Metrics
Which numbers decide whether Bitrate is working, which are diagnostic, and which are not allowed to be reported as evidence of anything.
The measurement problem here is specific: it is very easy to produce impressive-looking numbers in a product with no users. Signups, page views and catalogue size all go up with effort and none of them indicate the business works.
The north star
Artists who complete a release through Bitrate.
Not signups. Not uploads. Not monthly actives. Completed releases.
It qualifies because it is the only number that cannot be moved without the product actually working. A completed release means an artist trusted Bitrate with real work, the workflow held together end to end, and something reached the outside world. Everything the vision claims is either true or false at that moment.
Its natural companion, and the one that decides whether there is a business at all:
Artists who complete a second release.
One release proves curiosity. Two proves the product worked. Watch the ratio between them more closely than either number alone.
The funnel
The steps that matter are c → d and d → e. Everything above c is marketing and everything below f is a consequence; the two conversions in the middle are the product.
| Stage | Question | Failure means |
|---|---|---|
| Reached → signed up | Is the positioning legible? | The message is wrong, not the product |
| Signed up → activated | Does the first step feel worth taking? | Onboarding, or the wrong audience |
| Activated → completed | Does the workflow hold? | The core product does not work |
| Completed → returned | Was it worth doing? | It is a tool, not a workflow — the model is wrong |
| Returned → paid | Is the value worth the price? | Pricing, or insufficient value |
| Paid → referred | Is it good enough to stake a reputation on? | Adequate, not loved |
Metrics by phase
Different phases have different honest metrics. Applying phase 5's metrics in phase 1 produces noise and false confidence.
| Phase | The metric | What is explicitly not the metric |
|---|---|---|
| 1 — Validation | Interviews completed; problems that recur across them | Anything quantitative — the sample is far too small |
| 2 — MVP definition | Scope stability | Features specified |
| 3 — Launch prep | Can one artist complete a release end to end? | Readiness percentages |
| 4 — First market | Completed releases; second releases; first revenue | Signups, page views, catalogue size |
| 5 — PMF | Cohort retention; ARPU; CAC; payback | Total registered users |
| 6 — Scale | Growth rate; margin; listener retention | Anything that looks good in isolation |
Numbers that are not allowed as evidence
Each of these has been used by someone to prove a product was working when it was not:
- Signups. Free and uncorrelated with value.
- Registered users, cumulative. Only ever goes up, including while the product dies.
- Page views and sessions. Measure curiosity.
- Catalogue size. Bitrate can seed it. It says nothing about demand.
- Total streams, without unique listeners behind them.
- Uploads, without completed releases behind them.
- GitHub stars. Interest from developers, who are not the customer.
- Test coverage percentage. A diagnostic, never a target.
PRODUCT.md already commits to this discipline: "Claim nothing the product cannot back."
That is a rule about metrics as much as about copy.
What can be measured today
Honestly: almost none of the above, and that is worth stating rather than working around.
There are no users, and the analytics substrate does not exist — no event taxonomy, no product analytics, no warehouse, and metrics counters that nothing scrapes. The listening records that do exist capture user, track and timestamp, and not how much was listened, from where, or on what device — which is too thin to answer most of the questions above.
That is Stage 3 of the tech roadmap, and it is a prerequisite for phase 4, not for phase 1. Phase 1 needs interviews, not instrumentation.
The exception, computable today with no users at all, is cost to serve one artist for one month. It should be computed now.
Instrumentation principles
For when the event layer is built, so it is built once:
- Define the taxonomy before emitting anything. A named, versioned schema. Events invented ad hoc at call sites become unusable within months.
- Name events after what happened, not what was clicked.
release.completed, notsubmit_button_clicked. - Every event carries the actor, the subject, and the context. Missing context is what makes historical data unanswerable later.
- Instrument the funnel above before instrumenting anything else. Six transitions are worth more than a hundred incidental events.
- Personal data in events is a legal decision, not a technical one. Behavioural profiling is what triggers a DPIA.
Reporting discipline
A number reported without its denominator, its time window, and its sample size is not a finding. At the scale of the first fifty artists, individual cases carry more information than any aggregate does — so read the cases, and use the aggregate to know which cases to read.