What 2,760 Live-Picked Songs Say About How I Choose Music (and Where Gemini Got It Wrong)

—

,

This is a follow-up to Instagram Artifacts, my daily music and visual storytelling practice. Every day I build a playlist live, on the fly, one song at a time. After a few months of that, I wanted to know what my own choices looked like from the outside: what I actually reach for, how I get from one song to the next, and whether any of it was changing.

So I ran an experiment with two AI models. Gemini gave me a fast first read. Then I had Claude parse and count every song across three months of playlists, and check Gemini’s claims against the real numbers.

How I build a playlist

Context matters here, because the playlists aren’t passive listening logs:

  • Everything is picked live. On big days I’m DJ-ing for myself, jumping to the next song before the current one ends. I often skip after a minute or so; sometimes I listen all the way through.
  • Some days are deep dives. When I find an artist I like, I add a run of their songs and save them so the music app’s recommendations learn from it.
  • Life leaks in. When I was watching Bob’s Burgers, songs from the episode (or ones that matched it) went straight into the playlist. On sluggish mornings, I reach for heavy metal, especially Judas Priest, to wake up.

So a song count measures choices, not listening time. That’s exactly what I wanted to study.

The data

  • 17 dated playlists from August 31 to November 15, 2026, exported from YouTube Music as plain text.
  • 2,760 songs picked: 206 in late August and September, 1,480 in October, 1,074 in November.
  • Each song’s genre is assigned by its lead artist. That’s an approximation (a Britney remix still counts as pop), but it’s consistent across all three months.

Step 1: Gemini’s first read

I gave Gemini four recent days and asked what patterns it saw in how I pick the next song. Its read was quick and genuinely useful in places. It spotted that TV moments trigger songs, that comedy tracks act as a reset between high-energy runs, that I sometimes chain songs by a shared word in the title (Nickelback’s “Rockstar,” then N.E.R.D.’s, then LISA’s), and that remix producers like Purple Disco Machine keep the groove going across artist changes.

It also gave me percentages, a list of “keywords,” and a confident explanation of what all this says about how my mind works. Those are the parts that needed checking.

Step 2: Counting everything

Claude parsed the full export into a table of songs, then measured the mix by period and every song-to-song transition: same artist, shared featured artist, shared title word, and whether a song had appeared in an earlier playlist.

Share of songs picked, by genre
Aug–Sep: 206 songs · October: 1,480 · November: 1,074. Genre assigned by lead artist.
Aug–SepOctNov
Pop & Dance
34%
35%
37%
Hip-Hop & R&B
21%
32%
15%
Rock & Metal
26%
15%
10%
K-Pop & Asian Pop
13%
7%
9%
Comedy, TV & Film
2%
5%
11%
Jazz & Blues
0%
4%
10%

Pop and dance music is the constant: about a third of everything, every month. Around it, the mix moved a lot. I started out rock-heavy, went through a hip-hop phase in October, and by November jazz and comedy each made up about a tenth of what I picked.

How the next song gets picked
Top three rows: share of song-to-song transitions. Bottom row: share of songs that are new to the archive.
Aug–SepOctNov
Next song by the same artist
6%
29%
21%
Next song linked by a shared title word
11%
5%
4%
Next song shares a featured artist
3%
7%
7%
Songs not in any earlier playlist
97%
73%
63%

What the numbers showed

1. I moved from word games to deep dives

Early on, title words drove a larger share of my song changes (11% in August–September). The best example is a six-song run on September 27: Whatta Man, Sharp Dressed Man, Any Man of Mine, Rocket Man, Iron Man, Piano Man. Across all 17 playlists there are 31 chains of three or more songs linked this way, including “Money” (six songs), “Louie Louie” (five), “Fly” (five), and “Chicago” chains in two different playlists.

By October, the dominant move was going deep on one artist: back-to-back songs by the same artist jumped from 6% of transitions to 29%, and runs of five or more songs went from one to 26. The longest runs: 13 WWE entrance themes, 12 Britney Spears songs, and 10 each from Judas Priest and Chris Brown. Some of that is the deep-dive saving I described above; the data can’t tell listening from saving, and I can.

2. My sessions grew an opening ritual

Every playlist from October 27 to November 7 opens with jazz: Take Five twice, My Favorite Things, and Under a Blanket of Blue. Two of the next four open with comedy instead. I didn’t notice this habit until I saw it in a list.

3. A core rotation is forming

In the first playlists, almost every song (97%) was new to the archive. By November, that was down to 63%. Two hundred songs now appear in three or more playlists. The top anchor is The Donnas’ Take Me to the Backseat, in seven playlists; Toxic, Melatonin, Bad Reputation, and LISA’s Born Again remix each appear in six. Sixty-three artists show up in all three periods, led by Britney Spears, Dua Lipa, Ariana Grande, and The Donnas.

Where Gemini was right, and where it wasn’t

On the same four November days Gemini analyzed (559 songs), here’s how its numbers compare with the counts:

ClaimGeminiActual
Songs in the sample~560559
Pop & dance35%40%
Latin & reggaeton18%5%
Rock & metal14%13%
Comedy & TV10%14%
  • Right: the song count, the rock share, and most of its qualitative observations (TV triggers, comedy resets, title chains, remix continuity).
  • Wrong: Latin was overstated almost fourfold, partly because it filed Lexie Liu, a Chinese artist, under Latin. Several of its “keywords” don’t appear anywhere in the data.
  • Missed: the biggest patterns only show up across months: the shift from word chains to deep dives, the jazz openers, and the growing core rotation.
  • Overreached: its explanation of why metal wakes me up (drums that “mimic a surging heart rate”) sounded scientific but wasn’t. Fast, loud music raising alertness is plausible; the rest was decoration. Same with “hyper-literate stream of consciousness”: flattering, not measurable.

What I took away

  • AI is a good first reader and a bad accountant. Gemini’s quick read gave me the right questions. The answers came from counting.
  • Patterns live in the transitions, not the totals. Genre percentages say what I like. Song-to-song links say how I think while picking.
  • Check the confident parts first. The most specific-sounding claims (percentages, keywords, physiology) were the least reliable.
  • My taste is consistent; my method isn’t. Pop is a fixed core, but how I move between songs changed a lot in three months.

Tools: Gemini (first-pass analysis), Claude (parsing, counting, verification and charts), YouTube Music (playlist exports). Part 1: Instagram Artifacts: A Live Music Diary Made with My Own Creative AI.