A hand-marked jam start on every track in Phish history makes it measurable: the improvised share of a show rose from under a third in the early '90s to about half. Jam time per show nearly doubled, and it has held there since 1997.
Every jam in Phish history has a hand-marked start time on this site — the exact second the written song ends and the improvisation begins. Sum those markers across the whole catalog and you can answer a question no setlist can: how much of a Phish show is actually made up on the spot?
The answer, for a modern show, is about half. In the early 1990s it was under a third. The whole jump happened between 1994 and 1997, and in the nearly three decades since, the share has barely moved.
Here is the improvised share of stage time, year by year: total jam seconds (track length minus the tagged jam start, counting only tracks that were jammed) divided by all the time the band spent playing.
From 1987 through 1993 the line sits flat near 29.8% — under a third of the night made up on the spot. Long songs existed, but most of the setlist was played as written.
Then a fast climb. Between 1994 and 1997 the improvised share rises year for year, from the mid-30s toward half. This one window carries the entire move.
In 1997 the line first crosses the 50% "half the show" mark. Half a Phish show is now improvised.
And then it holds. Across 2015–2025 the improvised share averages 51.8% — hovering around half for nearly three decades. The shape is a single step up to a plateau.
The two straight bridges in the line are the 2000–02 and 2004–09 hiatuses (no touring). Per show, the absolute jam time nearly doubled. In the early 1990s a night held about 41 minutes of improvisation; a modern show delivers about 87. (The 1.0-era average in the table below, 51.5 min, already folds in the post-1994 climb.)
| Era | Shows | Jam min / show | Show min | % improvised |
|---|---|---|---|---|
| 1.0 (pre-2000) | 1052 | 51.5 | 140.9 | 35.9% |
| 2.0 (2000–04) | 116 | 85.3 | 163.1 | 51.6% |
| 3.0+ (2009–) | 682 | 87.0 | 174.6 | 49.9% |
The rise came from jamming more songs rather than stretching each one longer. The share of songs that get a jam at all makes its move in the same window, from 37.7% in 1993 to 57.6% in 1997, climbing year for year with the improvised-share line above.
After 1997 the two measures part ways. The songs-jammed share keeps climbing, reaching about 70% by 2025, while the improvised share of the night holds near half (Fig 3). They stop rising together because the average jam, once it happens, slipped from 7.6 minutes in 1997 to 6.4 across 2015–2025 — just enough to cancel the higher rate. That slip is a change in the mix, not the veterans shortening: the established jam vehicles run as long as ever, while the newer songs pulled into the rotation carry shorter jams (6.1 minutes against 6.7 for the veterans). Improvisation kept spreading across the setlist even as each new jam ran a little shorter.
The whole finding rests on one assumption: that the tagging is even across eras. It is. 99.3% of 1.0 tracks, 99.6% of 2.0, and 97.2% of 3.0+ tracks carry an explicit decision — a marked jam start, or a “played straight” flag. The early years aren’t under-tagged, so the low early share is real.
Some songs are almost all improvisation once they start: a short written head, then open playing. Rank every song with 30 or more tagged jams by the share of its runtime that falls after the jam start and you get the catalog’s purest launch pads.
| Song | Jams | % jam | Median jam (min) |
|---|---|---|---|
| Tube | 190 | 84% | 5.2 |
| Ruby Waves | 40 | 83% | 12.4 |
| Timber (Jerry the Mule) | 106 | 80% | 5.6 |
| Weekapaug Groove | 463 | 79% | 6.1 |
| 46 Days | 150 | 78% | 6.2 |
| Scent of a Mule | 137 | 76% | 6.8 |
| Run Like an Antelope | 472 | 75% | 8.4 |
| Undermind | 64 | 75% | 5.9 |
| Waves | 57 | 74% | 7.0 |
| A Song I Heard the Ocean Sing | 50 | 74% | 5.9 |
| Plasma | 35 | 74% | 6.3 |
| No Men In No Man's Land | 83 | 74% | 7.6 |
| Maze | 340 | 74% | 7.6 |
| Set Your Soul Free | 44 | 74% | 9.9 |
| A Wave of Hope | 34 | 74% | 12.8 |
“Phish jams more now” is the kind of thing everyone knows and no one had measured. Measured, it’s a near-doubling of jam time per show, most of it in place by 1997, and a share that has sat at half ever since. The claim covers quantity, counted in minutes; it says nothing about the quality of the playing or where the best of it happens. What makes it possible to state at all is the jam-start tag on every track.
Without a composed-to-jam boundary, “half the show” is a feeling. With one, it’s a number.
The metric. Improvised share = summed jam seconds ÷ total played seconds. Jam seconds for a track = track duration − the hand-tagged jam_starts_at_second, counted only when the track is marked as jammed. Tracks flagged “played straight” (jam_start_excluded) contribute zero jam seconds but their full duration to the denominator. Sets 1, 2, 3, and encores; soundchecks excluded.
Where the jam ends. The metric treats everything after the jam start as improvisation, including composed material that resolves back at the end — the Antelope outro, the return of the Mike’s or Weekapaug theme, a written coda. That slightly overstates the absolute jam minutes. Because the effect is roughly constant across eras, the trend and the era comparison hold; read the per-show minutes as a modest ceiling, not a precise count.
Coverage. The finding depends on even tagging across eras. 99.3% / 99.6% / 97.2% of tracks in the three eras are adjudicated (a marked jam start or an explicit no-jam flag), so the era comparison is not a coverage artifact. See the jam-start tagging piece for how the markers were set.
Per-year line (Fig 1). Years with at least 200 tagged main-set tracks; the current partial year is excluded. The x-axis is calendar year and the two long segments bridge the 2000–02 and 2004–09 hiatuses (no touring).
Eras. 1.0 = before 2000, 2.0 = 2000–2004, 3.0+ = 2009 onward.
Jam length over time. The mean jam among jammed tracks slips from 7.6 min (1997) to 6.4 min (2015–2025). To check whether that is the old vehicles shortening or newer songs joining with shorter jams, “established” songs (jammed ≥5 times in years ≤2000) are compared then and now: they hold their length (6.8 min in 1994–2000, 6.7 min in 2015–2025), while newer songs average 6.1 min. The slip is a change in the mix, not fatigue in the veterans.
Table 2 percentages. The “% improvised” column is the mean of each show’s own improvised share, not the ratio of the two minutes columns beside it. Averaging per-show ratios weights every show equally, so it does not exactly equal jam minutes ÷ show minutes (51.5 ÷ 140.9 = 36.6%, versus the 35.9% shown). The per-year line (Fig 1) instead pools all seconds within a year.
Jam share of runtime (Tbl 4). For each song with at least 30 tagged jams, the mean of (duration − jam start) ÷ duration across its versions — the fraction of a typical version spent improvising. Untitled “Jam” listings, Big Ball Jam, and Llama are excluded as unrepresentative of open improvisation.
Reproducibility. All aggregates are precomputed in scripts/analysis_improv_share.mjs and served as JSON at data.json. Data as of 2026-07-24.