Algorithmic Awareness

Spotify pays royalties per play.

This sounds like a fair model — listeners discover music, artists get paid in proportion. But every algorithmic market eventually reveals the gap between what it's optimizing and what humans think it's optimizing.

Spotify isn't optimizing for music quality. It's optimizing for play counts. And in 2017, an aspiring musician named Michael Smith noticed.

Smith's insight wasn't musical. It was algorithmic.

The system that decided his income didn't listen to his music. It counted plays. If a play was a play — and the algorithm couldn't tell whether the listener was real — then the most direct path to royalties wasn't writing better songs. It was producing more plays.

This is the central move. Look past the surface of the market to the actual reward function underneath.

Step one: build the demand side. Smith bought thousands of email addresses in bulk, used them to create Spotify accounts, and wrote scripts to have those accounts stream his music in rotation. The accounts looked plausible enough — varied IPs, varied listening patterns, varied playlists.

The first bottleneck was supply. He couldn't produce songs fast enough. One song played a billion times would trigger fraud detection. So he needed many songs, each streamed just enough.

Step two: build the supply side. In 2018, Smith partnered with an AI music company to generate thousands of machine-composed tracks. Each one was a believable-but-forgettable song with a fake artist name and a fake bio. Each one was streamed by his bot army a few thousand times — below the fraud-detection threshold.

And then the strangest thing happened: real listeners started streaming them too. Spotify's recommendation algorithm — seeing all this play activity — started suggesting his songs to humans. Real humans then unknowingly added to his royalty pile.

At peak, Smith's network was generating 661,440 fake streams per day. Royalties accumulated. By the time the FBI charged him with wire fraud in 2024, he had collected over $10 million.

It's tempting to call this just fraud and move on. But the move he discovered isn't fraud-specific. It's a recipe with three ingredients: identify the reward function, manufacture supply that satisfies it, manufacture demand that triggers it. Smith's version was illegal. The same recipe is being executed legally, at scale, across every algorithmic market in 2026.

SEO content farms — produce thousands of articles optimized for whatever Google rewards, not for what humans need to know. TikTok creators who study the algorithm more than their craft and out-perform actual filmmakers. LinkedIn posters who optimize for format over substance. Amazon sellers who buy reviews and stuff keywords. AI agents that file thousands of grant applications, all just below the detection threshold.

None of these are fraud. All of them use Smith's move. The leverage isn't in being better. It's in knowing what the algorithm rewards.

The leverage came not from traditional assets but from understanding how digital platforms allocate rewards.

— From Reshuffle, Chapter 7

If your business sits inside an algorithmic market — and if you can't immediately name the market's reward function, you do — then your competitive position is being decided by whoever can name it. That's the new leverage. Algorithmic awareness as a strategic capability.

It's not glamorous. Smith's case is gross. But the version of him operating in your category, today, isn't going to court. They're going to compound.

Sangeet on this in Chapter 7 ↗