Conversations
Every exchange a visitor has with your bot is stored so you can read it back. The text is kept for 7 days and then deleted; the measurements taken from it are kept, which is what lets your analytics cover a longer period than the transcripts do.
Reading a conversation
Open your project and go to Chatbot → Conversations. Each row is one visitor session: the question they opened with, how many messages they sent, how many of those the bot failed to answer, and when they were last active.
Open a row for the full transcript. Assistant turns carry what we measured at the time — whether the answer was given or a fallback, which fallback and why, how closely your content matched the question, how long the reply took, and links to the pages the answer was drawn from.
You can search the text of every retained message, and filter to just the threads where the bot fell back at least once. That filter is usually the fastest read on a bot that isn’t working: it shows you the sessions where somebody wanted something and left without it.
What gets kept, and for how long
A job runs every night and erases the wording of every message older than 7 days — what the visitor typed and what the assistant replied. Once erased it cannot be recovered, by you or by us.
What is kept after that is the measurement, not the content:
- When the exchange happened, and which session it belonged to.
- Whether the bot answered, and the reason if it declined.
- How closely your indexed content matched the question.
- Response time, token counts, and the model that served it.
- Which of your own pages were cited. These are your URLs, not visitor input, so they stay.
That split is why the numbers on your analytics page cover 30 days while the transcripts only cover 7. The unanswered-questions list is the one panel made of message text, so it uses the shorter window too — a longer one would only show you blanks.
What we never do with it
- We do not train models on your visitors’ conversations.
- Conversations are never shared between accounts. Every retrieval and every query is scoped to a single project.
- We do not ask visitors who they are, and the widget sets no advertising or cross-site tracking identifiers. The only identifier attached to a thread is a random value the widget generates in the browser.
The full statement is on the privacy page. If you need a specific thread removed before its 7 days are up, email support@tekltalk.com. Deleting the project removes everything it holds immediately.
Getting more out of it
Conversations are the cheapest research you will ever run. Two habits are worth forming:
- Read the unanswered filter weekly. Every thread in it is a question your site does not answer, in your visitors’ own words.
- Read the answered ones occasionally too. A technically correct answer assembled from the wrong page is invisible in the numbers and obvious in the transcript.
Because the text only lives 7 days, this has to be a habit rather than something you do when you get round to it. Anything you want to keep, copy out.