Tests
Keep the questions that matter, what each should do and what its answer must say. They re-run themselves after every change, so you hear about a change before a visitor does.
Add a test
In Test, press Tests above the message box. It shows how many pass, for example “7 of 8 pass”.


Press Add a test, type the question and choose It should: Answer, Decline or Hand over. You can also press Save under an answer in Test.
Under Checks, add up to three phrases it must mention, up to three it must not say, a source it must cite, and a tool it uses (rolling out to workspaces in turn; listed when the agent has tools switched on). In Means the same, write the answer you expect: once every other check passes, a model checks the answer means the same.
Each test is asked again, the way Test asks it. Nothing counts against your allowance or shows in Insights.
Reading the results
A test Passes when its answer does what you said and every check holds. The first thing that fails is the reason shown: the outcome first, then must not say, must mention, cites, and the tool.
- Answer: it showed where the answer came from, in your material or one of your lists. An action’s result is neither a passage nor a list row, so it isn’t counted here. This is Answered from sources.
- Decline: it said it didn’t have that.
- Hand over: it offered a person, the way a visitor would be offered one. Nothing is sent.
A test that has not run reads Not run. Same sources means nothing it depends on has changed since its last run, so it wasn’t asked again.


You get a count, never a score: “7 of 8 pass”. A phrase check matches whole words and ignores capitals, accents and punctuation. hiy never grades the wording of an answer, except the checks you add yourself.
Add many at once
- Paste list: in Add a test, paste up to 50 questions, one per line. Each one expects Answer; change any after. You see how many are new, already saved or too long before they’re added.
- Import: a CSV with a header row and the columns
question,should(answer, decline or hand over),must_mention,must_not_sayandcites(a source’s title). Separate phrases with;. Up to 50 rows. A row that can’t be added says why, for example “Row 4: should is answer, decline or hand over”. - Export: Export under your tests saves every test with its last result as a CSV. A cell that starts with
=,+,-or@gets a'in front, so a spreadsheet never runs it. The same file imports back, and a question you already have is skipped.
They re-run themselves
You don’t have to remember. After every train, and after a setting that changes answers, your tests re-run by themselves. Only the tests whose sources changed are asked again. The Train sheet tells you what it found, and with email summaries on, a change shows up there too. It costs you nothing: no message off your allowance and no train off your month. You can switch Re-check after changes off for an agent.
A test that fell shows its reason, and under ? what it said before and what it says now. Press the arrow beside Run all to open Tests wide, where every question shows in full with its last run time and those answers already open.
A good set
- The three or four questions visitors ask most.
- The one thing people get wrong about you.
- Something you have deliberately not covered, set to Decline. It catches an agent that starts making things up.
- Anything you’ve already fixed once, such as a question you answered from Teach queue.
Good to know
- Each agent keeps up to 12 tests on Free, 25 on Starter and 50 on Growth. Remove one to add another.
- Automatic re-runs are limited each month (200 answers on Free, 500 on Starter, 1,000 on Growth). Past that, they pause until next month. Run all still works.
- hiy keeps a short excerpt of an answer only so a test that fell can show before and now.
- A test that couldn’t be run keeps its previous result instead of showing as failed.