THE SHORT ANSWER
Run a shortlist panel every two weeks. Write five to ten workload questions the way a buyer's agent would get them, with no brand names, and run each through the assistants and coding agents your buyers use. Log four things: whether you were named, where you ranked, the sentence attached to you, and the phrase that eliminated you. Jason Lemkin's October 2026 SaaStr post is the worked example. His agent wrote five hosting requirements from the codebase, compared five options, picked Render, and struck through Vercel on one criterion. He had barely heard of Render. In June I recommended a simpler test, asking a model to describe your product, and his post shows why that falls short: the agent was asked about a workload and no vendor's name was in the question. Sort the eliminating phrases into docs tickets, where the phrase is wrong or stale, and roadmap items, where it is true. Then add a category to win/loss that sales cannot report: never shortlisted.
A buyer whose agent eliminates you doesn't open a deal, so nothing lands in the CRM. The panel exists to make that loss visible while it's still a sentence and not yet a revenue number.
The test I gave in June, asking a model to describe your product, still works as a floor. It checks whether the model knows you. It doesn't check whether you survive a requirement list an agent wrote from someone's codebase.
The sheet, the six steps, and what Patrick Collison's point about product quality leaves out are in Find Your Strikethrough: The Agent Shortlist Panel. The agent that the result feeds is the Win/Loss Analysis Agent.