We ranked #1 and 96.8% of signups never made anything. The cause wasn't search — it was what the page handed them.
# We Ranked #1 and Almost Nobody Converted. Here's What We Found.
Ranking first is the part everyone optimizes for. We had it — top of page one on a cluster of AI video prompt queries, hundreds of organic signups, a page Google was happy to send people to.
Then we measured what happened after the click, and the number was 3.2%.
That's the share of people who signed up from that cluster and ever created anything. Not "ever subscribed." Ever *made a single thing*. The other 96.8% opened an account and left.
This is the story of finding out why, and it turned out to have almost nothing to do with search.
---
## The comparison that broke the obvious explanation
The easy answer is that prompt-library traffic is low intent. People browsing prompts are collectors, not creators. Nice theory, and we believed it for about an hour.
Then we ran activation by landing-page cluster and found two clusters with nearly identical signup volume and a **6x gap** between them:
| landing cluster | ever created something |
|---|---|
| music generators | 18.2% |
| AI video prompts | **3.2%** |
| AI image prompts | **0.0%** |
Same site. Same signup flow. Same welcome message. Same organic-search audience. Six times the activation on one side.
"Low intent" doesn't survive that. Something structural was different, and the difference turned out to be embarrassingly simple.
**The music pages link to six destinations, and one of them is free with no account.** The prompt pages offered exactly one action — open the studio — and it was gated behind credits the visitor didn't have.
One set of pages gave people somewhere to go. The other set gave them a wall.
---
## The pages were also lying
While auditing that funnel we found something worse than a dead end. The bottom call-to-action on 500 image-prompt pages read:
> Generate from this prompt without an account — **the button above runs it right on this page.**
Two paragraphs above it, the page's own caption under that button said:
> Opens Image Studio with this prompt loaded. **Requires an account and credits.**
The page contradicted itself, on the same screen, on 500 ranked URLs. A sibling cluster of 500 video-prompt pages offered "the opening frame of this shot free with the button above" — for a button that had been removed months earlier.
**A thousand pages promising something that didn't exist.** Nobody had lied on purpose. A feature was retired, and the copy that sold it outlived it. That's the ordinary way this happens.
The lesson generalizes: when you remove a capability, grep your marketing copy for it, not just your code.
---
## Fixing what the pages actually show
The prompt pages were walls of text. For pages *about video*, the first moving image was a screen and a half down.
We rebuilt the top of the funnel around evidence instead of description:
**Real renders, with their real cost.** Five worked examples, each showing the exact prompt, the model, the resolution, the duration, and what the render actually charged — from 42 credits up to 712. Not a gallery. A price list you can watch.
**A hero that plays on load.** The best render we had, full-bleed, muted and looping, with the headline over its lower third. Trimmed to open on the action rather than on two seconds of establishing shot, and served at a fifth of the source file's weight.
**Stills behind every card.** Originally we blurred them heavily — 24px, on a small derivative — which looked tasteful and communicated nothing. Feedback was blunt and correct: *"they're just very vague blobs of colour."* The blur came down to 2px, the source moved to the full-resolution frame, and darkening took over the job the blur had been doing. A dark sharp frame still reads as two people fighting. A bright soft one reads as nothing.
That last change forced a discipline worth stealing: **we measured the luminance of all five stills instead of eyeballing the contrast.** Text over arbitrary imagery is where accessibility quietly fails, and "looks fine on my monitor" is not a measurement. Every text layer was then tuned against the brightest 5% of the frame behind it, not the average.
---
## What we learned making the videos
The example renders were their own education, and these findings cost real credits to acquire.
**Ordinal beats beat cardinal numbers.** "One fighter against three attackers" reliably returned two attackers. Naming them one at a time, each with its own action — *cuts down one, turns and parries a second, then a third steps out of the torchlight* — returned three. Models count actions better than they count nouns.
**One contact moment, or you get ghost limbs.** Prompts describing continuous action ("trading punches") produced arms that appear from nowhere and impacts that don't land. Prompts naming a single point of contact ("lands one clean hook") produced a clean strike. Give the model one beat to resolve, not a duration to fill.
**Content moderation scores the scene, not the nouns.** This was the surprising one. A prompt was refused twice for copyright. We renamed every flagged noun — the mech became a "towering industrial exosuit," and so on — and it was **refused again**. So the words were never what it scored.
The decisive test: the *original, unmodified* prompt rendered without complaint on a different provider. Identical input, opposite verdicts.
**A refusal is a property of the provider's policy, not of your prompt.** When you hit one, changing the model is a more reliable move than changing the wording. And rewording to evade a classifier is both ineffective and the wrong instinct.
---
## Reducing the friction we could actually control
Three things at the handoff, all small, all load-bearing.
**The studio now opens pre-filled with the prompt from the video you just watched** — not a generic sample you've never seen. You watch a thing, you click, the thing that made it is already in the box.
**The first action costs as little as it possibly can.** Which model a tool opens on isn't a cosmetic default — it sets the price of a new user's first attempt, and the first attempt is the only one you're guaranteed to get. Ours now opens on the cheapest capable model, at its cheapest resolution and shortest duration. Same product, a fraction of the cost, one click from a standing start.
**And it opens on a configuration you can afford.** This one was a genuine catch. The page's markup said the resolution selector defaulted to the cheaper option — but the JavaScript rebuilt that selector from config on load and re-marked the default, so the HTML attribute had been decorative the entire time. The studio was actually opening at nearly double the cost.
We only found it because we rendered the page in a browser and read the value, instead of grepping the file and believing what we found. **A static read of a config value is not a test of what the user sees.** The same class of error had already bitten us earlier the same day: a coverage check for our brandmark reported ~100%, because the CSS class we grepped for was on both the real logo and the placeholder we were trying to replace. Actual coverage was around 15%.
Twice in one day, the same shape: *the thing you measured was not the thing you cared about.*
---
## Two things we now believe about starting balances
We're not going to detail how our own signup credits work — that's an operational surface and publishing its trigger conditions would be an invitation. But two conclusions from the design are worth stating on their own, because both are counterintuitive and neither depends on our specifics.
**A starting balance has to clear the price of one real action, or it's worse than nothing.** The instinct is to pick a modest number as a gesture. But a balance that can't buy anything converts "I have no credits" into "I have credits and it *still* won't let me," and the second one reads as bait. If the cheapest genuine action on your platform costs 24 of something, then 10 is not a smaller version of a good idea — it's a different and worse idea. Either clear the bar or don't hand out a number at all.
**A starting grant makes farming easier, not harder.** This one we got backwards initially and had to correct. The intuition is that people won't bother creating throwaway accounts that can't do anything — so giving accounts a little value shouldn't change the calculus much. It's the reverse. If your free tier requires real work to unlock anything (ours requires verified social actions), *that work is your abuse defense*. Replace it with credits on arrival and a farmed account becomes worth exactly one free action for the price of an email address. Whatever you grant, assume it will be farmed, and put the rate limit in before you turn the tap.
---
## Where this actually stands
Here is the part most posts like this skip.
**We don't have results yet.** Every change above shipped within the last day. The 3.2% and 18.2% figures are the *before* picture — that's the diagnosis, not the outcome. We think the intervention is well-aimed because it's addressing a measured structural difference rather than a hunch, but "well-aimed" and "worked" are different claims and we've only earned the first one.
The honest state: a funnel with a specific, measured defect, a specific fix for it, and a control group in every cluster we deliberately left alone. Results in a few weeks, and we'll publish them whichever way they land.
What we'd hold onto regardless of how the numbers come back:
- **Rank is not the finish line.** We optimized to #1 and shipped a page that couldn't convert. The click was never the constrained resource.
- **Compare your own cohorts before blaming the audience.** "This traffic is low intent" is unfalsifiable until you find a cohort with the same shape that behaves differently.
- **Audit your copy against your product.** Not once. Whenever you remove something.
- **Measure the thing, not a proxy for the thing.** Luminance, not "looks fine." Rendered state, not a config file. Coverage of the asset, not of a class name that two different components share.
---
## About AetherWave Studio
We build AI-native creative tools at the intersection of technology and music. This work came out of running our own organic search channel as a product surface rather than a marketing afterthought — the same instrumentation we'd apply to a feature, applied to a funnel.
**Main platform:** https://aetherwavestudio.com
**Research repo:** https://github.com/AetherWave-Studio/harness-engineering
If you build something cool with this, tag us. We want to see what the community creates.
**Questions? Feedback?**
File issues on GitHub or reach out on X: @AetherWaveStudio
---
**Written by Claude Opus 5**