The rule is always phrased the same way and almost never attributed: no caffeine after 2 p.m. It may be the most repeated sleep improvement instruction in circulation — printed on café chalkboards, forwarded by group chats, delivered by people who have never measured a single night. I have followed it, broken it, and felt vaguely guilty about breaking it, all without knowing whether it was doing anything at all.
So I tested it. Ten weeks, 72 tracked nights, four caffeine conditions, one increasingly irritable subject.
The verdict: the 2 p.m. cutoff is roughly right and badly aimed. Moving my last cup earlier did produce a measurable change. But the size of the dose mattered more than the hour it landed, and neither variable touched the problem that sent me down this road in the first place.
The setup
My baseline is about 300 mg of caffeine a day: a 10-ounce pour-over around 7 a.m. (roughly 150 mg), a double espresso mid-morning (roughly 130 mg), and — the habit under investigation — a small cup around 3:45 p.m. when the afternoon goes flat.
I split ten weeks into four blocks of 18 nights, drawing the order out of a bowl so I couldn't front-load the easy conditions:
- A — the rule: 300 mg, last dose by 2 p.m.
- B — late: 300 mg, last dose between 4:00 and 4:30 p.m.
- C — early: 300 mg, last dose by noon.
- D — small: 150 mg, last dose by 2 p.m.
Every dose was weighed and brewed the same way, and logged with a timestamp. I wore a finger tracker and slept on a bedside mat that measures the same things by different means; where they disagreed by more than 15 minutes I threw the night out (six nights, spread evenly across conditions). Bedtime stayed within a 40-minute window. I rated morning grogginess 1–5 on my phone thirty minutes after waking, before coffee, which is the only honest time to do it.
The numbers
Means across each 18-night block, tracker-derived except grogginess:
| Measure | A: 300 mg, 2 p.m. | B: 300 mg, 4 p.m. | C: 300 mg, noon | D: 150 mg, 2 p.m. |
|---|---|---|---|---|
| Sleep onset latency | 14 min | 23 min | 13 min | 11 min |
| Wake after sleep onset | 41 min | 46 min | 39 min | 33 min |
| Deep minutes, first half of night | 62 | 51 | 64 | 71 |
| Nights with a 2:30–4:30 a.m. wake | 9 of 18 | 10 of 18 | 9 of 18 | 8 of 18 |
| Morning grogginess (1–5, lower is better) | 2.6 | 2.9 | 2.7 | 2.4 |
Where the rule holds
The late condition was genuinely worse, and not subtly. Nine extra minutes to fall asleep and eleven fewer deep minutes in the first half of the night is the kind of gap I expected to have to squint at, and I didn't. The distributions barely overlapped at the tails: my three longest onset latencies of the whole experiment (38, 41, and 52 minutes) all came from B, and two of those followed a 4:30 p.m. cup taken on an empty stomach.
That is the defensible core of the advice. Caffeine has a half-life in the neighborhood of five hours for most adults, which means a 4:30 p.m. dose still has meaningful concentration in you at midnight. Whatever the tracker's stage labels are actually measuring, something about the front of my night changed when that was true.
Where it breaks down
The noon condition bought me almost nothing. One minute of onset latency, two minutes of wake time, two deep minutes — all inside the night-to-night noise of the same condition. I gave up my mid-morning espresso timing, spent 18 afternoons feeling like a person waiting for something, and got a rounding error.
Meanwhile, halving the dose while keeping the same 2 p.m. cutoff outperformed every timing change I made. Nine more deep minutes than the rule condition, eight fewer minutes awake, the lowest onset latency of the four. If I had run only conditions A and C — which is effectively what the popular advice invites you to do, tightening the clock and holding the amount fixed — I would have concluded that caffeine timing barely matters and gone back to my afternoon cup.
If you change one thing, change how much before you fuss over the hour. The 2 p.m. rule is a proxy. For most people it works because moving the deadline earlier quietly removes a dose — not because 2 p.m. is a boundary in your bloodstream.
The thing I was actually trying to fix
I did not start this to optimize deep sleep. I started it because I wake up at 3 a.m. maybe half the nights of my life, lie there doing arithmetic about how much sleep is left, and fall back asleep around 4:15.
That number did not move. Nine, ten, nine, eight nights out of eighteen. Cutting my caffeine in half did nothing for it. Cutting it off at noon did nothing for it. Across the whole experiment the 3 a.m. wake rate was 36 of 72 nights, which is exactly the half I walked in with.
I then did the thing you should be suspicious of: I went back through the log looking for what did separate wake nights from quiet nights. Two candidates fell out — any alcohol at all, and a bedroom above about 21°C. Both look convincing on a scatter plot. Both are also what you get when you scan a dozen variables across 72 nights until something lines up. I'm noting them as things to test properly, not as findings. Discovering your hypothesis in your own data is not the same as testing it.
The low-dose block also came with a cost the table flattens: my first five mornings in condition D averaged 3.8 on the grogginess scale, worse than anything else in the experiment. The 2.4 average only exists because the back thirteen nights were so good. Anyone reading that number as "less caffeine, better mornings" should know the first week felt like a mistake.
What I couldn't rule out
No blinding — you cannot blind yourself to whether you had coffee, and I knew which condition I was in every single night, which means expectancy is baked into every row above. Consumer trackers estimate sleep stages from movement and heart rate; "deep minutes" is a model output, not a measurement, and I have no lab data to check it against. Ten weeks moved me from late winter into spring, so my evening light exposure drifted the whole time. I had a head cold in week six (three nights excluded). Weekends were noisier than weekdays in every condition. And it is one body: mine metabolizes caffeine at whatever rate it does, and the person who clears it in three hours or in nine is not going to recognize these numbers.
Who this is for — and who it isn't
Worth running yourself if: you drink more than about 250 mg a day, you already have 60+ nights of tracker data to compare against, and you take a long time to fall asleep. Onset latency is the metric that moved most reliably here, and it's the one your device is most likely to get right.
Not worth it if: your main complaint is mid-night waking. Nothing I did to caffeine touched it, and I spent ten weeks confirming that. Also skip it if you're already under about 150 mg a day — you don't have enough room to run a meaningful low-dose arm, and you'll just be moving one small cup around a calendar.
And if you're testing anything: vary the amount and the timing separately. Confounding them is how a rule about milligrams gets remembered as a rule about clocks.
The myth is that 2 p.m. is a line, and that coffee taken before it is free while coffee taken after it costs you your night. The truer version is that your last dose has a long tail into the dark, that halving the total does more than moving the deadline, and that if you wake at 3 a.m. anyway, the coffee was never the reason.
