Why AI Agents Aren't Picking Your SDK (and How to Fix It)

You're in the listicles. Your GitHub stars are healthy. And yet when a developer tells their agent "add search" or "add payments," your SDK isn't the one that gets imported. What's going on?

Nine times out of ten it isn't awareness. The model knows you exist. The problem is Discovery: the agent can't read you well enough, right now, to choose you with confidence. Here's how to diagnose and fix it.

Start with the Probe/Serve gap

Pull your logs and look at AI crawler traffic. Are bots like the ones behind Claude, GPT, and Gemini hitting your docs? Now look at what you returned them: a 200 with full content, or a JS shell, a 403, or a thin stub? The gap between how much they probe and how much you serve is the single most common cause of known-but-not-picked. They're knocking. You're not answering.

Ship the machine-readable surface

An agent can only pick what it can parse. The cheap, high-leverage fixes:

• llms.txt — a plain-text map pointing agents straight at your most useful docs.
• OpenAPI — a real, current spec so the agent knows your endpoints, params, and types without guessing.
• MCP server — where it fits, expose your tool as something an agent can call directly, not just read about.
• robots.txt — confirm you're not accidentally blocking the crawlers you want.

Test with unbranded prompts

The trap is testing with branded prompts. "Show me how to use Pinecone" will always produce Pinecone code — that measures whether the model can use you, not whether it picks you. Test the way real developers actually prompt: "add a vector database," unbranded, many times. That's where the truth lives, and it's exactly how Pickrate scores.

When you do, you'll often find the ranking is closer than the marketing suggests. In vector databases, the leader wins around 30% of trials while the open option behind it takes 13% — a gap that real Discovery work can close.

The diagnosis loop

Fixing Discovery blind is guesswork. The faster path: measure your Pick Rate, see which prompts you lose and to whom, ship the machine-readable surface, then re-measure and watch the number. That regression — inputs you control against the Pick Rate outcome — is the whole game. Check where your SDK stands and start closing the gap.