Why testing matters
You want to catch a wrong answer, a broken branch or a missing piece of knowledge before a customer does. Two things make that easy, and they are not the same thing: one screen tells you whether the AI has your content at all, and the Playground lets you talk to it. Start with the first — most "the AI gave a bad answer" reports turn out to be content that never reached it.
First, check what the AI actually stored
Open Brand → Knowledge, the one screen listing everything your brand knows. It answers the question before you ask it, in two places.
At the top, a summary card shows Sources (how many sources Yaplet reads), Indexed (how many searchable pieces it stored from all of them together) and Website last read. Below it, every source sits in a group — Knowledge bases & documentation, Website, Files, Text snippets, Questions & answers and API tools — and a group with a source that needs attention opens by itself. On a row, look for:
- For a knowledge base or documentation, how many pieces are indexed, for example 40 pieces indexed, or Not indexed yet when nothing is stored, which means the AI cannot answer from it at all.
- For a website, how many of its pages are learned, for example 12 of 40 pages learned, and when it was last checked.
- A status badge on websites, files, text snippets and question-and-answer rows. Ready — stored and searchable, which is what you want. Not indexed — nothing was stored from it. Reading… — we are still working through it, so wait and refresh. Warning — read, but not cleanly. Failed — a file, snippet or answer could not be stored. Stopped — a website whose reads kept failing, so its automatic checks switched off.
Hover a badge to see why. One article usually becomes several pieces, and only published articles count — a draft is invisible to the AI. Once a group has more than ten rows, a search box and a status filter appear above it, and if a website stopped, a warning appears when the page loads.
Try Yaplet in the Playground
- Go to Brand → Playground, the first tab of the brand menu.
- On the left, a status card says whether Yaplet is Trained or still learning your website. Once your site is learned, a list of Test questions written from it follows, then Yaplet's Model, Knowledge, Actions and Instructions. The test chat is on the right; in a narrower window, switch between the two with the Test / Settings buttons.
- Write a message as if you were a visitor. Or open a test question, press its send button, and compare the reply with the Answer on your website it shows.
- Yaplet answers from the brand's actual knowledge, and it already uses instructions and a model you have not saved yet — change a line, ask again, and save when you like the result.
Test chats in the Playground stay out of your inbox and your reports, so there is nothing to clean up afterwards. They are not free, though: each answer is paid for like any other AI answer, and a hand-off shows a card instead of reaching your team. The whole walkthrough is in The Playground: try Yaplet before your visitors do. To see the genuine widget the way a visitor does, use the live preview at Brand → Chat widget instead (on a wide screen it is already open beside the settings; in a narrower window, switch to it with the Preview button at the bottom of the page, and back with Edit). That one is not a simulation: whatever you write there becomes a real conversation in your inbox, and a real AI answer that counts towards your usage — so keep those test chats short, and delete them afterwards if they clutter the view.
What to check while you are in there
- Does it answer correctly, and does it link back to the right article?
- Does it refuse questions it should be able to answer? That points back to the Knowledge page.
- Does it behave the way you intended when it finds nothing? That is set on AI agent → Human hand-off, under When Yaplet has no answer, which offers two choices: Politely decline and Connect to a person. The second one is grayed out until Visitor asks Yaplet for a person is switched on, just above it in the same section.
- Does it hand over to a person at the right moment? (After a tapped Yes the handoff line is a fixed line in the visitor's language, and after a typed yes Yaplet phrases it itself — either way there is nothing to proofread, only the timing is yours.)
There is no preview for a workflow
Scripted workflows have no preview anywhere in the dashboard. The way to test one is to switch it on and open the widget yourself — on a staging page, or on a page where the widget is already installed. Then walk the whole thing:
- Click every button and answer every question, following each branch to its end.
- Test branching steps both ways — the "yes" path and the "no" path.
- Check that collected fields validate properly (email format, required fields).
- Confirm the request-agent step really does reach a person.
If the workflow runs on other channels, test it there too
The Runs on column on Brand → Workflows names the channels each workflow can start on: Chat, Social (Facebook Messenger and Instagram together), Email and Calls, or All channels. If it lists Social, Email or Calls, test it there as well. A channel missing from the list is one the workflow cannot run on — a Start a form step, for example, rules out Facebook and Instagram — and hovering the small info icon beside the list tells you exactly which step is in the way.
Test on a real page
For the most realistic run, install the Yaplet snippet on a staging site and open the widget in a fresh browser tab as a visitor would. This is the only way to catch triggers that depend on the page address, time on the page or visitor data — none of which the preview can reproduce.
Next step
Found a gap? Fill it with Add knowledge to your brand. Happy with the results? Turn Yaplet and workflows on or off shows you the switches.