Adaptive or Just Rescheduled? How Coaching Apps Handle a Missed Run

You get a cold in week 9 of a 16 week marathon build. Seven days, no running. You open the app on the Monday and the calendar looks… fine. Tidy. Week 10 is sitting there exactly as it was, 58km with a 29km long run and 6×1km at 4:35/km on the Tuesday. Nothing acknowledges that you have done nothing for a week.

That moment is the single most useful test of any AI plan you are paying for. So: do running apps adapt to missed runs, or do they just move the tiles around and hope you keep swiping? We ran the same two disruptions through several of the popular options and the answer splits the market almost perfectly in half.

Two things get called “adaptive”

Rescheduling is calendar arithmetic. You tell the app you can’t run Tuesday, it puts Tuesday’s session on Wednesday, the week’s total stays identical, and the following 7 weeks are untouched. Useful! Not adaptation.

Rebuilding means the plan recalculates what you can absorb from here, and then changes something structural: next week’s volume, the date of the peak long run, the length of the taper, the paces, or the race target. A rebuild always has a visible cost. Something moves. If nothing moves, nothing was recalculated.

Most apps marketed as AI coaches do the first thing and let you believe it’s the second.

The test

A 16 week marathon plan, target 3:45, built in each app from the same inputs: 38 year old, 42km a week base, half marathon PB 1:44, four runs a week, long run Sunday. Then two interventions at week 9.

Test A: skip one key session (Tuesday, 5×1km at 4:35/km) and leave everything else alone.

Test B: skip all 7 days of week 9. No runs logged, no activity in Strava or Garmin Connect.

Then look at exactly three numbers in the regenerated plan: total distance for the next 7 days, the date and distance of the peak long run, and the goal pace.

Test A was less dramatic and more revealing. Every app in the list let the missed interval session simply expire, and not one of them reduced the following week’s intensity to compensate for a lost quality day. Which is arguably correct: one missed session is noise. But the three rebuilders logged it against cumulative load, and the reschedulers didn’t log it at all, and that difference compounds.

The missed week, in actual numbers

Here is the diff that matters. Same athlete, same week 9 absence, two different philosophies about week 10.

COMPLETED
  wk 6   48 km   long 24 km
  wk 7   52 km   long 26 km
  wk 8   54 km   long 27 km
  wk 9    0 km   long  0 km   <- illness

RESCHEDULER (plan as originally generated)
  wk 10  58 km   long 29 km   6x1km @ 4:35 + 8km tempo @ 4:55
         jump vs last completed week: +7.4% volume, +7.4% long run
         jump vs last 7 days:        +infinity

REBUILDER (TrainAsONE, regenerated Monday morning)
  wk 10  41 km   long 21 km   1 easy quality day, no tempo
         76% of last completed week, then +12%/wk back to 29 km by wk 13
         peak long run:  wk 13 -> wk 14
         taper:          3 weeks -> 2 weeks

The reschedulers aren’t obviously wrong on paper. After 7 days off, aerobic fitness has barely budged: VO2max losses in the first week of inactivity are small, in the low single digits. But plasma volume drops meaningfully inside 7 to 14 days, which is why the first run back feels awful at a pace that was easy a fortnight ago, and why heart rate for a given pace can sit 5 to 10 beats high. Your fitness is mostly intact. Your tolerance for a 58km week with two quality sessions, immediately, is not.

What a human coach does here is make a trade. They decide whether to pay for the missed week out of the build (push the peak long run back, shorten the taper) or out of the goal (accept a slightly slower target and keep the taper). TrainAsONE and Athletica both made that trade explicitly and showed me the cost. Runna and Coopah made no trade, because they never registered that anything happened.

The catch-up trap is worse than the missed run

Rescheduling actively creates a specific injury risk that doing nothing wouldn’t. Say you miss Sunday’s 27km long run in week 9 and drag it to Monday, which every calendar-based app happily lets you do. Week 10’s long run is still on the Sunday at 29km. You have now run 56km of long running inside 7 days, with the second one arriving six days after the first, and no app in the reschedule column warned me about it.

Runna’s session-moving interface is genuinely the slickest of the lot, and that is precisely the problem: it is so frictionless to drag a 29km run into a bad spot that you’ll do it without thinking. A tool that makes an unwise thing easy and says nothing is not neutral.

Garmin’s split personality

Garmin is worth separating out, because it ships both failure modes in one watch. Garmin Coach gives you a periodised block with a race date and a structure, and that structure is close to inert: after the missed week it nudged paces by about 4 seconds per kilometre off the updated VO2max estimate and left the sessions alone.

Daily Suggested Workouts do the opposite. They genuinely respond to what you did, using acute load, HRV status and training readiness, and they have no block and no race in mind. Seven days of zero training reads as “recovered” plus “anaerobic shortage” in the load focus gauge, so on day 2 back the watch offered a threshold session. The reasoning is internally consistent. It is also the last thing you want on your second run after illness.

Neither half of Garmin is dishonest about it. If you want the comparison across the whole field rather than the two scenarios here, our tested comparison of AI coaching apps covers how each one actually generates a block in the first place, which largely determines whether it can rebuild one.

The ChatGPT plan is the sharpest version of this problem

A plan you built with an LLM and pasted into a spreadsheet has no idea you missed anything. Ask it to adjust and it will confidently hand back a modified week, and in our attempts it consistently produced week 10 as “week 10 minus the missed sessions”, with the 29km long run intact. It has no running record of your last four weeks unless you paste them in, and it will not volunteer that a 24km to 32km jump after seven days off is a bad idea.

It will, though, do the arithmetic properly if you give it the constraint. Paste your last four weeks of actual volume from Strava and ask for a re-entry week capped at 80% of your last completed week, with the long run capped at 110% of your last completed long run, and tell it which of the taper or the goal time you’re willing to give up. That is the whole coaching decision, made by you, executed by the model. It works. It just isn’t the model adapting.

Three checks, five minutes

Open your plan and mark a session as missed. Note the total distance of the next 7 days before and after. Delete a whole week and do it again.

• If the next 7 days total is identical to within a kilometre, your app reschedules.
• If the date of your peak long run never moves no matter how much you skip, your app has a fixed skeleton and the “AI” is in the pace calculation only.
• If the goal pace changes but nothing else does, you have a calculator with a calendar attached.

None of that makes a fixed-skeleton plan useless. A Runna block executed at 90% compliance will get most recreational runners round a marathon faster than nothing will, and the pace targets are reasonable. What it means is that the judgement about cumulative load stays with you, permanently, and you should know that going in rather than discovering it in week 14 with a niggle in your Achilles and a taper that assumes eight weeks you never ran.

So set the rule now, while you’re healthy and nothing has gone wrong: after any absence of four days or more, the first week back is 80% of your last completed week, the long run goes up no more than 15% from the last one you actually finished, and the week you lost comes out of the build, not the taper. Write it at the top of the plan. The app won’t.