Activation was a note someone else received

Login was the wrong activation

A login is easy to count and hard to mean. It fires when someone is curious. Harbor’s retention problem was being discussed as an activation problem, and activation had been defined as curiosity. Nora’s customer sentence was stricter. The product worked when another person received a note. Until that happened, the account was a tour.

The question was whether a simple classifier, given only early events, could recover the same cut the interviews were making by ear. If login separated returners, the empty state was fine and the problem lived later. If a delivered note separated them and login did not, the first screen was the wrong lesson.

Events, and the words for “it works”

Three events were enough, kept apart: login, draft started, note delivered. A draft that is never sent is not a quieter version of a send. It is a different outcome, and mixing them would have let the tour claim credit for sentences no one received. Delivery meant the recipient’s client got the note, not that the sender pressed a button into an error.

Twenty-two customer interviews were coded for the phrase that meant the product had worked. The code book was short on purpose. Recipient named, reply described, or neither. We were not mining attitudes. We were checking whether the human label and the event label were the same cut.

Notes sent to yourself do not count

Test notes sent to the same account were removed. So were notes sent to a second address the sender controlled, where we could see it. A person can complete the gesture of sending without completing the social fact the interviews described. Tour clicks with no recipient were kept as their own path, not dropped. They are the behavior of the old empty state, and deleting them would have hidden it.

Accounts that only ever logged in from a company onboarding session, with a success manager driving, were excluded. Those logins are staff. Week-two activity was censored if the account was not yet two weeks old at extract time. A young account is not a failure to return.

A classifier with an unimpressive feature list

The classifier compared login-only accounts with accounts still active two weeks later. Features were the early events and almost nothing else. No firmographic enrichment, no email-open score. Those would have made the model look sophisticated and would have pointed the interface at a person-type instead of a behavior. A delivered note dominated. Login did not separate returners. A draft without delivery barely did.

The interviews used the same cut without seeing the model. Twenty-two people named a recipient. The empty-state heatmap explains why the product was not producing that event. Attention stays on the tour at the top. The send control in the lower right is a cold corner. The first screen was teaching the tour.

A hot tour, a cold send

The heatmap is the old empty state. It is not a map of a feature people love. It is a map of a paragraph they were given instead of a person to write to. The tension on the compose blueprint sits on that tour, dashed and held, under the recipient field. The field itself is quiet. The question is the interface.

Who should receive this

The screen is a compose, not a product tour. The title is the question. One recipient field. Send note. The tour remains on the drawing so the thing we held back is still visible, and it is not in the way. Before, send was a control after the explanation. After, the explanation waits until a note has been delivered. Nora’s line was the layout: ask for a person, do not explain the product to an empty room.

The founder’s sentence, checked

Nora already had the label. The work was to show that the classifier, which had never heard the interviews, drew the same line, and then to change the first pixel to produce that event. Leah’s condition was twin movement. Delivered notes up, and week-two return up with them. A send that nobody came back from would have been a trick.

The send was the new activation

New accounts were split for four weeks. A delivered note in week one replaced login as the activation measure. Week-two return was the hold. We did not announce the metric change only in a dashboard while leaving the empty state alone. The measure and the screen moved together, which is the only way to know the measure was about the product.

Accounts with a delivered note rose 12%. Week-two return rose with that send. Qualitative and quantitative research had named one event. The interface was changed to make that event the first pixel.