The 66-day habit number is a median of 39 people

You have probably corrected somebody about this. I have. Somebody says 21 days to build a habit, and you get to say the 21 comes from a 1960 plastic surgery book, and the real figure is 66 days, from a study at UCL.

That correction is right about the 21. It is doing something slightly dishonest with the 66.

I read the paper, all 12 pages, pulled as a PDF from a Portuguese university repository after Wiley and UCL both refused me. Link at the bottom.

96 people started. 82 gave enough data to analyse. Five separate exclusions cut that to 39, and the 66 days is the median of the 39 who were left. Not one of the habit-app pages I read while writing this says 39.

Everyone reports 96. The number came from 39

I checked three habit-tracking apps that cite this study on their own blogs, on 3 September 2026. Every one prints the participant count, and every one prints 96. TaskCoach: "Phillippa Lally's 2010 study of 96 people". Keelify: "tracked 96 people for 12 weeks and observed a range from 18 to 254 days".

96 is the number who chose to take part. The results section is blunt about what came next:

On the next page: "Table 2 shows the values for the curve parameters for the 39 participants for whom the model was a good fit." That table is where the median of 66 and the range of 18 to 254 live. Both belong to the same 39 people, so quoting them with n=96 attached puts a denominator on them 2.5 times too big.

Keelify goes further: "Half of Lally's participants reached automaticity in fewer than 66 days." Half of the 39 modelled well did. Of the 96 who signed up, that is at most 20 people, or 21 percent, which turns the sentence around.

Who is actually in the 66

0

chose to take part, the figure every citing page prints

0

gave enough data to analyse at all

0

the median of 66 days is computed from these

The five filters, and what each one removed

The drop from 82 to 39 was five separate exclusions. I added them up, because the paper never does.

• 12 participants where SPSS "was unable to find an optimal solution after 100 iterations". A software convergence failure.
• 8 where the model returned zero for the b constant, meaning a flat line: these people "did not increase their automaticity ratings over the study period".
• 16 whose R² came in under 0.7, a threshold the authors set by eyeballing the histogram of fits.
• 4 whose modelled plateau was below 21, the cutoff for calling something a habit at all.
• 3 whose modelled plateau came out above 49, on a scale whose real maximum is 42.

12 plus 8 plus 16 plus 4 plus 3 is 43, and 82 minus 43 is 39. It reconciles, which is a good sign for the paper and a bad one for anyone quoting it loosely.

Look at the middle two. The 8 flat lines repeated a behaviour daily for 12 weeks and reported no rise in automaticity at all. The 4 low plateaus rose and never reached a level the authors would call a habit. Those 12 are a finding, removed before the headline number is computed.

And the filters do not remove people at random. Table 1 of the paper breaks compliance down by group, and the pattern is one direction all the way. The 39 who survived had a median compliance of 88 percent. The 16 with poor fits: 60. The 8 flat lines: 63. The 4 low plateaus: 61. Weight those three by size and you get 61 percent against the survivors' 88, a gap of 27 points, and the paper's own Mann-Whitney superscripts mark it significant.

So the filters preferentially strip out the people who did not do the thing. Repetitions tell the same story: a median of 63 among the good fits against 40 among the flat lines, 1.6 times as many. 66 days is the median of the compliant, and the compliant are who is left, because a curve only fits people who generated one.

The paper says the adjacent thing plainly in the discussion:

Half. Among postgraduate volunteers paid £30 who chose their own behaviour, often something as small as a glass of water with lunch. Worth holding next to any claim that a daily language app goes automatic on a schedule, because the stall people report around week 6 looks a lot like the flat lines this study dropped.

The study stopped on day 84

Participants were asked to perform the behaviour every day for 84 days. That is the whole observation window. So how does a study that runs 84 days report someone taking 254?

By modelling it. The authors fitted Mitscherlich's law of diminishing returns to each person's daily scores, y = a − be^−cx, then solved for the day the curve would reach 95 percent of its own asymptote. If someone was still climbing on day 84, the curve keeps climbing after the data stops, and the answer comes out of the equation rather than anybody's life.

The paper is upfront about it. The high modelled plateaus, it says, "were often due to participants not reaching a plateau within the 84 days and the fitted curve estimating a high value for the asymptote." And of the poor-fitting high performers: "It is probable that these individuals were relatively slow in forming their habits and would have reached a plateau if the recording had continued for longer." Probable. Their word.

So I worked out how much of the 254 anyone watched. The paper prints the minimum rate constant, c = 0.010, and the formula. Solving backwards from a 254-day answer at that rate gives a b/a ratio near 0.63, putting that person at roughly 73 percent of their modelled asymptote on day 84, the last day they were seen. The remaining 170 days are the curve talking, 67 percent of the figure. That reconstruction assumes the slowest person sat at Table 2's minimum rate constant, and the table reports each column separately, so those may be two different people. The conclusion survives either way, because observation stopped on day 84.

The reported range against the 84 days of observation

18 fastest 39 Q1 66 median 84 last day observed 102 Q3 254 slowest

One flag for anyone reproducing this. The paper prints the time-to-95% equation as [ln(a/20b)]/c. Feed it the Table 2 medians (a = 35, b = 30, c = 0.042) and you get −67.7 days. Invert the fraction to ln(20b/a)/c and you get +67.7, next to the reported median of 66. The published equation is upside down, probably a typesetting slip, since the corrected form reproduces the reported values.

None of this makes the study bad. It is careful work that says clearly what it did. It describes 39 curve fits with a modelled tail, which is a different object from "it takes 66 days to form a habit".

The author was asked in January and said no

Phillippa Lally is now a senior lecturer at the University of Surrey. In January 2026 her own university put the question to her.

She goes on: "what we really showed was that how long habit formation takes is highly variable." Earlier in the same piece she calls it "sometimes frustrating when a finding, for example the statement that 'it takes 66 days to form a habit', is taken out of context."

The spread is the result, and the 66 is the summary statistic that ate it. 254 over 18 is a 14-fold gap, and even the interquartile range, 39 to 102 days, is a factor of 2.6.

One more claim gets repeated as settled. The exercise median was 91 days against 59 for drinking, which the discussion calls "one and a half times longer". Check the test: Kruskal-Wallis, p = 0.328, on cells of 10, 15 and 13, and the paper says that finding "should not be considered to be definitive" because the study was not powered for it. TaskCoach meanwhile prints "a median of around 20 to 30 days" for simple drink behaviours. The paper's drinking median is 59, and that range appears nowhere in it.

The finding your streak counter ignores

Here is the part of the paper habit apps quote least, for a reason that will be obvious shortly. The authors found 140 missed opportunities across 55 participants: a skipped day after three consecutive days of doing it. Automaticity fell 0.29 points on a 0-42 scale. When people resumed the next day, their score sat 0.55 points above where it was before the miss, against 0.79 for three unbroken days.

0.55 against 0.79, and the gap was not significant on a Wilcoxon signed rank test. The paper's conclusion: "missing a day resulted in a non-significantly lower SRHI score the following day, but there were no longer-term costs associated with a single omission." When in the process the miss happened made no difference either.

Now the awkward bit, which I would rather write myself than have somebody point out.

Verbamor has a streak. I looked up the code before writing this. It advances when yesterday was your last review day, and otherwise sets you back to 1. Miss a Tuesday and a 40-day streak becomes a 1, which is the opposite of what this paper's missed-opportunity analysis found: you lost 0.29 points of automaticity out of 42, and got it back the next day.

I will not pretend that counter comes from evidence. It is there because a visible run is motivating and it was easy to build. The evidenced part of the app is the scheduling: FSRS choosing which cards come back today, and retrieval rather than rereading when they do. A card you owe on Thursday is still owed on Friday, whatever the streak says.

So if you miss a day, the honest reading of the paper being sold to you is: do it again tomorrow, and don't treat the counter as a measurement. A missed week is its own problem, with cards piling up and real forgetting behind them, and it deserves a plan. A broken Tuesday deserves a Wednesday.

And notice the shape. Automaticity climbs steeply early and flattens late, so the last 5 percent costs more days than the first 60. If you are 30 days into Spanish and it still takes effort, you are on the curve, in the stretch 39 postgraduates in London sat through too.

---

Sources

Every load-bearing claim Verbamor makes is traced to its paper on the research page.