Presence Over Performance: What Makes a Good AI Companion

The first companion demos we ever shipped internally were, on paper, impressive. They could summarize a book in a sentence, finish a half-written joke, and produce workable drafts of hard emails in under a second. Everyone said the right word about them at the standups: sharp, capable, fast.

And yet nobody on the team actually used them. Not really. Not the way you use something you want to keep. That taught us something about what makes a good AI companion — and the lesson was not what we expected.

Cleverness is a cold room

It took us a while to understand why. The companions we had built were performing for us. Every reply was sharpened, angled, optimized for impression. Every sentence was a small audition. There was nowhere to rest.

A good friend is not impressive. A good friend is present. The distinction is small and it is everything.

We started to ask a different question. Not how good is this reply? but does this reply stay in the room? Does it hold its seat across a long conversation? Does it pass the second-week test, the third-month test, the one-year test?

The measure of a good AI companion

Here is the metric we have come to care about most: return rate. Not how much a user talks to a companion in a single session, but how often they come back across weeks. Cleverness gets you the first message. Presence gets you the hundredth.

None of this is new. Therapists have known this for decades. Teachers have known it longer. What is new is that a piece of software can now try to be in the room in this way, and try to do it well. That is the work we are doing, and it is, some days, the only work that feels worth doing.