The $100 question

dailybedtimestory.com is a side bet inside my kids-games platform: a fully automated pipeline that publishes a new illustrated, narrated bedtime story every night for about $0.30 an episode. A year of evenings went into it. Traffic: effectively zero. Search Console over two weeks: 0 clicks.

So I set a budget and a rule. $100 to answer one question: does anyone actually want this? The kill criterion, written down before spending a shekel: one organic visitor who returns three nights in a week, or three parents saying their kid asked for it again. Neither → the project dies, and $100 is a cheap funeral.

Buying honesty on Mumsnet

Research said Mumsnet’s paid “Non-member requests” board was the only verified channel where you can pay real parents to test a product — £30, and the replies are famously unvarnished. We posted honestly (the AI disclosure in paragraph three, because they will find out), attached two screenshots, offered a £25 voucher draw, and waited.

Three days of silence, one bump later: 33 replies in about a day. Reader, they did not like it.

The roast, sorted into piles

The trick with 33 hostile replies is refusing to read them as one signal. They split cleanly:

People who actually pressed play found real defects. The narration was “too flat” next to the voice actors on Calm and CBeebies — every single real tester said some version of this first. Audio died when the phone screen slept. A line repeated between scenes. The karaoke text drifted out of sync on older episodes (a known debt item — they found it in one evening). One person called the stories “AI word vomit.” Harsh; directionally useful.

People who would never press play objected to the category: support human authors, screens before bed are wrong, bedtime stories are my moment with my child. Several noted the board had seen a wave of AI-app launches lately — we’d paid £30 to walk into a pre-existing backlash. Nothing we fix converts this group.

And the metric that mattered? Zero “my kid asked for it again.” Zero returners. The kill criterion was heading for a clean miss.

Four reasons to quit

That evening I had four arguments for breaking the project, and they were good ones: the anti-AI segment of the audience is real and permanent; kids apparently want the same story repeated, not a new one nightly; nobody wants a screen at bedtime; and the entry bar is on the floor — anyone can ask a chatbot for a bedtime story.

I’d also lost faith in the original growth thesis. The whole plan was “generate content daily → Google long-tail → organic users.” In 2026, with AI answers eating informational queries, that plan is a field being paved over. If Google doesn’t buy content volume, why generate content?

The dashboard I almost didn’t open

Before writing the post-mortem, I opened Apple Podcasts Connect for the show’s RSS feed — a thing we’d shipped three weeks earlier and never promoted. Not once. No launch, no links, nothing.

98 plays. Seven listeners in the United States (I’m not in the United States). Two hours of listening time — and the split that made me sit up: 100% of it came from people who follow the show. Somewhere out there, a handful of American families had found a zero-follower podcast, subscribed, and were playing episode after episode. Flat robot voice and all.

That’s the exact user the $100 was hunting. They existed the whole time — in a different room.

Spotify said zero. That was the clue.

The same show on Spotify: 1 all-time play. Same episodes, same feed. The difference is how each platform does discovery. Spotify is a recommendation engine; a zero-follower show is invisible until you bring your own momentum. Apple still has a working search box — and “daily bedtime story” is a literal query match.

Which reframes the dead SEO thesis nicely: discovery didn’t die, it moved. Google’s long-tail is gone, but Apple Podcast search is an under-crowded search surface where an exact-match name already ranks. The premise “publish daily, let an algorithm distribute” still works — you just have to feed the algorithms that still distribute.

The same-day pivot

Solo founder plus AI agents means a pivot doesn’t need a quarter. Same day, in order:

• Voice casting. Generated the same 30-second passage in seven voices — including our current narrator on the newer, more expressive TTS model, which isolated a finding: half the “flat” complaint was the cheap model, not the voice. Picked a two-narrator cast: George for dinosaurs, trucks and dogs; Matilda for princesses, rainbows and cats. Same voices, native in Spanish, if that expansion ever unlocks.
• The podcast became the product. Each episode already generated a sung lullaby that only website visitors heard. Now the podcast cut is story → 2.5s crossfade → lullaby → sleep. We built a spoken outro too (“follow the show…”), listened once, and deleted it. Nothing should talk after the story.
• Engineering that respects the constraint. The episode generator runs in a slim container we deliberately keep ffmpeg-free, so the audio stitching happens at deploy time in CI. First live run promptly taught us that GitHub’s current runner images no longer ship ffmpeg — and that a fail-fast guard written as ffmpeg -version | head -1 doesn’t fail fast at all, because the pipe swallows the exit code. The episode gracefully degraded to plain narration, we shipped the two-line fix, and the next deploy self-healed. Degrade-don’t-break earned its keep on day one.

By bedtime, the first new-format episode was live: a dog story, expressive narration, ending in its own lullaby. 4:46 of exactly what those seven American families signed up for.

The gate

The bet stays cheap and the exit stays sharp: ~$0.40 a night to generate, zero hours a week to run, and a four-week gate. If Apple followers become visible and plays trend toward 20+ an episode by late August, the full audio-first version gets built — favorites shelf, human-QC badge, maybe the Spanish twin. If the line stays flat, the project dies with a clean conscience, and the pipeline gets recycled into my other games.

What I’d tell you

Paid forum feedback buys you opinions — loud, honest, useful for finding defects, terrible for measuring demand. Dashboards buy you behavior. When they disagree, believe the behavior: the same week 33 people told me nobody wants this, seven strangers were quietly using it exactly as intended.

And check every distribution surface you own before you write the post-mortem. Your users might not be where you shouted. Mine were asleep by the third scene — which, for once, was the whole point.