Which App Would You Trust With a 16-Week Marathon Build?

A marathon is decided by four things: how your long run grows, how much hard running sits around it, what happens in the last three weeks, and what the plan does when life removes a fortnight from the middle of it. Everything else is decoration. Yet almost every comparison of the best app for marathon training plan generation spends its word count on whether the calendar view scrolls nicely and how the haptic cues feel at 4:58/km.

So here is a different test. I built one athlete, fed them into the major AI coaching apps, and audited only the four things above.

The test athlete

Forty-two years old, two kids, four runs a week, 42 km of honest base volume, a 21:40 parkrun from August, and 16 weeks to a first attempt at sub-3:30 (4:57/km, so a 3:29 finish with a little change). No injury history worth naming, one sensitive left achilles that grumbles after anything over 28 km on road.

This is the median person buying a running app. They are not fragile and not elite, and the plan has to earn about 8% improvement out of them in four months.

Test one: does the long run actually progress, or just get longer?

Progression is not a synonym for increase. A good long-run ladder steps up for two or three weeks, steps back, then steps up past the previous high. Here is what the apps produced, in kilometres, weeks 1 through 16:

Runna and Garmin Coach+ pass. Both step back roughly every fourth week, both cap the peak (Runna at 32 km, Garmin at 2:45 regardless of pace), and both touch that peak twice rather than once, which is the bit that matters. Doing 32 km once is a stunt. Doing it in week 10 and again in week 12 is a training adaptation.

Garmin’s decision to prescribe in time rather than distance is quietly the strongest thing in this table. Our athlete runs long at about 5:50/km, so a 2:45 long run lands around 28 km. If they were slower, it would land at 25 km and take the same 2:45 out of their week and their legs. Distance-based ladders punish slow runners with four-hour training days, and four-hour training days at week 12 are how achilles grumbles become achilles injuries.

Nike Run Club fails on cap discipline: 35 km in week 10 for a 3:29 target is more than the race demands of you per hour, and it is a static plan so it does not know how the 32 went. The ChatGPT plan fails on shape. Look at it again: fourteen weeks with only one genuine step-back, a 26 km long run in week 4 off a 42 km base, and a 38 km peak that nobody asked for. This is the signature failure of a prompted plan. The model knows what a marathon block looks like in aggregate and produces a monotonic staircase, because “reduce this by 30% for one week” is not a thing the text of training books emphasises as loudly as “build the long run.”

Test two: count the hard days, not the sessions

Here is week 7 from the ChatGPT plan, which the athlete had been running before we started:

Tue  6 x 1 km @ 4:30, 90s jog             hard
Wed  8 km easy                            easy
Thu  3 km wu, 8 km @ 4:45, 2 km cd        hard
Sat  27 km, last 8 km @ marathon pace     hard
     ----
     3 hard days out of 4 runs, 55 km total

Three of four runs are quality. That is not a marathon week, it is a 10K block with a long run stapled to the end of it, and the athlete had been doing versions of it for six weeks with a resting heart rate that had drifted up four beats.

The comparison week 7 from Runna: one interval session (5 x 1 km at 4:35), one easy 8 km, one steady 10 km with 4 km at threshold, and a 27 km long run entirely easy with no marathon-pace work until week 9. Two hard days, and crucially the marathon-pace segments arrive in the long run only after the long run itself has stopped growing. Garmin Coach+ is more conservative still, typically one true quality session plus a daily suggested workout that it will downgrade to recovery if your HRV and sleep scores are poor.

The number to hold onto: on four runs a week, two hard days is right and three is the ceiling you only reach for a fortnight. If your app is giving you three every week for eight weeks, it is not coaching you, it is entertaining you. Our pillar comparison, AI Coaching Apps, Tested Against Each Other, scores this same intensity-distribution question week by week across a full block if you want the raw tables.

Test three: the taper, where plans quietly stop being plans

A taper should cut volume hard and keep intensity almost intact. The rough target from the literature is roughly 40% to 60% volume reduction across two to three weeks, with session quality preserved.

Runna’s last three weeks came out at 79%, 61% and 43% of peak volume, with a 4 x 1 km at 10K effort nine days out and 6 x 400 m on the Wednesday of race week. Good. Garmin Coach+ ran 76%, 58%, 41% with strides retained through the final week. Also good. Nike Run Club dropped to 72%, 55%, 38% and kept a tempo, which is defensible for a static plan.

The prompted plan tapered by deleting the hard running: 32 km long run in week 14, then “easy running only, 3 to 4 runs” for the last fortnight. Two weeks of nothing but easy running before a marathon leaves you flat, slightly heavy, and unable to find 4:57/km on race morning because you have not touched it since week 13. Ask any model for a taper and it will reliably cut the wrong variable, because “rest and recover” is the loudest sentence in its training data about tapers.

Test four: the missed fortnight

Our athlete gets a chest infection in week 9 and runs nothing for 13 days. This is the test that separates software from a coach, and it is the one nobody demos.

Nike Run Club does nothing useful. The calendar sits there, week 11 waiting with a 32 km run, and the athlete either does it or abandons the plan. Runna offers to reschedule, which shifts the remaining sessions forward and compresses the build. Better than nothing, and it will now ask about the missed time, but the compression is real: you cannot recover a fortnight in a 16-week block by sliding the blocks along.

Garmin Coach+ handles this best of the subscription apps because it was never reading from a fixed calendar. It recalculates from your actual acute and chronic load, so after 13 days off, your training status reads “unproductive” or “detraining”, the daily suggested workout comes back at 45 minutes easy rather than 6 x 1 km, and the long run rebuilds from around 18 km rather than resuming at 30. TrainAsONE, which generates every session from your history rather than from a template, does the same thing more aggressively and will bluntly reduce your target finish time. That last behaviour is the most coach-like thing in any of these apps, and the most unpopular.

The prompted plan will do whatever you ask it to. Tell ChatGPT you missed two weeks and it will cheerfully hand you a revised 7-week build that still reaches 38 km, because it has no model of your accumulated load and every incentive to agree with your goal.

Audit your own plan before Sunday

Pull your last eight weeks from Strava or Garmin Connect and write down four numbers:

longest run, weeks 1-4 avg :  __ km
longest run, weeks 5-8 avg :  __ km   (want +12 to 18%, not +30%)
hard days per week, mean   :  __      (want 2.0-2.5 on 4 runs)
weeks with a step-back     :  __      (want 1 in every 3-4)

If your step-back count is zero, your plan has no recovery architecture and the peak weeks will find that out for you around week 12. If your hard-day mean is above 2.5, the fix is not more sleep, it is deleting a session. And if you missed time and your plan never noticed, stop trusting it with the taper, which is the one part of the block you cannot re-run.

The next fortnight is the cheapest place to make that change.