The Metrics That Sold Me on AI Marketing Are the Ones I Trust Least

I had four browser tabs open, all marketing-automation vendors, all making their pitch to me as I built the content agent that would become Ballad. One tab: 312% ROI within eighteen months. Another: AI-optimized content pulling 5.8% engagement against 2.1% for anything picked by hand. A third: AI-personalized email hitting 34.9% open rates, more than double the static-template number next to it. Every stat was real, sourced, footnoted. And every one of them answered a question I hadn't asked.

I wasn't deciding whether AI beats a human at picking subject lines. I was deciding whether to spend my own hours writing posts or hand the job to software. Those numbers compare AI to some other company's average marketer, running some other company's playbook, on traffic I don't have yet. None of it told me what happens when a technical founder with no marketing background tries this alone.

These are the numbers that sell AI marketing. They're the ones I trust least.

Lift is easy to measure and easy to fake

Take the 5.8% versus 2.1% engagement number. That 2.1% is the baseline for "manually selected content" — which is vendor-speak for someone dragging blog posts into a newsletter with no strategy at all. Nobody thoughtful runs that. So the test isn't AI versus a competent marketer. It's AI versus nobody trying. Same story with the 34.9% versus 22.1% open-rate lift: the loser in that comparison is a "static template," not a founder who actually knows their audience and writes a subject line that means something.

Any A/B test looks like magic when the control group is barely trying. That's not a property of the AI. It's a property of the comparison. If I swap a lazy baseline for a slightly-less-lazy one, the number goes up regardless of whether anything intelligent happened in between.

And the softer stats don't rescue it. 91% adoption tells you the tool is common, not that it works — everyone using something is a herd fact, not a performance fact. 312% ROI within 18 months is a vendor picking the window and the denominator. None of these numbers tell you what happens when the baseline is you, doing it well.

What a founder should actually watch

I stopped optimizing the agent for demo-friendly engagement numbers in 2025 and started asking something plainer: did anything come back? Not likes, not opens. Signups I could trace to a specific post. A reply that turned into a call.

Those numbers are smaller, and they show up slower. Most of the value from content arrives as pipeline, not as a spike. The engagement metrics keep climbing—and that's exactly the problem. They keep climbing even when the thing underneath has stopped working.

The obvious objection: engagement lift is real, so why ignore it? I don't ignore it, I just don't aim at it. Lift is a fine byproduct and a dangerous target. Watch whether the machine runs without you pushing it. Watch whether a real person shows up because of something specific it wrote. That's the only lift that pays rent.

The number I'd want on the dashboard

So what's the metric worth keeping? Something almost embarrassingly plain: did the content go out on schedule, and did a real person who matters to the business show up because of it and come back.

That's it. Not engagement rate, not open rate, not any number an AI system can quietly push up by optimizing for the number itself. Cadence and arrival. Did you publish when you said you would, week after week, without it depending on your willpower. And did the people who arrived look like the people you sell to, not just people who click.

Everything else on the dashboard is a proxy, and proxies are easier to move than to trust. That's the whole problem with this generation of tools. They're built to make the proxy go up.

The numbers that make AI marketing look smart measure the tool. The numbers that make it worth paying for measure whether a stranger showed up and stayed.