One opened at twice the other. Three months later it was nine times smaller.
A renewal decision is made on an opening. We measured what happens to that opening across thirteen 2026 streaming titles, month by month, from each title’s own release. Median retention after three months is 7.3%. The range is 2.2% to 44.9%. Every title, every monthly figure and every rejected term is on this page, including the two of four predictions we registered beforehand that failed.
At release Euphoria was 2.2× The Pitt. Three months later The Pitt was 9.1× Euphoria.
Euphoria season 3 opened at 3,350,000 searches a month and kept 2.2% of it. The Pitt season 2 opened at 1,500,000 and kept 44.9%. A twentyfold swing in relative position, between the two largest titles in the panel, inside one quarter.
The two of them, on the calendar and from their own releases
The left panel is the calendar, which is how the data arrives. The inversion is not visible on it, and that is the point: the two titles opened three months apart, so a calendar chart compares April with January rather than opening with opening.
The right panel sets each title’s own release month to 100% and counts forward from there. It is the same two series, aligned to the thing the decision is about.
Every value in both panels is in the two tables below, because a figure that exists only as a shape cannot be checked.
| Months since release | Euphoria season 3 | The Pitt season 2 |
|---|---|---|
| release month | 100.0% | 100.0% |
| +1 | 36.4% | 30.0% |
| +2 | 13.4% | 44.9% |
| +3 | 2.2% | 44.9% |
| +4 | 1.2% | 11.0% |
| +5 | 1.0% | 4.9% |
The Pitt reads 44.9% at both +2 and +3, and that is not a transcription error. Both months return 673,000, which is a Google band value rather than a count. Two months landing in the same band are not measurably different, and the flat segment in the right panel is the banding showing through rather than a plateau we can vouch for.
| Month | Euphoria season 3 | The Pitt season 2 |
|---|---|---|
| 2025-10 | 201,000 | 135,000 |
| 2025-11 | 165,000 | 201,000 |
| 2025-12 | 246,000 | 450,000 |
| 2026-01 | 673,000 | 1,500,000 |
| 2026-02 | 301,000 | 450,000 |
| 2026-03 | 450,000 | 673,000 |
| 2026-04 | 3,350,000 | 673,000 |
| 2026-05 | 1,220,000 | 165,000 |
| 2026-06 | 450,000 | 74,000 |
| 2026-07 | 74,000 | 74,000 |
| 2026-08 | 40,500 | 49,500 |
| 2026-09 | 33,100 | 49,500 |
The decision this is for
A streamer’s content lead deciding whether to renew. It is the most frequent consequential decision in the category, it is made on incomplete information, and it turns on whether attention persists or collapses.
The two answers lead to opposite moves: renew on a strong opening, or wait for the curve because the opening does not predict the tail.
On this evidence, the opening tells you roughly who the small titles are and nothing reliable about which of your big ones is still alive in a quarter.
The correlation is high, and it inverts exactly where decisions are made
Rank correlation between the release month and three months later is 0.841 across thirteen titles. That sounds like opening order is reliable. The top four say otherwise.
| Rank | At release | Three months later |
|---|---|---|
| 1 | Euphoria season 3 | The Pitt season 2 |
| 2 | Bridgerton season 4 | Bridgerton season 4 |
| 3 | The Pitt season 2 | Monarch: Legacy of Monsters s2 |
| 4 | Daredevil: Born Again s2 | Euphoria season 3 |
| 5 | Stranger Things: Tales From '85 | Stranger Things: Tales From '85 |
| 6 | The Night Agent season 3 | Daredevil: Born Again s2 |
The correlation is carried by the tail, where small titles stay small. At the head it reshuffles — and the head is where a renewal conversation happens.
Every title, best retention to worst
| Title | Released | Peak / month | +3 months | Retained | Sep 2026 |
|---|---|---|---|---|---|
| The Pitt season 2 | 2026-01 | 1,500,000 | 673,000 | 44.9% | 49,500 |
| Monarch: Legacy of Monsters s2 | 2026-02 | 165,000 | 74,000 | 44.8% | 3,600 |
| For All Mankind season 5 | 2026-03 | 90,500 | 33,100 | 36.6% | 6,600 |
| House of David season 2 | 2026-03 | 40,500 | 8,100 | 20.0% | 4,400 |
| Stranger Things: Tales From '85 | 2026-04 | 301,000 | 40,500 | 13.5% | 90,500 |
| The Night Agent season 3 | 2026-02 | 246,000 | 18,100 | 7.4% | 8,100 |
| A Thousand Blows season 2 | 2026-01 | 12,100 | 880 | 7.3% | 390 |
| Bridgerton season 4 | 2026-01 | 1,830,000 | 90,500 | 4.9% | 22,200 |
| Daredevil: Born Again s2 | 2026-03 | 823,000 | 40,500 | 4.9% | 12,100 |
| The Legend of Vox Machina s4 | 2026-06 | 27,100 | 1,300 | 4.8% | 1,300 |
| Devil May Cry season 2 | 2026-05 | 165,000 | 5,400 | 3.3% | 3,600 |
| The Lincoln Lawyer season 4 | 2026-02 | 110,000 | 3,600 | 3.3% | 1,000 |
| Euphoria season 3 | 2026-04 | 3,350,000 | 74,000 | 2.2% | 33,100 |
Median retention at three months is 7.3%. Half the panel keeps less than a fourteenth of its opening demand within a quarter — and the spread around that median is so wide that the median describes almost nobody in it.
What we said beforehand
| # | Registered | Prior | Outcome |
|---|---|---|---|
| E1 | Opening predicts the tail. | roughly even chance · 45–55% | Holds. Rank correlation between the release month and three months later is 0.841 across thirteen titles. The falsifier required weak correlation and did not fire. |
| E2 | Decay rate is the signal, not level. | likely · 55–80% | Holds. Retention at three months runs 2.2% to 44.9% — a 42-point spread across the panel. |
| E3 | Everything decays at about the same rate. | unlikely · 20–45% | Dead. A twentyfold range in retention is not a common rate. |
| E4 | The signal is too short-lived to support the decision. | roughly even chance · 45–55% | Dead. Every title is still measurably above its own pre-release baseline at three months, and at September, five to eight months out. |
A defect in our own registration, kept on the page. E1 and E2 were written as competing explanations — either the opening predicts the tail or the decay rate carries the information. They are independent properties and both hold. Rank survives precisely because the levels differ by two orders of magnitude, so the largest title stays largest even while decaying fastest. Asking which one was true was the wrong question, and a false dichotomy returns two holds and scores neither.
Four titles were rejected, and our own screen was too blunt
A title whose name is a common word or belongs to an older work does not measure itself. The screen registered before the pull was: a 2026 title carrying search volume in 2019 to 2024 has a contaminated term.
Applied literally it rejects twelve of nineteen — because it cannot tell a different referent from legitimate anticipation for the measured season. Searches for a fourth season are about the fourth season, however early they start.
The line that works is a rule rather than a threshold fitted to the result: volume in 2019, the earliest full year, before any title in this panel existed or was announced. A term with real 2019 volume refers to something that was already there.
| Term | 2019 searches / month | What it already was |
|---|---|---|
| A Knight of the Seven Kingdoms | 4,625 | A Game of Thrones novella collection — and it spikes in April 2019, when that show’s final season aired |
| Scarpetta | 16,792 | A restaurant group and a long-running novel series |
| The Night Manager season 2 | 4,392 | A decade of viewers asking for a second season of a 2016 series |
| Young Sherlock | 347 | A 1985 film and the books behind it |
Both figures are published. The registered screen rejected twelve; the refined screen rejects four. The registered one gave the wrong answer and the record says so rather than presenting the refined one as what was always intended.
The comment coder, and what it is worth
An earlier pass of this study used trailer comments and the coder failed its own test on all five terms. Trailer comments are published before release, so they cannot carry a question about what happens after it. This pass uses post-release review and reaction videos instead.
| Corpus | n | binge | rewatch | boring | quit | disappointing |
|---|---|---|---|---|---|---|
| The Night Agent season 3 | 294 | 0.68% | 0.34% | 1.36% | 0.34% | 0.00% |
| Ted Lasso season 4 | 442 | 0.23% | 0.68% | 0.90% | 0.23% | 0.23% |
| Scarpetta | 378 | 0.53% | 0.00% | 1.59% | 0.53% | 1.06% |
| NEGATIVE CONTROL | 404 | 0.74% | 0.00% | 0.25% | 0.25% | 0.50% |
Two of five terms now discriminate, against none in the earlier pass. Boring separates cleanly: the panel runs 0.90% to 1.59% and the control runs 0.25%. Binge separates in the opposite direction. The other three sit inside the panel range and do not.
What this does not do is explain the decay. No term maps onto the retention figures, and the strongest discriminator may simply be that a scripted drama invites the word “boring” and a sports documentary does not. It is reported as context, not as evidence.
What this study cannot tell you
- Search volume is interest, not viewing. Nothing here observes a minute watched. A streamer’s own data would answer this better and is not public.
- Renewal is decided on viewing, completion and cost. This measures one input and does not claim otherwise.
- Google bands and rounds these figures and may rewrite a past month. Values cluster on band boundaries — 90,500, 110,000, 135,000 — so small differences between titles are not resolvable.
- The panel is declared, not drawn from a register. It cannot make the “every filer above a floor, nothing hand-picked” claim that the filings studies on this site make, and the inclusion rule is published above rather than implied.
- Thirteen titles with a full three-month window. Enough for a rank correlation, thin for a distribution.
- Eleven of thirteen are returning seasons. A first season has no prior audience and may decay differently; this panel cannot say.
- One title of twenty returned no series at all and is recorded as absent rather than as zero.
How this page changes
A standing study, not an article. Corrections are published in place and dated, never applied quietly.
| Measurement | Cadence |
|---|---|
| Monthly demand, every panel title | Quarterly |
| Whether the three-month ordering holds at six months | Q1 2027, scored on the page |
| Rejected terms and their 2019 baselines | Every re-run, republished in full |
| Post-release comment corpora | Quarterly, with the control |
| Date | What changed |
|---|---|
| 2026-10-25 | First publication. An earlier pass of this study used pre-release trailer comments, whose coder failed its own test on every term; that pass is superseded and the reason is in the section above. The contamination screen registered before the pull rejected twelve titles and was too blunt; both it and the refined screen are published. |
This page is re-run each quarter, and the three-month ordering is scored again at six.
Sound Decisions carries each re-run and each correction, with the registered question written before the data is pulled.
Subscribe to Sound Decisions