My Own Homepage Was Overselling What Proof Does

I went back through the homepage FAQ this week and found a line I'd written that wasn't true anymore. Maybe it never was.

It said PodGlue Proof "checks every piece against your voice and brand before anything goes out." That sounds like a real safety net. It wasn't one. Proof didn't run on the server at all until a couple weeks ago, and even now that it does, it doesn't check voice or brand. It checks two specific things: em dashes and a banned-words list.

What was actually wrong

Three things, once I sat down and compared the claim to the code.

"Every piece" wasn't true. The check runs on the newsletter and social posts. Show notes, carousels, and short-form video scripts don't go through it yet.

"Voice and brand" wasn't true either. Nothing in PodGlue checks your writing against your voice profile or brand guidelines. That's a much harder problem, and I haven't built it.

And even for the two things it does check, nothing gates on it. If a piece of copy trips the banned-words list, PodGlue notes it in the background. It doesn't stop the post, and there's nowhere in the app yet where you can see that it happened.

That FAQ answer also gets read by AI search tools as structured data on the page, not just by a person scanning the site. So the overclaim wasn't only misleading a visitor, it was something ChatGPT or Perplexity could repeat back as fact.

What I changed

The line now says what's real. Em dashes get stripped out of everything PodGlue writes for you, no exceptions. And your posts and emails get checked against a banned-words list, the one you build yourself plus a few hundred overused corporate terms loaded by default (I'd been saying "about 150" in a couple of the help docs too, which was also wrong; it's over 300).

I went through the rest of the site while I was in there. "PodGlue Proof" as a named feature only shows up in that one spot. Two help articles described the banned list as a filter, something that removes words automatically. It doesn't. It's an instruction the writing model gets, plus a check afterward that flags what slipped through. I fixed both of those too.

There's one thing I found and didn't fix: the onboarding screen still has a setting called "Em-Dash Grammar" that implies you can choose how the AI handles em dashes. Since the underlying check now strips them unconditionally, that setting doesn't actually change anything anymore. Leaving a wrong sentence on the page is a copy problem. Leaving a control in the product that quietly does nothing is a product decision, and I didn't want to make that call in the same pass as fixing the words around it.

What's next

The honest version of Proof today is narrower than what I was claiming. I'd rather it read small and true than big and wrong. Widening it, checking voice and tone for real, covering more content types, showing you what it caught, is the actual roadmap. Until it's built, I'm not going to describe it as built.

If you use PodGlue, this doesn't change anything you're doing today. It changes what I'm willing to tell you about what's happening behind the scenes.