01 · The situation
The old site could publish pages. It could not carry the product.
The useful experience crossed too many boundaries for a page builder: member identity, payment state, music libraries, listening sessions, product evidence and a catalogue that changed faster than a small team could research it.
The front door and the product had become different things.
The WordPress site still held search visibility and member access while the new application carried the real experience.
Thousands of products could not be researched one at a time.
Brand, model and evidence coverage had to grow without turning generated guesses into published facts.
A listening impression is not a score.
The product needed to preserve subjective evidence without collapsing it into rankings or a generic “best”.
Each connection carried a different failure mode.
Music services, payments, partners and playback could not be treated as interchangeable API calls.
02 · What changed
The website became a product system rather than a collection of pages.
Rebuild the foundation
The membership experience moved into a dedicated progressive web app with authentication, payments and product data under one operating model.
Connect the listener
Music-service connections, collection data and listening sessions became inputs to the experience rather than separate destinations.
Use AI for translation
Raw listening notes could become a structured profile with per-axis evidence, but only as a draft a reviewer had to confirm.
Use models for breadth
Research agents expanded the product catalogue while publication rules separated discovered material from evidence strong enough to show.
03 · What failed
“Completed” did not mean the catalogue work had completed.
The useful failure
A bulk image job reported success while most batches had silently failed.
The job’s own summary looked healthy. Comparing what was found with what reached the database exposed the loss. Other research runs attached the wrong product image to a valid record. Coverage alone was not a quality measure.
Automation made the catalogue larger. Reconciliation made it trustworthy.Working principle from the product
04 · What we refused to automate
Some attractive features made the product less honest.
The product describes character and fit. It does not pretend taste is a leaderboard.
An unreviewed product says there is not enough evidence instead of displaying a plausible guess.
The model structures the listening evidence. The reviewer confirms every dimension.
A moving hardware control surface made the consequence larger than the convenience.
05 · The result
A platform broad enough to discover with, narrow enough to trust.
06 · What this does not prove
A sophisticated product is not the same as a proven market.
The honest limits
- Current member and monthly-active-user counts have not been confirmed for publication.
- Most reviews currently come from the founders and the platform’s own account.
- Catalogue growth was strong; complete, correct image coverage was not.
- This is an owned product story, not independent client validation.