The case, in the open
Everything we believe is on this site already — the argument in the manifesto, the product in how it works, the honest inventory on agents. This page is the short version, plus the part most companies leave out: what would prove us wrong.
Why now
Three clocks are running, and none of them are ours
Discovery is being replaced in public on a two-year timescale. The trust window that lets an agent write to a production repo opened eighteen months ago. And a feedback-loop moat only compounds while it runs — a month not compounding is permanently gone.
69%
of searches end without a click
Up from 56% the year before.
−61%
click-through where an AI answer appears
The rank still exists. The traffic does not.
90%
of B2B buying agent-intermediated by 2028
Gartner. Roughly $15T through agent-to-agent exchanges.
>40%
of agentic-AI projects cancelled by 2027
Gartner — on cost and unproven value. Proof is the constraint.
Sources: AI traffic decline, 2026 · Gartner via Oro · Gartner on agentic project cancellations.
The argument
Five beats
Each one is expanded, with its evidence, in the manifesto.
Two things broke at once, and they compound
Producing marketing work went to roughly zero — a model writes the page, the post, the sequence. And the reader stopped being reliably human. The two things marketing software has always sold, making the asset and putting it in front of a person, both lost their scarcity inside eighteen months.
Value moved to judgment, and judgment cannot be bought
When you could produce five things a quarter you did not need to choose. At five hundred, choosing is the whole job — and nobody can, because the feedback loop was never built. Judgment about your product in your market requires outcome data about products like yours, and that data does not exist on the internet.
Three positions exist. Two of them are races
Generate is standing on the input going to zero. Measure — the funded AI-visibility category — is right about the problem and ships a dashboard: it cannot make you visible, cannot prove a change worked, and does not get smarter with use. Closing the loop is empty, because it requires execution rights nobody hands a stranger.
The wedge is a trust problem, not a market-size problem
Founders building with Cursor, Claude Code and Lovable already let an agent write to their repository. Their product is a repo, so the highest-leverage marketing changes are code changes, reachable by a pull request they review in the tooling they already use. It is the one segment where the trust problem is already solved.
The asset is the only thing that appreciates
Every closed loop leaves a fact: for a product of this shape, at this stage, this action moved this metric by this much. Cross-project, that is a genuine multi-tenant network effect — and a frontier model will never have it, because the fact was never published.
The moat
The Outcome Graph is the compounding asset
One measured outcome is an anecdote. Ten thousand, anonymised and connected, is the only empirical prior on go-to-market that has ever existed — because everything the industry runs on today is a survey, a case study written by the winner, or a consultant's memory. Its defining property is that it is cross-project: your result improves the first recommendation the next founder gets.
Products, linked by learned similarity
IllustrativeThe model
Free intelligence. Paid hands.
We never charge for the part whose price is collapsing. The plan, the analysis, the diagnosis and the assets are free and stay free — that is the acquisition layer. Executed actions are metered, and charged on success. You pay for hands and for judgment, never for words.
The number we run on
Weekly active projects with an approved, executed action and a measured outcome
One metric that only rises if the whole thing works: adoption, trust and proven value at once. It is also, exactly, the rate at which the moat compounds. Notice that it is not revenue — that is deliberate.
What would prove us wrong
A thesis you cannot falsify is a slogan
Four things would move us, in rising order of severity. We would rather you test them with us than discover them later.
Zero-click reverses
Regulation, publisher lawsuits or search restoring referral traffic. It weakens the urgency of the discovery half; it touches neither the agent-buyer half nor measurement. Cost of being wrong: low.
Agent-mediated buying stalls
Forecasts about agent adoption have been wrong in both directions. This delays the value of the distribution work rather than erasing it — a sequencing error, hedged by never letting distribution outrun measurement.
A foundation-model lab ships the loop
The real one. Our defence is structural, not technical: they have the model, not the write access, not the per-product measurement, and not a reason to take on the liability of executing inside somebody else's business. That defence has a shelf life, and we would rather say so.
The graph does not transfer
The deepest risk: what if what worked for product A genuinely does not predict product B? Then the cross-project layer is decoration. It is testable — measure whether a prior drawn from similar products beats the cold start — and we intend to know that answer early rather than late.
Getting in touch
The memo, the build state and the current numbers on request
This page is deliberately the argument and nothing else — no round size, no terms, no traction slide. Those belong in a conversation where we can show you the loop running on a real product and tell you plainly what is built and what is not.
hello@bladesmith.io