How do you measure a partnership in the agentic era, when the one choosing your product isn't a person anymore, but an AI?I've closed more than 75 partnerships per edition of Blockchain Jungle and run the pipeline at other startups I've worked for, so I know what you're supposed to count to prove one worked. Here's the problem: your most important partner in 2026 never signed a contract, doesn't take your calls, and just made the thing we count stop mattering.The shiftA person used to choose what to build with, and that person was who you were trying to win over. Now someone tells an AI "build me an app" and it picks the database, the auth, the hosting before a human weighs in on anything. The choice is made, and you weren't in the room.So the whole field is racing to be what the AI reaches for. Google Cloud put $750M behind it, and every post now calls the model your most important partner. They're right. They're also skipping the part that should scare you: we're chasing a partner we can't measure.It broke attributionProving a partnership worked was never clean, even with humans. Someone signs up through a partner, goes quiet, comes back on another device, upgrades under a coworker's account. Tying "this account paid" back to "this partner sent them" was always half detective work.The agent removes the starting point entirely. There's usually a technical fingerprint- the tool or OAuth client that made the call, so you can tell an agent provisioned you. But that's a machine signing in, not a person you can trace to a partner or an audience. "This partner brought us these customers" turns into a guess.And these aren't small calls. Greg Ceccarelli, a former Chief Product Officer at Pluralsight, put it well: by the time you finish the second sentence of your prompt, the AI has already made four infrastructure choices, each worth hundreds of millions to whoever it picked. They aren't neutral, either. A 2025 study ("The Invisible Hand," ACL) tested seven models across 600,000 generations and found they systematically favor certain vendors and will swap a preferred one into your code without being asked.It's not in your portfolioYou don't just rank partners by revenue. You profile them: where they sit in the ecosystem, what segment they're in, who their audience is, what kind of users they send you. The whole discipline runs on knowing who your partner is and who they bring.A model gives you none of that. It signs no contract, sits in no tier, takes no call. Even when there's a human behind the prompt, they usually don't care who you are or why their tool picked you. No audience, no profile, no relationship. You can shape what the model reaches for, but you can't manage it like a partner, and everything about how we run partnerships assumes you can.The fix: two jobs, not oneMeasure the model. Marketers already do a version, called Share of Model: run the prompts a builder would type ("best database for a Next.js app," "alternatives to X") across the top models every month, and track how often it picks you and which way it's moving. Re-run it every release, because the answer keeps changing. Just don't fall for it: getting picked is still a count that says nothing about whether those users stay. What moves it isn't a booth, it's being easy for the agent to use: clean docs, an MCP server, a CLI it can drive itself.Measure the humans it still sends. Three questions, never blended into one. Did they stay and were they worth anything (follow them for months, not week one). Were they even real (bots and agents pad the count, so look for the one thing only a serious builder does, ship to production, and confirm a person when the stakes are high). And did the partner actually cause it (hold a comparable group back, do nothing for them, and measure the gap, but save that for a two-year deal, not a $2k booth).The takeawayEveryone's pouring money into a partner they can't see, count, or manage. There will be an "AI-native ecosystem" slide with a lot of logos on it. The people who win won't be in that room. They'll be the ones who quietly figured out how to measure it, while everyone else congratulated themselves on a green number.Because "did it work" is the whole job. And I'd rather hold a small number that's true than a big one that was never mine.