Deriving audience segments without any personal data

Segmentation in music marketing usually implies a CRM full of names and email addresses. That is one way to do it, and it makes you responsible for a large amount of other people's personal data.

There is another way that gets most of the value. Aggregate platform data already describes an audience's shape, and you can name and act on the resulting groups without ever holding a record about an individual.

Five segments available from aggregate data

Geographic core markets. Derived from listener geography per city or country, filtered by a minimum share and a minimum absolute size. The output is "your audience in Manchester", sized in listeners, not a list of people.

Platform-native audiences. The audience that lives on one platform and behaves differently there. A large SoundCloud following with modest Spotify numbers is a real segment with its own norms, and it is addressable through that platform's own tools.

Ticket buyers. Derived from aggregate box-office data by show and market. Not who bought, but how many, where and at what pace. Enough to know which markets have converted attention into money.

The opted-in inner circle. Your mailing list and your followers: people who took an explicit action. Sized from aggregate counts held by the email platform, with no need to hold the addresses yourself.

Discovery-stage listeners. People arriving through algorithmic and editorial placement who have not yet converted. Derived from the gap between listener growth and follower growth, which is the same ratio that separates rented from earned attention.

None of those requires a personal record. All five are actionable, because the channels you would use to reach them are also aggregate: a platform's own targeting, a market-level ad campaign, a local press push, a support slot.

What every segment has to carry

A segment with a number and nothing else is a liability, because the reader cannot tell a solid figure from a guess. Ours carry five things.

Size with its unit. 14,200 listeners is a different claim from 14,200 people or 14,200 ticket buyers. The unit is part of the number.

An estimate flag. Some sizes are counted, some are inferred from a share. If a figure was derived rather than read, say so in the data structure rather than in a footnote nobody sees.

Confidence, with a reason. High, medium or low, plus the reason: how many independent sources contributed, and whether the volume is above the floor where the arithmetic is meaningful.

Contributing sources. Which connections produced this segment. It lets a reader judge it and it makes the missing-source case obvious.

A trend direction. Growing, holding, shrinking, or unknown. Unknown is a legitimate and common answer when history is short.

The rules that keep it honest

Never sum the sizes. Segments overlap by design: a Manchester ticket buyer on your mailing list appears in three of them. A total audience figure built by adding segments is wrong, and it is wrong in the direction of flattering you. We state this in the data the AI assistant reads, so it cannot casually produce a total either.

Enforce floors. Below a minimum share and a minimum absolute count, a geographic segment is noise. Ours require both, and the thresholds are named constants rather than magic numbers buried in a query.

Fail open on every source. A project with no ticketing connection should get four segments and an explanation, not an error. Every read in the assembly path degrades to nothing rather than throwing.

Say what is missing. An honest empty state names the signals that are absent and what connecting each one would add. That is more useful than an invented segment and it doubles as onboarding.

One derivation, three surfaces

The design decision that matters most is architectural rather than analytical.

The derivation lives in one pure function with no database access, unit-tested against synthetic inputs. A server module assembles the inputs from each source, and three consumers use the same result: the page a manager reads, the API, and the tool the AI assistant calls.

The alternative, which is what happens when a feature grows organically, is a query behind the page and a slightly different query behind the assistant. Then a manager asks the assistant about a segment they are looking at and gets a different number, and the credibility cost of that is much higher than the engineering cost of avoiding it.

What this cannot do

It cannot tell you who to email, because it holds no addresses. It cannot build a lookalike audience from your own customer list. It cannot do individual-level personalisation.

For a manager deciding where to tour, where to spend, and which platform deserves attention this quarter, it does the job. And it does it without making you the custodian of a database of other people's personal information, which is a trade most small teams should take.