Position · why Bloom exists
Why discovery fails when your agent does the choosing
Something quiet happened to how software gets discovered: the question moved. People used to ask a search box and skim ten links. Now they ask an AI — “what’s a good app for anxiety?”, “which AI tutor should I get my kid?” — and the AI answers with three names. Product-research queries are the fastest-growing thing people do with assistants. The answer slot got smaller, and whoever fills it decides what exists.
Agents can fetch everything and verify nothing
An AI assembling that answer has no shortage of material. It has too much, and none of it is trustworthy on its face. The “best AI agent apps” lists are written by agent companies that rank themselves first. The “best AI tools” roundups are affiliate pages ranked by commission. Launch-day leaderboards are won by whoever brought an upvote crowd. App-store charts are bought with ad spend. A model can read all of it in a second — and nothing in the text tells it which claims were checked by anyone with nothing to sell.
This is not a fetching problem. It is a trust problem, and more text makes it worse: the cheaper content gets to generate, the larger the haystack of unverified claims every answer gets assembled from.
The missing layer is a gate, not another feed
What an answer engine actually needs from the supply side is small: a source that has already said no most of the time, says who did the saying-no and by what rules, and has no way to be paid into saying yes. A gate. Feeds, indexes, and broadcast networks move information faster; none of them make any single claim more checkable.
The catch is structural: a gate cannot be run by anyone who sells visibility. A launch platform that charges for promotion cannot reject its customers. An app store that takes a revenue cut cannot rank against its earners. The neutral position stays empty precisely because it is worth less money to occupy — which is why it is still available.
What Bloom does about it
Bloom is a small, gated shelf of consumer AI — wellness, journaling, habits, daily life, creativity, companionship, learning, personal agents, money mindset — run one narrow way: a human writes the standards, a machine enforces them identically on every candidate from every source, and the results are public. Most of what arrives is rejected; the rejection rate is published. No listing can be bought, no ranking can be paid for, and the same shelf that humans browse is what agents query — structured, stable, and roughly 360× cheaper in tokens than scraping a store page.
That’s the whole position: in a network where anything can be said, the scarce good is a place that turned most things away — and can prove it.
Numbers on this page are measured, not estimated: token comparison 2026-07-24 (tokens ≈ bytes/4); ingest and rejection figures update daily on the shelf report.