A Close Reading of the Most-Cited Habit Formation Study
If you’ve ever heard that “it takes 66 days to form a habit,” the number comes from one study: Lally et al. (2010), published in the European Journal of Social Psychology. It’s one of the most-cited empirical studies on habit formation, and its findings have been repeated in thousands of articles, books, apps, and TED talks.
The problem is that almost no one who cites the study has actually read it. And when you do read it, carefully, with an eye on what the data actually show, the picture is far less encouraging for the habit formation industry than the headline suggests.
Here’s what the study found, what it didn’t find, and what it means for anyone trying to change their behavior. (This article is a companion to our complete guide to behavior change, which covers the full evidence base.)
The Setup
Ninety-six university volunteers (mostly postgraduate students, mean age 27) were asked to choose a new healthy behavior, eating, drinking, or exercise, and perform it daily in the same context for 84 days. Examples: “eating a piece of fruit with lunch,” “drinking a bottle of water with lunch,” “running for 15 minutes before dinner.”
Each day, participants logged onto a website and reported whether they’d performed the behavior, then completed a modified version of the Self-Report Habit Index (SRHI). The researchers used 7 of the SRHI’s 12 items, the ones measuring automaticity specifically (e.g., “I do automatically,” “I do without thinking,” “I would find hard not to do”), excluding items about repetition history and identity. This gave an automaticity score ranging from 0 to 42.
The researchers then fit an asymptotic curve to each individual’s automaticity scores over time, using Mitscherlich’s law of diminishing returns. The idea: automaticity should increase rapidly at first, then slow down, eventually plateauing at a maximum. The curve parameters tell you the plateau height (how automatic it gets), the rate of change (how fast it gets there), and the time to reach 95% of the asymptote.
Behavior type breakdown: 27 chose eating, 31 chose drinking, 34 chose exercise, 4 chose other (e.g., meditation).
Finding #1: More Than Half the Participants Failed
This is the finding that never makes it into the popular summaries.
Of 96 participants, 14 dropped out. Of the remaining 82:
- For 12 participants, the statistical software couldn’t even fit the asymptotic curve after 100 iterations
- For 8 participants, the model produced a flat line, meaning their automaticity never increased at all
- For the remaining 62, the curve could be fit, but:
- 16 had poor fits (R²
- 4 had asymptote values below 21 (out of 42), meaning they showed some increase but never reached what the researchers considered habit-level automaticity
- 3 had unrealistically high modeled asymptotes (above 49 on a 42-point scale, the model was extrapolating beyond the data)
That left 39 participants, 48% of those with adequate data, for whom the asymptotic curve was a “good fit.”
Let that sink in. In a study specifically designed to form habits, with motivated volunteers who chose their own behaviors and were paid to participate for 12 weeks, more than half did not show the expected habit-formation pattern.
The researchers noted that the participants whose curves fit poorly had “typically carried out the behaviour fewer times during the study.” In other words, inconsistency (the thing that plagues everyone trying to build a new behavior) was associated with poorer habit formation outcomes, even under idealized conditions.
Finding #2: The “66 Days” Is a Median of the Successful Minority
The famous 66-day figure comes from the 39 participants with good curve fits. Among those 39:
| Parameter | Median | Range |
|---|---|---|
| R² (model fit) | 0.88 | 0.72 – 0.98 |
| Asymptote (max automaticity, out of 42) | 35 | 21 – 48 |
| Time to 95% of asymptote | 66 days | 18 – 254 days |
The range is 18 to 254 days. That’s not a minor spread, the slowest habit-former took 14 times longer than the fastest. And 254 days is the modeled estimate; the study only ran 84 days, so for participants whose modeled time-to-95% exceeded 84 days, the asymptote is a statistical projection, not something actually observed.
When someone tells you “it takes 66 days to form a habit,” they’re giving you the median of the 48% of people who succeeded, in a study of simple daily behaviors, extrapolated from a model that often ran beyond the actual observation period. That’s a lot of caveats.
Finding #3: Exercise Took 1.5x Longer (And Probably Longer Still)
Among the 39 participants with good fits:
| Behavior Type | N | Median Time to 95% Asymptote | Quartiles (Q1:Q3) |
|---|---|---|---|
| Eating | 10 | 65 days | 35–106 |
| Drinking | 15 | 59 days | 39–75 |
| Exercise | 13 | 91 days | 44–118 |
Exercise took 1.5 times longer than drinking and eating behaviors. The difference wasn’t statistically significant (p = 0.328), but the study explicitly noted it was “not powered for sub-group analyses”, meaning the sample was too small to detect a real difference even if one exists. The authors themselves highlighted this trend: “It is notable that the exercise group took one and a half times longer to reach their asymptote than the other two groups.”
Here’s the critical problem: the study only ran 84 days, but the median exercise asymptote was 91 days. That means more than half the exercise participants hadn’t plateaued when the study ended. Their “asymptote” is a model extrapolation, not an observed plateau. The third quartile was 118 days, meaning 25% of exercise participants were projected to take nearly four months, well beyond the observation window.
Compliance was also lower for exercise: 86% vs. 93% for drinking. And compliance was the only performance variable significantly correlated with model fit (r = 0.34, p = 0.035). Less consistent performance → worse model fit → less evidence of habit formation. Exercise is harder to do consistently, which makes it harder to form into a habit. This isn’t surprising, but it’s rarely acknowledged.
Finding #4: They Measured “Automaticity,” Not “Habit”
This is the most important and most overlooked finding in the paper.
The researchers used the Self-Report Habit Index to measure automaticity, a participant’s subjective sense that a behavior feels automatic, effortless, and unintentional. But automaticity and habit are not the same thing.
The gold-standard definition of a habit, from decades of animal and human research (Dickinson, 1985; Daw & O’Doherty, 2014), is a behavior that persists when its outcome is devalued. If you devalue the reward (e.g., make a food taste bad, remove the benefit of a behavior) and the person keeps doing it anyway, that’s a habit. If they stop, it’s a goal-directed behavior that merely felt automatic.
The SRHI cannot distinguish these. When a participant reports that doing 50 sit-ups “feels automatic,” they could mean:
- The decision to start feels automatic: “When I finish my coffee, I automatically think ‘time for sit-ups’ without deliberation.” This is initiatory automaticity: the cue-response link for getting started has been learned.
- The performance itself feels automatic: “I do 50 sit-ups without conscious effort or attention.” This would be true motor automaticity.
- The routine feels established: “This is just what I do now. It’s part of my day.” This is goal-directed automaticity: a subjective sense of routine that remains tied to the person’s goals and would stop if the goals changed.
The authors were fully aware of this ambiguity. In their discussion, they cited Wood and Neal (2007), who proposed that repeating complex behaviors in a consistent setting could develop goal-directed automaticity rather than habit, meaning the behavior becomes routinized and feels effortless, but remains flexible and tied to conscious goals.
The authors wrote explicitly:
“Using the SRHI in this study means that we have assessed the development of automaticity for performing these behaviours rather than specifically habit. New measures will be needed to disentangle these two forms of automaticity.”
This is an extraordinary admission that almost never appears in popular accounts of the study. The researchers are saying: We don’t actually know if the behaviors that scored high on automaticity were true habits. They might just be well-established routines.
For simple behaviors like drinking a glass of water, the distinction probably doesn’t matter much. The behavior is simple enough that genuine habit (cue → response with minimal deliberation) is plausible. But for exercise behaviors? The person still has to decide to start, overcome inertia, sustain physical effort, count reps, manage fatigue, and choose to continue rather than stop. That the behavior feels routine does not mean it’s a habit in any scientifically meaningful sense.
Finding #5: The “50 Sit-Ups” Example Is More Ambiguous Than It Appears
The study’s Figure 3 shows three example curves from participants with good model fits. One is “Doing 50 sit-ups after my morning coffee,” which shows a clean asymptotic curve reaching about 35 on the 42-point automaticity scale.
This is one of the most reproduced figures in the habit formation literature, and it creates the impression that even exercise behaviors reliably become automatic. But several things are worth noting:
It’s one cherry-picked example. The paper shows three individual curves out of 39 good-fit participants (out of 82 total). We don’t see the exercise participants whose curves didn’t fit, whose automaticity never increased, or whose asymptotes were still rising at day 84.
The asymptote of ~35 is hard to interpret. A score of 35 out of 42 means the participant averaged “agree” on statements like “I do automatically” and “I do without thinking.” But what does it mean to “agree” that you do 50 sit-ups “without thinking”? You don’t do 50 sit-ups without thinking, you do them without deciding whether to do them. The initiation is automatic. The execution is effortful. The SRHI conflates these.
The devaluation test would be informative here. If this participant’s doctor told them sit-ups were bad for their back, would they keep doing them anyway? If so, it’s a habit. If they’d stop, it’s a well-established routine masquerading as one. We have no way of knowing from the SRHI data.
50 sit-ups is more complex than “drinking water” but less complex than “exercising regularly.” It’s a bounded, specific, single-movement behavior performed in one context. It’s not the same as “going to the gym” or “maintaining a varied exercise routine”, the kinds of exercise goals that people actually care about. The gap between “50 sit-ups after coffee” and “a sustainable fitness practice” is enormous.
What the Study Actually Supports
To be fair to Lally et al., the study made genuine contributions:
- It killed the 21-day myth. Even the successful minority took a median of 66 days, with massive variation.
- It showed that missing a single day doesn’t derail habit formation. A missed opportunity reduced automaticity by less than half a point, and scores recovered quickly. This is practically useful and reassuring.
- It confirmed that the asymptotic curve fits individual-level data. Prior work (Hull, 1943, 1951) had only shown this for group averages. Lally et al. demonstrated it for individuals.
- It showed that simple behaviors in stable contexts can start to feel automatic. Drinking water after breakfast, eating fruit with lunch, these are plausible habit candidates, and the data supports that.
What the Study Does NOT Support
Here’s what the popular interpretation gets wrong:
“It takes 66 days to form a habit.” It took a median of 66 days among the 48% who succeeded, mostly for simple behaviors, with a range of 18 to 254 days. For exercise, the median was 91 days. For many participants, the “asymptote” was a model projection beyond the observation window.
“Any behavior can become a habit with enough repetition.” More than half the participants did not show the expected habit formation pattern. The study’s own authors acknowledged that complex behaviors may develop goal-directed automaticity rather than true habit. Verplanken (2006), cited in the paper, had already shown that “for the same number of repetitions a simple behaviour has a higher habit score than a complex behaviour.”
“Exercise can become a habit.” The study provides weak evidence for this claim. Exercise participants took longer, were less compliant, and the study wasn’t long enough to capture whether they’d truly plateau. The automaticity they reported may reflect routinization rather than genuine habit formation. The authors explicitly flagged this ambiguity.
“If you just stick with it long enough, it becomes effortless.” The study’s final sentence states: “creating new habits will require self-control to be maintained for a significant period before the desired behaviours acquire the necessary automaticity to be performed without self-control.” But for complex behaviors, the evidence suggests the “necessary automaticity” may never arrive, and the self-control requirement may be permanent.
The Implication Nobody Wants to Hear
If the most-cited habit formation study shows that:
- More than half of motivated people didn’t show the expected automaticity pattern for simple behaviors in 84 days
- Exercise takes 1.5x longer and the study wasn’t long enough to capture it
- The measurement tool can’t distinguish true habits from established routines
- And the authors themselves say they measured automaticity, not habit
…then the popular advice to “just build the habit” is far less grounded than it appears.
For simple, bounded behaviors in stable contexts, taking a vitamin, drinking water, eating a piece of fruit, habit formation is real and achievable. The Lally data supports this.
But for the behaviors most people actually want to change (exercising regularly, eating well, writing consistently, managing their finances) the evidence suggests these are goal-directed behaviors that require ongoing conscious engagement. They may feel more automatic over time as you develop routines and reduce decision fatigue. But they’re unlikely to become true habits in the scientific sense: behaviors that fire automatically in response to cues and persist even when their outcomes are devalued.
The practical implication is significant. If your target behavior is genuinely goal-directed, then the most important question isn’t “how do I make this automatic?” It’s “have I chosen a version of this behavior that I can sustain with ongoing conscious effort?” That shifts the focus from habit engineering to what behavioral scientists call person-behavior fit, finding the specific implementation of a behavior that works with your personality, preferences, constraints, and life circumstances.
The habit formation industry has built an empire on the assumption that any behavior can become effortless with enough repetition. The Lally study, the industry’s own flagship citation, suggests otherwise.
If you want to understand what actually works for lasting behavior change, including how to find the right version of a behavior for your personality, constraints, and life circumstances, I wrote an entire book on this: Real Change: Achieve Lasting Transformation. It covers the science of person-behavior fit and why matching beats forcing, every time.
References
Daw, N. D., & O’Doherty, J. P. (2014). Multiple systems for value learning. In P. W. Glimcher & E. Fehr (Eds.), Neuroeconomics: Decision making and the brain (2nd ed., pp. 393–410). Academic Press.
Dickinson, A. (1985). Actions and habits: The development of behavioural autonomy. Philosophical Transactions of the Royal Society of London B, 308(1135), 67–78.
Hull, C. L. (1943). Principles of behavior: An introduction to behavior theory. Appleton-Century-Crofts.
Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W., & Wardle, J. (2010). How are habits formed: Modelling habit formation in the real world. European Journal of Social Psychology, 40(6), 998–1009.
Verplanken, B. (2006). Beyond frequency: Habit as a mental construct. British Journal of Social Psychology, 45(3), 639–656.
Wood, W., & Neal, D. T. (2007). A new look at habits and the habit-goal interface. Psychological Review, 114(4), 843–863.
