How we decide which surface to build next

From outside, a company that ships a job search, a skills assessment, an employer-of-record desk, a chat surface, a sports and news companion, a career astrology product, an inbox and an on-device document tool looks like it cannot decide what it is.

The list is not the strategy. The test each one had to pass is.

The gate

Before anything gets built, it has to clear four things. Not three of four.

One: people are already asking the agent for it. Not a survey, not a market map. Observed drift in what our own users type. A growing share of what people ask the agent has nothing to do with jobs any more, and that share is not evenly distributed. It clusters. Those clusters are the roadmap, and they are the only source of it.

Two: it adds no inference cost. This is the constraint that does most of the work. A candidate surface has to run on the models and the router we already have. If it needs a bigger model, a different stack or a per-query bill we do not currently carry, it does not get built, however good the idea is. That single rule is why we can afford every surface we run and a competitor with the same headcount cannot afford three.

Three: it makes another surface better. A product that stands alone is a distraction wearing a roadmap's clothes. Every surface has to feed something. Skill Tests make Job Search more accurate. The employer desk creates listings nobody else has. Mail sells into companies the hiring business already knows.

Four: it answers a question about the business, not just about the user. This is the one people find strange, and it is the most useful.

Surfaces as experiments

We had two questions we could not answer from the hiring business, and both were existential.

Would people come back to the agent on a day when nothing about their career was at stake? And would an Indian consumer pay, in cash, with no subsidy and no employer standing behind the invoice?

We could have run studies. Instead we built Signal to answer the first and Luma to answer the second, and had both answers inside ninety days, from people we had already acquired once.

Both are now revenue lines. That was not the point of them. The point was that a shipped product answers a question about consumer behaviour in a way that no amount of research does, and if you choose the product carefully, you get the answer and keep the asset.

What we said no to

The gate is only meaningful if things fail it, so: a standalone assessment business that would have sold to companies not hiring through us, failed three. A general consumer feed with no career context, failed one and four. Several products a large model would have made possible and our cost base would not survive, all failed two.

The hardest rejections are the ones that fail only rule two, because they are usually good ideas. A surface that needs a frontier model to work is a surface that turns our best structural advantage into somebody else's revenue.

Why the list looks incoherent from outside

Because the coherence is underneath rather than across.

Judged as a product portfolio, a jobs board and a horoscope have nothing in common. Judged as a system, they are the same agent, the same models, the same router and the same cost base, pointed at different moments in the same person's life. The person who used us to get hired in March is the person reading a match on Signal in November, and the company that hired them is the one paying for Mail.

That is what the ten entries in the ecosystem actually are. One thing, met at ten different moments.