The ZeroFit Blog

They Randomised 70,000 Reminder Moments. Timing Made No Consistent Difference.

Every app that nudges you believes the nudge can be tuned. Send it at the right hour, mention the weather, use the person’s first name, phrase it as an invitation rather than a scold. It is one of the most confidently held beliefs in the category.

A study published in August 2026 tested that belief across more than 70,000 randomised decision points, and found no message feature that consistently predicted who moved.

The design, and why it is unusually strong

The VALENTINE study embedded a microrandomised trial inside a larger randomised controlled trial. A hundred and eight adults each chose four daily time points. At each point, a tailored text was sent with 25% probability. Over the study, that produced 70,552 individual randomisations.

The outcome was step count in the sixty minutes after each decision point. The researchers then asked which message features moved that number: time of day, weekday against weekend, walking-focused against anti-sedentary content, and whether the message used the participant’s name, alongside participant age, sex and baseline activity. Weather was used to pick indoor wording when rain or snow was forecast, and time since enrolment split the analysis into three rehabilitation phases, but neither was tested as a moderator.

The finding, in the authors’ words, is that no text characteristics or participant characteristics consistently affected responsiveness (JMIR mHealth and uHealth, PMID 42580690, August 2026).

That design matters. A microrandomised trial randomises the intervention at every decision point, so each person serves as their own control several hundred times, and the trial as a whole accumulates more than seventy thousand of those points. That randomisation count powers the question of whether a text moved anybody at all. It does not power the comparisons between message types, which rest on 108 people split by the kind of tracker they wore and then by how far into rehabilitation they were, and the authors say the subgroup samples were small enough that the analysis should be read as hypothesis-generating rather than confirmatory.

What it does not say

Three things, and they need stating before anybody runs with the headline.

It does not say notifications don’t work. The parent trial found the tailored texts raised activity overall during the initiation phase. What could not be established was why, using any of the variables that were tuned. The messages worked; the tuning didn’t.

The population is not you. Participants were cardiac rehabilitation enrollees, average age 59.5, medically supervised and enrolled in a structured programme.

The outcome is narrow. Steps in the following sixty minutes captures whether a text produced an immediate walk. It says nothing about whether anything became a habit. And this was a secondary analysis, so the moderator questions were not what the trial was powered for.

The pattern this fits into

On its own it would be an interesting null. It is more than that because three other reviews point the same way, though one of them is a fourteen-study scoping review with no pooled estimate at all.

A meta-analysis included 87 eHealth trials targeting physical activity and pooled 53 of them. The pooled effect was small at SMD 0.22 (95% CI 0.13 to 0.31), and the authors tell readers to treat that number with caution: the funnel plot was badly asymmetric (Egger’s P<0.001), heterogeneity was high at I2=85%, and the trials at low risk of bias produced about half the effect, SMD 0.13. When the authors coded interventions by behaviour-change technique, the five most-used were feedback, goal setting, social support, self-monitoring and prompts. Interventions containing each of the five improved activity, with subgroup effects from SMD 0.21 to 0.30 and every confidence interval excluding zero. What the data could not show is that any one technique added anything over interventions that lacked it (Digital Health, PMID 42534183, July 2026). The population was older adults, mean age 66.4, with only four of the 87 trials recruiting for frailty and none for sarcopenia. The authors do not call the technique comparisons underpowered; what they flag is that techniques appearing in both arms were dropped from the analysis, and that their coding missed how often and how intensely each technique was delivered. Either way, this is an absence of evidence rather than evidence of absence.

A scoping review of just-in-time adaptive interventions, the technical name for nudges that fire when a moment of need is detected, found fourteen studies in people with diagnosed chronic conditions, only three of them randomised, and concluded the evidence is preliminary. Six of fourteen showed positive effects on activity, which is a vote count rather than a pooled estimate, and healthy people were excluded from all of it (JMIR mHealth and uHealth, PMID 42566750, August 2026).

And a synthesis of 36 goal-setting trials, in adults with chronic musculoskeletal pain rather than the general population, put the effect on physical activity at SMD 0.88 with an interval running from -0.32 to 2.07: an estimate too wide to act on, which the authors themselves call unclear (The Clinical Journal of Pain, PMID 42584484, August 2026). The goal-setting piece covers it.

Four lines, one message: the average effects are small, one of the four could not pin down an effect on activity at all, and none of them identify which ingredient does the work.

What to do with your own reminder

If you set a reminder for yourself, most of what this evidence tells you is relaxing.

Pick a time you can genuinely keep, and stop redesigning it. In the VALENTINE study, participants chose their own time points, and which of those points a text landed on did not consistently change how people responded. No cleverer version has been shown to work yet, and the trial’s authors are explicit that their null is not evidence that tailoring is ineffective: the only personalisation they varied was inserting the person’s first name, and they call for systems that learn a person’s preferences over time. The redesigning is itself a decision cost, which the decision-fatigue piece covers.

Use whichever cue you will actually notice, a time or a thing you already do. Tying the session to something that already happens, like your morning coffee, is one reasonable way in, though the one randomised trial to test routine cues against clock-time cues head to head found the two about equally effective, across 84 days in 192 adults and on an eating habit rather than exercise (British Journal of Health Psychology, PMID 33405284). What predicted a habit forming there was repeating the plan rather than which kind of cue was chosen, and a coffee you skip at the weekend is no steadier than an alarm you swipe away.

The one directional hint worth taking: content that handed people an action trended better than content that reported a problem. Telling somebody to go and do something beat telling them they had been sitting. That trend was weak and inconsistent, so treat it as a hint.

For what it is worth, ZeroFit currently sends nothing. Notification work has not been built, and this research is a reasonable argument for building the simple version first rather than the clever one.

The larger question this sits inside, including what the published retention numbers look like, is in do fitness apps work.

When you do open it, ZeroFit asks two questions and hands you a session. No setup to redo. Get beta access.

ZeroFit content is educational and not medical advice. If you have an injury, a health condition, or you’re pregnant, check with a qualified professional before starting any new movement routine.