bumble bff was a feature. friendship deserves its own product.

If you have ever opened Bumble BFF for a week, matched with three people, and noticed that two of them were obviously looking for a partner and the third unmatched after one reply, this essay is the explanation.

Bumble BFF is friend-mode bolted onto a dating app. The matching layer is the same. The photos work the same. The swipe rhythm is the same. The user base is overwhelmingly people who came for romantic matching and decided to try the friend tab as a side experiment. The selection bias makes the friend tab a strictly worse version of the romance product, populated by people who were not really there to make friends.

We think the gap that adult friendship is supposed to fill is one of the largest unmet needs in modern life. We do not think you can fill it with a tab in the corner.

what a friendship-first product looks like

On Soulmate, the same five-step vibe questionnaire is used for friendship, relationship and community. The matching engine doesn't know which intent a soul selected at the embedding layer; the embedding is the same vector for everyone.

What the engine does know, after the embedding similarity is computed, is the intents both souls marked. A soul who marked friendship only is only surfaced to other friendship-only souls. A soul who marked relationship only is only surfaced to other relationship-only souls. A soul who marked both is surfaced to both.

The implication is structural. People in the friendship lane never have to wonder whether the person on the other side of the chat secretly wishes they were on a date. They marked friendship; the person on the other side marked friendship; the system enforced it.

what friendship matches actually look like

The conversations are different from romantic matches. Friendship matches tend to:

• Share specifics earlier. People say which city they live in, what their work is, what their weeknights look like. Romance matches dance around these details longer; friend matches just need them.
• Skip the flirty register entirely. There is no performative wit. People talk like adults who already share an interest.
• Move to plans faster. Most friendship matches surface a 'want to grab coffee next week' message within five exchanges. Romance matches do this on a much longer arc.
• Last longer if they don't lead to an in-person meeting. Friendship matches will sit at a 'we both like reading this one author' level for months without disappearing. Romance matches don't usually survive that pace.

rooms, which most apps haven't built

Soulmate also auto-forms rooms. When enough souls in the system share a centroid of values and passions, a room gets proposed. Compassionate Vegans. Quiet Souls. Wanderers. Stewards of the Earth. Builders.

Rooms are not Discord servers or Facebook groups. They are small spaces seeded by the algorithm and joined by people whose vectors match. People inside a room can write to the room or read it. They can also choose whether to be discoverable to other room members as one-to-one matches.

Most adult friendship, in the research, looks more like a room than a date. Knowing six other people in your city who read the same thing you read this week is closer to what people actually mean when they say they're lonely. The 1:1 friendship is the deepest layer; the room is the wider one. We think both are inputs to the same problem.

what bumble bff was trying to do

We are not dunking on Bumble BFF specifically. The feature was a reasonable hypothesis: people on a dating app sometimes want friends, the matching layer mostly works, why not let them try. The issue is structural. The same matching layer that optimises for romantic engagement produces a friend tab whose users are second-class citizens of a romantic product.

Friendship is not a fallback for failed romance. It is a separate primary outcome. The app has to be built around it from the engine up, not from the tab in.

The door is at byvibration.com. The first thing it asks is whether you want friends, partners, or community. The matching engine treats all three the same. The selection bias problem is gone.