Testing DM conversations

    All CollectionsDMsTesting DM conversations

    In your agent's settings, open Testing and choose DMs. You can write a conversation as a customer would, or import a real one and continue it, and check each message with the same decision steps your agent uses on real DMs: the same DM automations and their triggers, DM inbox handling, page settings, knowledge and checks. Nothing is sent, and nothing in your workspace changes. The result is what production would decide for that message right now; the AI can word a reply differently from one run to the next.

    Start a test

    1. Pick the page and platform. Only pages assigned to this agent are listed. Messenger and Instagram can behave differently (which automations apply, how long a message may be), so choose the one you want to check.
    2. Optionally add a customer name. The AI sees it the way it sees a real customer's name.
    3. Write the customer's message and click Send and test (or press ⌘/Ctrl + Enter).

    To set up earlier context, switch Next message from to the page and add what the page already said. Page messages you add are not tested; they only become part of the conversation.

    Import a real conversation

    Click Import, pick a conversation from the page, then choose the message to continue from. Messages after it are left out, and your next test message arrives right after it. Up to 40 messages before that point are imported; page messages older than that still count for automations that answer only the first message. Imported messages are given to the AI with their stored attachments, as in DMs. Importing only reads the conversation.

    If a teammate had taken the conversation over, the test starts the same way: the AI stays quiet. Click Give back to the AI to test as if the AI owned it.

    Reading the result

    Click any customer message to see its decision.

    • The outcome: replies automatically, saves a draft for review, no reply, automations conflict, needs attention, handed to a person, or test incomplete.
    • Who answers: the DM automation that matched (by name), or DM inbox handling, or nobody and why. For example: no automation matched and DM inbox handling is off, or the page does not allow inbox fallback, or two automations matched at once.
    • The reply or draft exactly as production would send or save it. It is always written by the AI.
    • A rejected reply, when the AI's text failed a check (for example a link that is not in your visible instructions). It is crossed out and marked as not deliverable, with the reason. Production never sends it; in a real conversation it would be marked as needing your reply.
    • What the AI wrote before cleanup, when production would send a cleaned-up version (extra spaces or unreadable characters removed). The reply shown above it is the one production sends.
    • The AI's note: its short explanation of the decision. It is never sent.
    • Would production attempt to send it now? For automatic replies: eligible to attempt, blocked (for example the page connection needs attention, or automatic DM sending is paused) or unknown (for example your settings changed while the test ran), with the time it was checked and each check. Eligible means production would try to send; whether the platform accepts the message is only known when it is sent.

    Continuing the conversation

    A generated reply or draft is not added to the conversation by itself. Click Use as the page's next test message to add it, or Edit first to change it (it is then marked as written by you). This matters for automations set to answer only the first message of a conversation: they count a page message only when it is in the test conversation.

    When a message is handed to a person (for example the customer asked for one), later messages are tested the way production handles them: the AI stays quiet until you give the conversation back.

    Testing an Opening + follow-up

    When a DM automation with an Opening + follow-up matches, the result shows the opening with its button, as the customer would see it. Click Use as the page's next test message to add it to the conversation; a Tap “…” action then appears under it.

    • Tap the button to test the follow-up. The result shows which automation sends it, the exact follow-up production would send (with its button, if it has one), and whether production would send it now.
    • Or type the button's text as the customer's next message. As in production, it counts as tapping the button when exactly one opening is waiting for a tap and the text matches its label (capitalization and spacing are ignored). Otherwise it is answered like any other message.
    • An opening sent after a comment works the same way: import the conversation, choose the opening (or a later message) as the point to continue from, and tap it.

    The same rules as production apply:

    • A newer opening from the same automation replaces an older one in the same conversation, so the older button no longer starts the follow-up.
    • A button starts one follow-up only. An opening already tapped before the point you imported from cannot be tapped again.
    • If the automation's publishing is Draft only, the opening is a draft: a person sends it as text, without the button, so there is nothing to tap and no follow-up.
    • The follow-up is written from the automation as it is saved now. If you removed the follow-up, or the automation no longer applies to this page, the test says so and nothing would be sent.
    • When a person owns the conversation, the follow-up is still written, but production would not send it, and the result says why.

    Unsaved settings

    Tests use your unsaved DM inbox handling, DM instructions and AI disclosure changes, so you can try a change before saving it. The Allow inbox fallback switch in the tester lets you try the page's setting (from Pages) for the test only. Other unsaved changes are not used by DM tests, and the tester says so.

    What a DM test can't know

    • Whether the platform accepts the message. Nothing is sent, so delivery is not tested.
    • Messages sent in quick succession. Each test message is handled on its own, as if it just arrived.
    • Send caps. The current values are shown, not applied.
    • A teammate taking over in the middle of a conversation.
    • Recovery of a button tap whose notification never arrived (production checks the conversation for it later).

    Typed messages are text only; imported ones keep their attachments.

    Limits

    Tests count toward the same per-workspace testing limit as comment tests: a few at a time and a limited number per minute.