Correlation is the honest answer in music marketing

A manager asks the only question that matters after a campaign: did it work. The truthful engineering answer is that nobody can tell you exactly, and the reason is structural rather than a gap in the tooling.

When someone hears a track on Spotify, Spotify does not tell your ad platform, your link shortener or your analytics tool that this play came from that advert. There is no passback. The click on your landing page and the stream that may have followed it live in two systems that never exchange identifiers, and for good privacy reasons.

So every claim of the form "this campaign drove 40,000 streams" is a model, not a measurement. Some models are reasonable. None is a receipt.

What you can actually measure

Two things, and they are worth more than a fake receipt.

Direct outcomes inside one system. Clicks on your own landing page, by platform, country and campaign parameter. Ad platform outcomes as the ad platform counts them. Email opens and clicks. Ticket sales through your ticketing partner. Each of these is measured end to end within one system, so each is real.

Co-movement between systems. Take ad spend per day, streams per day and ticket sales per day. Align them on one date axis. Ask how strongly each pair moves together. That is a correlation, and reported carefully it is defensible.

How to do the co-movement properly

Three things have to be right or the number is worse than no number.

Align on dates, and only on dates

Resist the temptation to build a join on anything cleverer. If you find yourself matching a click to a listener because both were in Manchester within an hour, you have invented attribution and given it a technical veneer. Country-level scoping is defensible: comparing clicks from Germany against listener movement in Germany is still date-and-place alignment rather than person matching.

Fix the lag before you correlate

If your streaming series is running two days behind reality and your ad spend is same-day, then every correlation you compute is comparing Tuesday's spend to Sunday's streams. On a short campaign that alone can invert the result. Shift the series by the known reporting lag first, or accept that the answer is meaningless.

This is the most common silent error in the category, because the correlation still produces a confident-looking number.

Require enough overlap

Pearson correlation on eight points will happily return 0.8 for two unrelated series. Set a floor on the number of overlapping days before you report anything, and when you are below it, say so rather than reporting a weaker number.

Phrasing that survives a meeting

The difference between a useful finding and an overclaim is usually one verb.

• Weak and unfalsifiable: "The campaign drove a 22 percent lift in streams."
• Honest and useful: "Spend and streams moved together closely through the campaign window, and streams stayed above the pre-campaign level for nine days after spend stopped."

The second sentence is more persuasive to anyone who knows the data, and it does not collapse when someone asks how you know. We enforce this in generated copy: every correlation finding is phrased as moved together, never as caused, because the model has no way to know the difference and neither do we.

Where correlation earns its keep

Ruling things out. If a market's listener numbers were flat through a campaign that spent heavily there, the campaign did not move that market. Negative findings are cheap, reliable and frequently the most valuable thing in a review.

Sizing. Loose bounds are still bounds. Spend of a few thousand alongside a movement of a few hundred thousand streams tells you something about magnitude even without a causal claim.

Sequencing. Which signal moves first is often more interesting than how strongly they move. In live music particularly, recognition and airplay signals tend to lead ticket sales rather than follow them, and knowing the order of arrival shapes an on-sale plan.

Finding the thing you forgot. A cluster of press coverage explains a week that otherwise looks like an algorithm gift. Correlation across several signals is mostly a memory aid for a busy team.

The one thing to refuse

If a tool tells you a specific number of streams came from a specific ad set, ask how. If the answer involves modelled attribution, fine, ask for the assumptions. If the answer is a confident restatement of the claim, stop trusting the number, because the passback that would make it true does not exist.