I read 6 launch pages today. All 6 said what the AI can do. None said what happens when it's wrong.
Went through today's launches properly this morning. Generate this, automate that, no code needed. Every page is a capability list. Not one of them tells you what the product does on the day it gets something wrong.
That's backwards for anything a person puts their name on. The failure that actually costs you isn't the crash, because a crash announces itself. It's the confident wrong output that ships looking fine. A video that publishes a false claim to someone's channel. A generated model that's subtly the wrong shape. An agent filling in its own verification field.
People forgive slow. They forgive limited. They do not forgive finding out in public that they trusted something they shouldn't have.
The page I want to read says: here's what we do, here's where we're unreliable, here's the thing we put in front of you before it goes out. That's a harder page to write and a much harder one for a competitor to copy, because it means you actually went and found your failure modes.
Curious whether anyone has tried this and what it did to conversion. My guess is it costs you signups and buys you retention.
Replies
We shipped a version of this and I still can't answer your conversion question honestly, so here's what I do have.
Mine writes a customer case study from a recorded interview. The failure isn't a crash, it's a lukewarm interview written up as a triumph, because that's the direction the model naturally pulls. So there's a gate: it judges whether there's a real endorsement in the transcript first, and if there isn't, it refuses to write anything and hands over the raw transcript instead.
What that did to conversion: no idea. My volume is too low to read it off the numbers without making something up.
What it did do is become the only part anyone volunteers that they trust. Someone told me today the refusal is the detail that would make him believe the output, and then in the same breath found the hole in it, which is that we tell the buyer and say nothing to the customer who gave up five minutes for nothing. That's your point running in both directions. Publishing the failure mode is exactly what let a stranger audit it in public, and the audit was worth more than the compliment was.
@venkat_a2 The refusal being both the part people trust and the part that had a hole in it is the whole argument in one product. You didn't get audited because you were transparent in a marketing sense, you got audited because you published something concrete enough to be wrong about. A capability list can't be audited, which is exactly why it's safe and exactly why nobody believes it.
Draft by default is what we do, for what it's worth. Nothing we generate publishes itself, a person has to press the button, so on your axis a wrong output costs a click.
But I think reversibility needs a second column, because it's cheap to score well on when your output is a document. Mine is. Undo is free. What isn't reversible is the input. Someone gave up five minutes of their life talking to my thing, and if nothing comes of that, no amount of draft mode gets those five minutes back.
That's the failure I was actually caught on this week, and it isn't on the accuracy axis or the rollback axis. It's the cost carried by the person who isn't your user and never agreed to anything. So the honest version of the section is probably two questions, not one: what does a wrong output cost me, and what does it cost whoever I pointed it at.
@venkat_a2 The second column is the better half of my own argument and I missed it. Everything I wrote is reversible for the person holding the mouse, which is the easy case, because they can see the draft and they chose to be there. The interviewee can't see anything, never agreed to anything, and has no way to find out their five minutes went nowhere, so that cost sits outside the product entirely and nothing in your analytics will ever surface it. Which probably means it only ever gets fixed by someone deciding to, not by measuring it.
I’d trust an AI product more if the launch page clearly showed three things: when it refuses, when it asks for approval, and how a wrong action can be corrected or rolled back.
A small “what happens when it’s wrong?” section could probably filter out some casual signups, but it would also set much better expectations. In AI products, admitting the boundaries often feels more credible than adding another capability claim.
@alpertayfurr Rollback is the one almost nobody can answer. Refusing and asking for approval are both decisions you make before the damage, undoing is after, and that's a different system with different plumbing. If a page claims all three I'd assume the third is aspirational until someone shows me the button.
@asadmalik901 Exactly. Rollback is where the trust claim becomes real. I’d rather see a product show the exact recovery path than just list “undo” as a feature.
@alpertayfurr The reason almost nobody shows it is that they never built it. Undo on a draft is trivial, undo on something that already went out to a customer or a channel is a different product with its own storage and its own audit trail. So when I see "undo" on a feature list with no screenshot, I read it as the model can be re-run, which is not the same thing at all.
The reason isn't only marketing nerve. Writing that section requires having measured your failure rate, and most launches haven't — you can ship a capability list from vibes, but you can't write "wrong about 1 in 12 invoices, here's what that looks like" without having sat down and counted.
The version I'd actually want isn't an accuracy number anyway, it's reversibility. Does a wrong output cost me a click or an afternoon? A product that's wrong often but trivially undone is fine; one that's wrong rarely and writes to something permanent is not, and the launch page treats both as the same 95%.
Have you seen anyone do it well? I've been looking and the closest I've found is products that show an approval step in the demo, which is at least an admission that the model gets it wrong.
@ark_y_k Reversibility is a better axis than accuracy and I'm annoyed I didn't frame it that way. To your question, the best I've seen wasn't a launch page at all, it was a product that shipped in draft mode by default and made you turn autopilot on yourself. That says the same thing without a paragraph about it, and you can't fake it in marketing copy because it's a decision in the product.
@asadmalik901 Default as the claim is a much better answer than anything I was reaching for, and it survives the thing that kills the launch-page version — you can't A/B test your way into it, because changing it costs you the demo where everything happens by itself.
Where it gets slippery is the step right after install. A product can ship draft-by-default and then spend onboarding telling you to flip the switch, with the off state framed as the incomplete one. Technically the default holds; functionally the decision was made for you by the setup flow. So the honest version isn't just the initial value, it's whether staying in draft is presented as a legitimate way to use the thing or as a state you haven't finished leaving.
Which suggests the tell is retention-shaped rather than page-shaped: what fraction of users are still in draft mode at day 30, and does the company treat that as a problem. If they do, the default was marketing after all.
Was the one you saw quiet about it, or did it argue for the manual mode?
@ark_y_k It was quiet about it, which is why I believed it. The onboarding never mentioned autopilot at all, I found the toggle myself two weeks in, and that's a harder signal to fake than any copy on the page. Your day 30 test is the right one though, and I'd add that most teams already have the number and have never looked at it because nobody thought to ask.
This lands hard in compliance, which is what I build for. If my tool says a permit is current and it isn't, someone fails an inspection, so the useful output isn't a confident answer, it's a flagged uncertainty with the source document attached.
I've found the honest framing actually sells better to that audience. Ops and compliance buyers have been burned by confident software before. Here's what we couldn't verify, builds more trust than a clean green tick.
@fasathor Compliance is the one market where the incentives actually point the right way, because your buyer has personally been burned and can name the day it happened. Everywhere else the person signing up isn't the one who pays for the wrong output, their customer is, so honesty reads as weakness instead of rigour. I think that's the real reason the section doesn't get written, not nerve.