How we choose the five questions you get each day

Five questions arrive each day, the same for both of you. People reasonably assume there is a model behind this trained on what couples write.

There is not, and there cannot be, because we cannot read what anybody writes. The selection works on much cruder information, and the crudeness turns out to matter less than expected.

What the selection can see

How long a couple has been paired. Which questions have already been sent to that pair. Which were skipped. What language the devices are in. Nothing else.

It cannot see answers, because there are none to see in readable form.

The shape of a day

Five is a deliberate number. Three is too few to allow a bad question, and seven becomes homework.

The mix is the important part. On a typical day: two light, two ordinary, one that requires a moment of thought. Never five serious ones.

That last constraint is the one we hold most firmly, because it is the difference between a habit and an obligation. A product that asks something significant five times a day is one that people admire and abandon.

The unserious questions are not filler. They are what makes the serious ones possible on a Tuesday.

What changes over time

New couples get a broader, lighter set. The first fortnight is when a habit either forms or does not, and the fastest way to break it is a question in week one that neither person is ready to answer.

Long-paired couples get more questions about change, because the questions about the past run out. A couple four years in has already answered everything anyone can ask about childhood, and the useful material is now in what has moved since last year.

Skipping is the only signal we have

When a question is skipped by many couples, we look at it.

Usually it is one of three faults. It is unanswerable without a confession, so people leave it. It has an obviously correct answer, so it produces compliance. Or it is badly worded in one language and fine in the others, which is the most common and the least visible.

This is a coarse instrument. We cannot tell whether a question produced a good conversation, because we cannot see the conversation. We can only tell whether people were willing to answer it.

We consider that limitation acceptable and it is genuinely a limitation.

What we do not do

We do not send questions based on inferred mood.

We do not send anything intended to provoke a reaction, which is a real pattern in engagement-driven products and produces short-term activity and long-term distrust.

We do not repeat a question to a couple who answered it, except deliberately and after a long interval, because the same question two years apart is a different question and the two answers together are frequently the most interesting thing in the product.

And we do not adjust the mix to increase usage. There is no version of this where more time in the app is the goal.

Written by people

Every question is written and revised by people, in English and Estonian first and then rewritten rather than translated into the other languages.

The rule that governs all of them is the same one: the question addresses you about your partner, never in the first person, because both of you receive the same text and it has to mean the same thing to both readers.

It is a slower way to build a question bank than generating one. It is the only way we know of to build one where a person on the other end can tell that somebody wrote it.