafted from your own pages, add your own, train them, see when ARCYBot follows robots.txt, and put a retired page back.
Users
Find one end user by id, name or email, check the attributes identify() sent, have ARCY write a summary of what they did, and open any session to read it row by row.