118 visitors, 0 signups, and the two bugs I only found by digging

Three weeks ago I launched Postmint, which turns one sentence into a finished social graphic in your brand colours. Eight days in: about 118 unique visitors, zero signups that weren't me. Here's the whole autopsy, because the posts that helped me most were the ones with real numbers in them.

Why I built it

My first prototype made marketing graphics straight from a text prompt. I generated twenty, looked at them, and wouldn't have posted one. You know the look. Half-empty canvas, a grey gradient doing nothing, a fake button painted on that leads nowhere, a number the model invented. Technically an image. Also obviously slop, and posting that damages a brand worse than posting nothing does.

So I deleted it and spent my time on the boring part.

The boring part, which turned out to be the interesting part

The model never touches pixels. It writes an HTML document, a real headless browser renders it, and then two things look at the result.

The browser measures. It walks the rendered DOM and reports eight faults it can prove:

• text past the canvas edge,
• text clipped by an ancestor that hides overflow,
• a box too small for its own text,
• text painted over by an opaque shape,
• text overlapping other text,
• contrast under the WCAG minimum against whatever is actually painted behind the glyphs,
• text sitting in a format's reserved strips (the bands Instagram covers with its own UI on a Story),
• word count over a density ceiling. Every one comes back with an element reference and a number attached. Nothing in there is a judgement call.

Then a vision pass looks at the screenshot and grades it against the brief and the reference style. That covers what no measurement reaches: whether the composition is empty, whether the thing reads like the kind of post somebody actually asked for.

Both feed a revise step, and the loop runs a couple of rounds before you see anything. If a render or a model call dies mid-loop it hands back the best version it has so far, so the loop can only ever improve a design that already works.

One thing surprised me. Asking for hex values isn't enough on its own. "Use these colours" gets drifted away from often enough that you can't ship on trust, so the critique pass that grades the layout also checks the palette against the brand's stored values and sends the design back when it wanders. That's a review, and a review is weaker than a deterministic recolour pass. I'd rather say so than claim a pass I haven't written.

The launch, honestly

• ~118 unique visitors over 8 days. Half DE, half US.
• About 2 of them reached the register page. Roughly 1.7%.
• Zero external signups.
• 80% logged as "Direct", because I tagged exactly nothing.
• My Show HN got flagged in 25 minutes. Baity title, AI-fatigued crowd, cold account, and a link straight into a signup wall. Entirely deserved.

The two bugs

The landing page was the first one. Visitors from social bounced in about five seconds, because the first screen explained the product where it should have shown what the product makes. Nobody scrolled far enough to find out.

The second one is worse. My analytics never recorded a single signup: the goal wasn't firing, lost across a redirect. So the one number I actually cared about, signup to first generation, was structurally unmeasurable for the whole launch. I spent that week reading a dashboard that could not have shown me a signup if there had been one.

Which is why I'll push this on anyone about to launch. Register a throwaway account in an incognito window and confirm every funnel event fires. Five minutes. I skipped it and burned my only launch spike.

What changed since

• Fixed the goal, cut a release, verified an incognito signup fires the events end to end. The funnel records now.
• Rebuilt the landing's first screen around output.
• Killed the blank canvas. A new user used to land on an empty brief box and get asked to invent the one sentence they don't yet know the shape of. It's prefilled with an example now, so the first render is a button press.
• Tagged every channel link, so this time I can tell where people came from.

Where I am now

One real signup since, on day 18, someone I never spoke to. They generated an image without anyone walking them through it. That's the entire customer-side data set: n=1, still zero paying.

The gallery is here if you want to poke at it, no signup needed to look: https://postmint.de/ih

And the question I actually want answered. The product is German-first and the buyers I have in mind are small businesses and shops. But the two most engaged audiences in my launch data were from IndieHackers and Product Hunt, English and technical, averaging 16 and 49 minutes a session. Twice now, sampled independently, and across three weeks that audience has converted zero times. Is that real interest worth chasing, or builders admiring the machinery who will never buy? I can't tell from here, and it changes what I build next.