It seems it got some new math too, because one of my songs is seeing a historical stream lift of 86,607%. Impressed? You shouldn’t be.

More on that later. First, from the new summary page:

Discovery Mode is a no-brainer, right? From the looks of it, I made about three times as much in royalties as I would have on the 5 songs in my campaign.

But wait, there are 29 songs in my campaign! What gives?

Spotify only counts the ones that got at least 500 streams the month before, which tend to be the winners. Pretty sneaky sis!

Those 5 songs accounted for about two-thirds of my Discovery Mode streams. The commission, though, came off all 29.

If you thought their commission calculation was deceptive, wait until you see how they calculated my historical stream lift!

Back to that 86,607% figure:

Historical lift compares last month against the 28 days before the song was enrolled in Discovery Mode.

“Where Tigers Are Said to Roam” and “The Rehearsal” were both released as singles before the album came out. I opted them into Discovery Mode as soon as they were eligible. A few months after the album release, I removed the singles and left the album versions.

Here’s where things get messy. The album didn’t exist during the 28 days before the songs were first enrolled, so Spotify is comparing last month’s streams to zero streams. Since you can’t divide by zero, Spotify swaps in a 1.

So in September, “Tigers” got 929 streams and “The Rehearsal” got 691 in 30 days. Spotify is comparing them to a baseline of 1 stream in 28 days. That’s how you get 86,607% and 64,392%.

The fix, if I cared enough to fix it, would be to leave it out of Discovery Mode for a month and then re-enroll it. Then it would take the stream and listener counts from the unenrolled month as the new baseline.

Many of the 29 songs in my campaign have come in and out of Discovery Mode, so they’ve adopted new baselines.

As for the 91% historical stream lift for the entire campaign, when you remove the broken songs, it comes in at 58%. Put them back in with the baselines they had before I removed the singles, and it’s roughly zero.

tl;dr Historical stream lift falls somewhere between useless and bogus.

My advice: ignore it.

One more item of business: which songs to enroll.

Spotify only recommends opting in 7 of my 26 eligible tracks for November. It only recommends tracks that it thinks will get more streams in Discovery Mode.

I almost always opt in everything. When I tried sticking to their recommendations, my Discovery Mode streams dropped precipitously, so I reverted.

Until today, I thought that Discovery Mode streams were bonus streams that I otherwise wouldn’t get.

I didn’t mind paying a 30% commission, because I was paying far more than that for Meta ads. Who cares about a third of a third of a cent for a stream compared to $0.20 for a click, with no guarantee of an actual stream, on Instagram?

But Discovery Mode streams aren’t freebies. They include all the streams those songs would have gotten from Radio, Autoplay, and Mixes anyway.

So when I dropped the songs that weren’t recommended, of course my Discovery Mode streams went down. Those songs were no longer in the Discovery Mode report, but that doesn’t mean they stopped getting algorithmic streams.

With the 30% commission, a song needs to get 43% lift just to break even on royalties. My recommended songs are well past that, except for “Time to Kill,” which is new this month.

The recommendation predicts more streams, not necessarily more money. For the other 19 songs, Spotify doesn’t even give me a lift estimate, so I would be paying 30% on faith.

So I’m running a test. I’m only enrolling the 7 recommended songs for November. When the campaign ends, I’ll compare the other 19 songs’ Radio, Autoplay, and Mixes streams to the previous month. If they are in the same ballpark, I would have been paying 30% for nothing.

I’ll report back in December’s email newsletter, so subscribe if you haven’t already! I’m not selling anything and I only email at the end of the month or when I’ve got a new post.

What’s your Discovery Mode strategy? Am I getting it wrong? Set me straight in the comments!

If so, my new website at colortheorymusic.com will feel instantly familiar.

Color Theory XP features highlights from the Color Theory catalog disguised as a Windows XP desktop, complete with boot screen, login, taskbar, Start menu, draggable windows, and fake apps that play real music.

A month ago, I wouldn’t have thought that this was even possible, much less that I could do it myself.

I suppose that’s overstating things. I didn’t actually do it myself. I asked Claude to do it, and it was up and running in an afternoon.

Here is the actual prompt I gave it:

I, Brian Hazard, have a recording project called Color Theory and I want to build a new website that I’ll host at colortheorymusic.com, leaving colortheory.com to host my email. For the purposes of this task, pretend that colortheory.com doesn’t exist and just pull images and info online.

I want you to build a website that looks like a Windows XP operating system, with a desktop and file navigation and apps to click on. Make it fun with lots of easter eggs. Try to tastefully encourage people to join my Patreon page and include WinAmp that can play my songs. Also include Cubase VST that can load and play my songs.

Bada boom, bada bing, done.

I figured it would need PHP and a database and a special hosting package, but to my surprise, it didn’t require hosting at all. It’s just HTML, CSS, and JavaScript that runs in the visitor’s browser.

Cloudflare, where I host my domains, has a Workers & Pages feature where you simply upload the files Claude spits out (a “deployment”), and they’re immediately live on the web with no monthly fee.

So I put it up and shared it, and it started spreading on social media! To capitalize on the momentum, I updated it to include my new single, “Couldn’t It Be Me?“

I rendered stems from Ableton so that they could be muted and soloed on a per-track basis in Cubase VST. Getting that to work took longer than everything else up to that point.

Ten separate audio files won’t stay in sync in a browser, so Claude made it download all of them up front, decode them into memory, and start them on one shared clock. On mobile it collapses the stems to mono at a lower sample rate to fit in memory.

As for traffic, Cloudflare reports 5,440 unique visitors in 30 days mostly from the US, UK, Netherlands, Brazil and Canada.

Is it an effective way to promote music? Who knows. Was it fun? Yes.

If you’re paying $20 a month for a Claude or ChatGPT subscription, you can do it too, though you might have to space out your sessions. You genuinely don’t need to know anything about coding or web design. Just tell it what you want in plain language.

My little demonstration video only scratches the surface. Feel free to play around to your heart’s content!

I had Claude prepare a full feature list, including every easter egg, in case you’d rather read than explore:

If you’ve got any ideas for what I should do next, let me know in the comments!

ICYMI, The Atlantic created an AI Watchdog search engine where you can check if your music was used to train AI models, for which you were almost certainly not informed or compensated.

In my case, 181 songs are in two datasets: LAION-DISCO (above) and SLEEPING-DISCO:

I should be really angry then, right?

Maybe I should be, but I’m like a 4 out of 10 tops.

For starters, I already knew my music was trained on. There’s a music AI service called Sonauto (now Treblo) that allowed you to prompt for specific artists until recently, most likely due to legal challenges.

When it first launched, I prompted for songs “in the precise sound and style of Color Theory” and the results were shockingly good. For free, mind you. Anyone could theoretically whip up a Color Theory album in a half hour.

I shared a track with my patrons and here are some sample comments:

Yeah I can’t tell BrAIan apart from you based just on a voice or lyrics, or the music style… Scary!

This is unreal, if you’d have posted this as your next demo I wouldn’t have thought twice To be fair the vocals are good, but the song hasn’t quite got that Color Theory Magic.

Yes we‘re doomed….a bit. I still can hear a difference, luckily…But to be honest: if you wouldn‘t have told me/us before, I might have not noticed it….

My friend Minute Taker was in there too so I whipped up some songs with him singing ridiculous things he’d never say, for our ears only of course. If nothing else, AI is always good for a laugh!

Even then, I wasn’t angry. I was amazed! And a little bit honored to have made the cut, I suppose.

I’m not saying we don’t need protections for our sonic identities. We absolutely do, and maybe we’ll even get some retroactive compensation via a class action lawsuit. I’m not holding my breath.

Another possible reason I’m not angry (who can really say, emotions being what they are) is that a couple of years ago, I opted in to AI training through a sync/radio platform called AudioSparx. I’ve been active there for many years, and I figured if my music was going to be trained on anyway, I might as well get paid for it.

To be clear, this isn’t how my music ended up in those datasets. The AudioSparx training is fully licensed, as an ethical alternative to the theft detailed above. So far I’ve made $168.85 in AI training royalties.

To sum things up, I suppose I’m pro-AI, anti-theft. You know I’ll be asking Claude to proofread this post before I hit publish!

How do you feel about AI training on our music? Are you angry? Tell me why I should be too in the comments!

I learned that I’ve spent over $12K on Spotify Marquee and Showcase campaigns since 2022.

And what did all that money get me besides a hole in the wallet?

Exactly 107,327 streams from 23,100 listeners at an average cost of $0.55 per listener.

And how do I know this?

It’s all thanks to the free Spotify Campaign Analyzer I built for you. Export your results from Spotify for Artists, drag the file onto the page, and you get everything S4A shows you, plus the analysis it doesn’t:

• Every campaign in one sortable table. Spend, listeners, streams, and intent rate side by side across your entire campaign history.
• The cost metrics Spotify doesn’t calculate: cost per listener, cost per stream, and total streams, including streams of your non-promoted releases.
• Side-by-side comparisons. Pick any set of campaigns and see them charted against each other, metric by metric.
• Benchmarks with context. Any campaign can be measured against your own history or against typical independent-artist results. Tiny, stopped-early campaigns can be disregarded (you set the dollar threshold) so as to not skew your results.
• Apples-to-apples mode. Campaigns that target potential listeners (cold audiences) always look worse on paper. One click recomputes the numbers from warm audiences only.
• A straight answer on in-progress campaigns. Enter the numbers from a campaign that’s still running and get a verdict with clear reasoning: keep going, watch it, or stop.
• Completely private. It’s a single web page. Your data never leaves your browser. Nothing is uploaded, and there’s no signup. There’s not even a donate button!

What my 123 campaigns taught me

Pulling my whole history into one place surfaced a few things I would’ve never spotted on Spotify for Artists:

• Latin America is a bargain. In Brazil I pay about $0.25 per converted listener and in Mexico $0.38, versus roughly $0.60 in the US and UK. Brazil and Mexico also post my highest intent rates (~22%), well above the US (~17%).
• Albums beat singles, by a lot. Album campaigns convert listeners at $0.44 and generate ~5 streams per listener, while single campaigns cost $0.60 for ~1.6 streams each. More tracks means more to explore, but the cost-per-listener differences wasn’t obvious. EPs land in between.
• Almost 40% of my streams came from my back catalog. Across these campaigns, listeners drove ~64k streams of the promoted release, plus another ~41k streams of my other releases. Kind of important to factor in, wouldn’t you say? But S4A keeps it tucked away.
• Marquee and Showcase are different tools. My Marquee campaigns earn higher intent (~22% vs ~15%), but Showcase drives nearly double the streams per listener (3.0 vs 1.7).

Getting your data in

The tool reads the same results file Spotify already lets you download. It never connects to your account or asks for a login.

1. In Spotify for Artists, go to Campaigns → Display.
2. Click the “…” menu next to + Create campaign and choose Download all results. That’s a single CSV covering every campaign you’ve ever run.
3. Drag that file onto the analyzer. Done.

A few things worth knowing: it’s one file for your whole history. There’s no need to download individual campaigns, but it’ll accept them if you do. Your data stays in your browser; nothing is uploaded. It reads both Marquee and Showcase data all the way back to my first campaign in 2022.

It even works on my iPhone via the Files app! That was a pleasant surprise.

Give it a try at passivepromotion.com/spotify-campaign-analyzer

And share your results and suggestions for improvement in the comments!

But are those things actually connected, or am I paying for what the algorithm would do anyway?

I vibe coded a tool to find out!

The Meta Ads to Spotify Analyzer is a single-page, browser-based calculator with no login required. Your data never leaves the page.

Try it here, or use the version baked into this post.

I’ve completely overhauled it, and this post, since I first published the tool. Instead of entering data manually, you simply drop in CSV files from Ads Manager and Spotify for Artists.

The tool re-runs the whole analysis on 300 random shufflings of your conversions and only claims an effect if random data can’t match it.

See for yourself! If you need help locating the proper files, click on “Where do I get these files?” in the import box.

Or if you just want to see what it can do, click the Load Example button below the table.

The minute you click Load Example (or Run Analysis on your own files), you’ll get a plain-English verdict. For the example (a real 44-day campaign plus a couple of baseline days), it comes back with:

 Streams: strong evidence your ads are driving this (chance matched this result in fewer than 1% of 300 trials)
 Intent (saves + playlist adds): suggestive, but not conclusive yet (chance matched this result in about 10% of 300 trials)

Then you get a pretty graph and lots of numbers before the part that actually matters: what’s attributable to the ads.

In this example, each conversion is generating about 8 streams, for a cost of $0.03 per stream.

Keep in mind that this example is catalog-wide. For any individual song, Spotify for Artists only lets you download daily streams. Saves, listeners, and playlist adds are for all of your songs.

So if you include a CSV for an individual song, the streams conclusions will be tighter and ultimately more honest.

I realize it sounds ridiculous to pay that much when Spotify pays us back a paltry $0.003 in royalties, but that doesn’t factor in future streams or algorithmic impact. Good luck measuring that!

When it comes to streams, there are loads of confounding variables:

• organic activity
• The Algorithm
• Marquee and Showcase campaigns
• playlist adds and removals
• proximity to release day
• weekdays vs weekends

It’s so messy, we’re lucky to see any signal at all in the noise, and some of these can fake a correlation as easily as hide one.

Intent paints a cleaner picture. A save or playlist add is a deliberate, same-day reaction to your ad.

In my experience, saves are the best way to judge the success of an ad campaign. If you don’t see an obvious relationship between conversions and saves, something is probably wrong, and this tool will help you spot it.

Alrighty then, that’s a lot of data talk and my brain hurts.

If you’ve got a few spare minutes, try it for yourself and share your experience in the comments!