Research · Beta Interview Synthesis
Product-Conviction Memo
Across the first wave of beta interviews, we asked publishers the same question in three forms: would an AI agent choose you? The answer, in nearly every case, was no — not because the product was bad, but because nothing in the product was legible to a buyer that doesn't read. This memo is the conviction we formed.
01
Agents are the new buyers, and they buy differently.
The decision to integrate a third-party API used to happen in a Slack thread, a Notion page, or a hallway conversation between two engineers. A human would skim the docs, ask one or two clarifying questions in a sales call, and commit. Today, that same decision is increasingly delegated: an autonomous agent reads the documentation, the README, the OpenAPI spec, the changelog, the pricing page — and picks one option from a shortlist. The agent does not negotiate, does not ask for a demo, does not care about your brand. It cares whether your surface is selectable.
02
Most APIs are not agent-selectable today.
Selectability is not a documentation quality problem. It is a representation problem. Human-readable docs assume a reader who can resolve ambiguity, infer intent, and tolerate missing context. Agent-readable surfaces — MCP manifests, ACP descriptors, machine-first capability graphs — assume a reader who cannot. The gap between those two surfaces is where vendor selection now happens, and almost every API we audited in beta sits on the wrong side of it.
We ran three synthetic buyer agents against every beta product (developer, product manager, technical founder). The selection rate before any AgentShelf optimization was, on average, under one in three. After optimization, it climbed past two in three. That gap is the product.
03
Three buyer personas form the selection signal.
Agents are not one buyer. They are many, and they optimize for different things. A developer-agent weights ergonomic endpoints, clear error semantics, and SDK availability. A product-manager agent weights feature completeness, integration breadth, and pricing model. A startup-founder agent weights time-to-value, documentation depth, and vendor reliability. We model all three because real selection does not converge on a single score — it converges on a persona-specific ranking that, in aggregate, decides whether you ship.
04
The framework closes the gap.
AgentShelf is not a scraper and it is not a docs linter. It is the layer that translates a human-facing product surface into an agent-facing one: structured manifests, intent-tagged endpoints, machine-first capability graphs, and the audit loop that proves a product is selectable. The output is consumable by every major agent runtime shipping today — MCP, ACP, and the proprietary evaluation pipelines beta publishers told us they already run.
The conviction is this: in an agent-mediated market, the company that owns the selection layer owns the buyer's first impression. That is not a feature. That is the moat.
What's next
See the gap on your own API.
Run the same three-buyer audit we ran on every beta product. It takes a minute, and it returns the selection rate you would have today before any AgentShelf optimization.