When making marketing costs nothing, the only thing worth paying for is knowing what to make

For thirty years go-to-market was one human persuading another, and software helped the human do more of it. Mailchimp sent more emails. Ahrefs found more keywords. Every one of those products is a tool for a person: its value is the leverage it gives to someone who already knows what to do.

Two things broke at the same time, and they compound. The cost of producing marketing work went to roughly zero. A model writes the email, the landing page, the ad variant, the comparison table. Not badly — well enough that the marginal value of “we can write this for you” is collapsing in real time, and will keep collapsing with every model release.

And the reader stopped being reliably human. Search answers instead of linking. Meanwhile the buyer is becoming an agent too. Taken separately, each is a headache. Taken together they invert the job: production became free and the audience became a machine, so the two things marketing software has always sold — making the asset, and getting it in front of a person — both lost their scarcity in the same eighteen months.

When one input goes free, value moves to whatever is still constrained. Two things are.

Being legible and citable to a machine. You could always buy your way to a keyword. You cannot buy your way into an answer. An assistant that summarises a category and recommends three products is making an editorial decision from whatever it can read, parse and trust about you.

Knowing which action actually moved a number. This is the deeper one. When you could only produce five things a quarter, you didn't need to know which one worked — you barely had a choice. When you can produce five hundred, choosing is the entire job, and nobody can do it, because the feedback loop that would tell you was never built. Analytics answers what happened. It does not answer which of the things I did caused it, and what should I do next.

So the bottleneck moved from production to judgment. And judgment is the one thing better models do not hand you for free — because judgment about your product, in your market, at your stage requires outcome data about products like yours, and that data does not exist on the internet. It only exists inside a system that both acted and measured.

Generate. Every AI marketing assistant — draft the copy, produce the calendar, spin the variants. This is where almost all the money and almost all the products are, and it is standing on the thing that is actively going to zero.

Measure. The AI-visibility category: how often a brand appears in AI answers, for which prompts, against whom. Real, funded, and correct about the problem — but it is a dashboard. It tells you that you are invisible; it does not make you visible, it cannot prove that anything you did changed the number, and it does not get smarter with use.

Close the loop. Measure, act, prove the action moved the metric, learn which actions work for which kinds of product, and carry that forward. Nobody owns this position, and the reason is not that nobody thought of it. Closing the loop requires execution rights — a key, a repo, a send — and execution rights are exactly what nobody wants to hand a stranger.

That difficulty is the opportunity. The first two positions are races. The third is a moat, and its entry fee is trust.

If the hard part is getting write access, the right first customer is not the biggest segment. It is the segment where the trust problem is already solved: founders building with Cursor, Claude Code and Lovable, who already let an agent write to their repository every day. They have crossed the line every other segment is years from crossing.

Their product is a repo, which means the highest-leverage marketing changes — the landing page, the pricing copy, the structured data that makes them legible to an assistant, the instrumentation that measures any of it — are all code changes, reachable by a pull request they review in the tooling they already use.

Every safety property follows from the same place. A pull request is the best consent mechanism ever designed. It is a proposal, it is diffable, it is approved by a human, it is auditable, and it is revertible. We do not ask for trust. We ask for a review.

But the repo is the door, not the ceiling. Most go-to-market happens outside the product — on Reddit, on Product Hunt, on LinkedIn, in review sites and marketplaces, in the inbox. What the repo buys is the first yes. Once a founder is approving work every day, the same one-click habit extends outward. The hard part of this business was never the second channel. It was the first grant of permission.

Every time a loop closes, it leaves a fact behind: for a product of this shape, in this market, at this stage, this action moved this metric by this much. One of those facts is an anecdote. Ten thousand of them, anonymised and connected, is the only empirical prior on go-to-market that has ever existed — because everything this industry runs on today is a survey, a case study written by the winner, or a consultant's memory.

We call it the Bladesmith Outcome Graph, and its defining property is that it is cross-project: your measured outcome improves the first guess we give the next founder whose product looks like yours, and theirs improves yours.

A frontier model will never have this. It will write better copy than us forever — good, we use it for exactly that. It will not know which of two onboarding emails converted for the four hundred products most similar to yours, because that fact was never published. It exists only where somebody both shipped the change and measured what happened.

We get better when they get better, on three axes at once: drafting quality rises for free, cost per action falls, and the ceiling on what an agent can safely execute rises — which converts things that are advice today into things that are done tomorrow. We are long model progress, not hedged against it.

Meanwhile the asset is untouched. A better model does not generate outcome data; it consumes it. If anything, the value of proprietary outcome data rises as generation commoditises — because when everyone can produce infinite plausible options, the scarce good is the evidence for choosing among them.

An argument you cannot falsify is a slogan. Four things would move us, in rising order of severity.

If zero-click reverses — regulation, publisher lawsuits, or search restoring referral traffic — the urgency of the discovery half weakens. It would not touch the agent-buyer half, and it would not touch measurement.

If agent-mediated buying stalls. Forecasts about agent adoption have been wrong in both directions. This would delay the value of the distribution work, not erase it — which is why we never let distribution outrun measurement.

If a foundation-model lab ships the loop. The real one. Our defence is structural rather than 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 is a defence with a shelf life, and we would rather say so out loud than pretend otherwise.

If the graph doesn't transfer. The deepest risk: what if what worked for product A genuinely does not predict product B? Then the cross-project layer is decoration. This is testable, and we intend to test it early — once there are enough projects, measure whether a prior drawn from similar products beats the cold-start recommendation. If it doesn't, the thesis needs surgery, not spin.

One test for every feature. Does this compound, or does it trend to free as models improve? Compounds, invest. Trends to free, build the cheapest honest version and move on.

Measurement is not a phase, it is the product. An action whose outcome we cannot measure has not been executed, it has merely been performed.

Honesty is a feature, and it is enforced. A button that looks like it does something must do it. An agent must be named for what it actually does. Gartner expects more than 40% of agentic AI projects to be cancelled by 2027 — not because the technology failed, but because the value was never proven. We are building the thing that proves it. We do not get to be sloppy about proof.