Find Your Strikethrough: The Agent Shortlist Panel

For years the worst line in a pipeline review was a deal marked closed-lost with no reason attached. There's a worse one now. The deal that never opened, because an agent crossed you out before the buyer learned your name.

Jason Lemkin published one this week, from the buyer's side.

The short version

Jason Lemkin asked his agent where to move hosting for SaaStr AI Connect. The agent wrote the requirements from the codebase, compared five options, picked Render, and struck through Vercel and Netlify on one criterion. Lemkin had barely heard of Render. In June I published a test for this exposure: ask a general model to describe your product and fix what it gets wrong. His post shows the hole in it. The agent was never asked about a brand. It was asked about a workload. The replacement is a shortlist panel: five to ten workload questions with no brand names in them, run every two weeks across the agents your buyers use, logging whether you were named, where you ranked, the sentence attached to you, and the phrase that eliminated you. A wrong phrase is a docs ticket, a true one is a roadmap item, and "never shortlisted" becomes a row in win/loss.

What the agent did

The post is on SaaStr under Lemkin's name. He asked the agent that works on SaaStr AI Connect a plain question: if we moved hosting, where should we go?

Before it named anyone, the agent listed what the app needs. An always-on web server. An always-on background worker. Long-running jobs, since their email builds take minutes. Logs kept for days. Deploys without downtime, on sized machines. Lemkin says he didn't write that list. The agent did, from the codebase.

Then it compared five options and put its choice at the top, with "my pick" next to Render in the table. Railway came second, judged a bit less mature for heavy production. Fly.io was more hands-on. The big clouds were overkill for now. Vercel and Netlify were struck through as "Not suitable," because they're built for short web requests and this app runs minutes-long jobs.

Lemkin had used Vercel many times and assumed it would be the default. He had barely heard of Render. Render gets tested first.

Two sentences in the post are worth more than the rest. He hasn't checked whether the belief that eliminated Vercel is the full picture of what Vercel does today, and for this purchase it doesn't matter, because the agent made the shortlist. And when a product gets filtered out that way, the vendor never finds out. No lost deal in the CRM, no closed-lost reason to review.

His recap of the same week's 20VC episode adds a number. Resend was his first agentic purchase, on his agent's recommendation, and by the figures he had on the show its MCP calls went from 106,000 in April to 3 million in September.

The hole in my own test

I wrote about this exposure in June. In AI Evaluators Are Your New PM Audience and again in Downside Exposure, Extended: When the Model Is Wrong About You, the practice I gave was this: pick your most important product line, ask a general model, cold, to describe it and say who it competes with, and treat the gap between the answer and the truth as your first ticket.

That test checks whether the model knows you. Lemkin's post shows the elimination happens somewhere my test never looks.

Nobody asked the agent to describe Vercel. Nobody asked it about any vendor. It built a requirement list from a codebase and matched products against the list. A model could hold a perfectly accurate description of your product and still cross you out on line three of a list you never saw.

There's a second difference. My test used a cold model. His agent was sitting in the repository. It knew the worker process existed and that log retention was a pain point. The question a buyer's agent answers is specific in a way a cold prompt isn't.

So the describe-me test stays, as a floor. It isn't the instrument.

The drop: a shortlist panel

Lemkin's post ends with three moves for vendors, and the middle one gave this piece its title: find your strikethrough. His 20VC recap has the cadence. Find ten people you trust who have agents in production, ask each one every two weeks to have their agent recommend a vendor in your category, and write down where you rank. He says he doesn't know of a product that does this yet.

It doesn't need a product. It needs a sheet and an owner. Here's the sheet.

And the steps.

1. Write the questions from your last ten wins. For each one, the customer's workload in a sentence, the way they'd say it to an agent. No vendor names, including yours. Keep five to ten.
2. Pick the agents. The general assistants and the coding agents your buyers use. Lemkin lists Claude, ChatGPT, Gemini, and the coding agents. For the runs with a codebase, borrow his panel of ten people, since those are the runs that look like his.
3. Run it every two weeks and fill every column. Copy the sentence and the phrase exactly. The wording is the data.
4. Sort the eliminating phrases. If the phrase is wrong or out of date, it's a docs ticket against the public page the belief would have come from. If it's true, it's a roadmap item, and the phrase is its title.
5. Add a row to win/loss: never shortlisted. My Win/Loss Analysis Agent reads won and lost deals from the CRM every two weeks and pulls seven categories of insight out of them. None of the seven can see a deal that never opened. The panel is the eighth input, and it's the only one that sees this loss before revenue does.
6. Log the date of every fix. I don't know how long a corrected page takes to change an agent's answer, and I'd distrust anyone who quotes a number. Write down when you shipped the fix and watch the next runs.

A caution on reading it. There's no benchmark for a good named rate, and I don't have one of my own to offer. Count yours, then count it again in two weeks.

Legible quality

The same week, Rich Holmes at Department of Product reported Patrick Collison's view that personal agents will act as a kind of structural subsidy for product quality. The reasoning, as Holmes gives it: if an agent researches every purchase, tactics that depend on inattentive buyers work less well, so making the product better becomes the stronger strategy.

I'd hold that next to the Render post, because the Render post is the first worked example I've seen.

Vercel wasn't eliminated for being a worse product. It was eliminated on a belief about long tasks that the buyer didn't verify and didn't need to. So the subsidy goes to quality the agent can read. A product that got better last quarter and didn't say so in a place an agent looks is, for this buyer, the product it was a year ago.

That makes two jobs out of one. Make the product better, and keep the written account of it current for a reader who never opens your homepage. The first has an owner in every company. The second usually doesn't.

This is the argument on AI Product Management in a small case: the build got cheap, and what's left is the part nobody was trained for. Nobody was trained to maintain a product's description for a reader that builds its own requirement list.

This week: write three workload questions from your last three wins and leave your name out. Ask two agents. Write down the phrase next to your name. If your name isn't there, write down whose is.

Related answer: How do you find out if AI agents recommend your product?

Sources: Jason Lemkin, "Our AI Agent Picked Render Over Railway and Vercel. I'd Barely Heard of It. That's How You Get Leads in The Agent Age.," SaaStr, October 12, 2026, and "Jason's Takes on This Week's 20VC," SaaStr, October 11, 2026, both read in full; the Render evaluation and the Resend figures are his. Rich Holmes, "Stripe's CEO says Agents will reward Product Quality," Department of Product, October 11, 2026, for Patrick Collison's view as Holmes reports it.