Chatbot readiness
Two independent lights: active versus draft, and trained versus not. Visitors need the first; good answers need both.
Browse topics
What it is
Every chatbot has two independent status lights, and almost every "why isn't it working?" question is really a question about which one is off.
Light 1 — Activation. Draft, active, or inactive. Decides whether the public page, the widget, and messaging channels exist at all.
Light 2 — Training. Never trained, pending, training, completed, updating, needs retraining, or failed. Decides whether the chatbot can answer from your knowledge.
They do not imply each other. Nothing about finishing training activates a chatbot, and nothing about activating one starts training. The four combinations behave completely differently:
| Activation | Training | What a visitor gets | What you get in Test Chatbot |
|---|---|---|---|
| Draft / inactive | Anything | Hosted page not found; widget refuses to load | A working chat |
| Active | Never trained | The page loads and says it is not ready yet | A working chat, possibly a setup-style reply |
| Active | Completed | A real answer from your knowledge | The same real answer |
| Active | Updating / needs retraining | The previous good answers, uninterrupted | The same |
Learn to read the failure, not guess at it. A not-found page is always an activation problem. A "not ready yet" reply is always a training problem. A blocked widget is neither — that is a domain or plan problem, covered in The widget is not showing.
When you would use it
- A link you shared says the chatbot was not found.
- Visitors say the chatbot will not answer, but it answers perfectly for you.
- Analytics shows a training badge you do not recognise.
Where to find it
Open Dashboard → Chatbots → your chatbot → Playground (activation) and Analytics (training health).
Reading light 1 — activation
New chatbots are created as draft. That is intentional: a half-configured chatbot should not greet a customer.
- Draft — never published. Everything public is off.
- Active — published. The public surfaces work.
- Inactive — published once, then paused. Behaves exactly like draft for visitors, but all your data is intact.
The Playground tab shows a note whenever the chatbot is not active, and it names the specific surfaces that are down: the public link, the website widget, the iframe, and messaging channels. Activate from that note or from Settings → Danger zone — see Activate or pause a chatbot.
Reading light 2 — training
Training is the processing that turns an uploaded document into something the chatbot can search. The states split into three groups:
Not ready — the chatbot cannot answer from this yet
- Not trained — nothing has ever finished processing.
- Pending / Training — work is in progress. Wait.
- Failed — something went wrong. Retry, and if it repeats see Knowledge stuck processing.
Ready
- Trained — the chatbot can answer from this source and cite it.
Ready, with an update in flight
- Updating — you changed knowledge and the new content is being synced. The chatbot keeps answering with the previous content until it lands. No action needed.
- Needs retraining — an update did not finish, so some answers may be missing the newest content. There is a retry.
The last group matters because it is the one people over-react to. Both states keep serving good answers. Do not deactivate the chatbot during a re-sync — that swaps working answers for a not-found page, which is strictly worse.
Steps
- Open Playground. Read the activation note (if any). Then open Analytics for the training status on the health card.
- If the chatbot is draft or inactive, activate it — from the note, or from Settings → Danger zone.
- Open Knowledge. Confirm at least one source shows as trained, not pending or failed. See Dataset status.
- Ask a real question through Test Chatbot in the chatbot header. This proves the knowledge works, but remember it proves nothing about whether visitors can reach the chatbot — Test Chatbot works on a draft.
- Open the hosted link in a private window. This is the step that actually tests the public path. See Share a hosted link.
- Only then embed the widget or share the link widely. Go-live checklist covers the rest.
Diagnose it in fifteen seconds
Match the symptom, not the guess:
| Symptom | Light | Fix |
|---|---|---|
| Hosted link says the chatbot was not found | Activation | Activate it, and recopy the current link |
| Public page loads but says "not ready yet" | Training | Add knowledge, wait for trained |
| Widget bubble never appears on your site | Neither — plan or domains | The widget is not showing |
| Answers exist but are wrong or generic | Neither — knowledge quality | Fix the source; check Analytics → Needs attention |
| Chatbot stopped answering everywhere at once | Neither — credits | Out of credits |
What you will see
The Playground tab states the activation problem and Analytics states the training problem, both in plain language rather than as codes. Knowledge shows a per-source badge. Channels shows a warning if you copy a link or a snippet while the chatbot is still a draft.
Limits and plan notes
Both lights are independent of your plan. Free can be fully ready — activated, trained, and answering real visitors on the hosted /agent/{handle} link. What Free does not include is the website widget, plugins, and messaging channels, and its knowledge cap is 400 KB per chatbot. See Plans and what you get.
Activation needs permission to edit chatbots. Adding knowledge needs permission to manage knowledge.
Common problems
The hosted link 404s but Test Chatbot works fine.
Classic. The chatbot is still a draft. Test Chatbot is the owner's private door and does not need activation; visitors do. Activate it.
The public page loads but says it is not ready.
Activation succeeded, training did not. Add at least one knowledge source and wait for the trained badge. See Untrained mode.
I removed my last knowledge source and the chatbot went back to untrained.
Correct — with nothing to search, it is untrained again. Add a source. See Remove knowledge.
Analytics says updating and I am worried visitors are seeing errors.
They are not. During an update the chatbot serves its previous answers. Leave it live.
Everything says ready and visitors still cannot reach it.
Then the problem is the door, not the chatbot: a widget domain, a plan gate, a stale handle after a rename, or access protection on the hosted page. Start with Hosted page 404 or Widget blocked.
Common questions
Does finishing training activate the chatbot?
No. Nothing activates a chatbot for you — that is deliberate, so a half-configured chatbot cannot greet a customer by accident.
How do I tell an activation problem from a training problem?
By the symptom. A not-found page is always activation. A "not ready yet" reply is always training. A missing widget bubble is neither — that is plan or allowed domains.
Should I deactivate while a crawl or re-sync runs?
No. Updating and needs-retraining both keep serving the previous good answers. Deactivating would swap working answers for a not-found page.
The link 404s but Test Chatbot works. Why?
The chatbot is still a draft. Test Chatbot is the owner's private door and does not need activation; visitors do.
Can a Free chatbot be fully ready?
Yes — activated, trained, and answering real visitors on the hosted link. Free excludes the website widget, plugins, and messaging channels, and caps knowledge at 400 KB per chatbot.
Was this article helpful?
Related articles
Activate or pause a chatbot
Turn the chatbot on for visitors, or pause it without losing anything. Training is an entirely separate status.
Untrained mode
An active chatbot with no finished training still loads and tells visitors it is not ready, rather than erroring.
Go-live checklist
Eight checks before your first real visitor: trained knowledge, your own wording, activation, a channel, and outside proof.
Dataset status and training
What each badge on the Knowledge table means, what the failure messages are telling you, and when to just wait.
Ready to try it on your own content?
Create a free workspace, add a document, and ask the questions your team is tired of answering.