Onboarding and habit formation: the science of making products stick

Most sign ups never become habits

Most people who sign up for a product never come back. That is not a bug. It is the default. The products that keep their customers are not always better. They have designed the path from sign up to habit. They know which early actions go with long term use, and they help every new customer reach those actions quickly.

Two things are easy to confuse here. Onboarding is the first few sessions, the part you can design screen by screen. A habit is what happens months later, when the customer opens the product without deciding to. Finishing onboarding is not the same as forming a habit, and good products work on both.

This article takes the frameworks in turn: finding the action that predicts retention, removing the steps in front of it, the Hook model, how long habits really take, habit stacking, unfinished tasks, head starts, the goal gradient, and the Irish rules on onboarding emails. One invented business runs through all of it: Tallybook (tallybook.example), an invoicing app for tradespeople in Limerick, €12 a month after a 14 day free trial. Every number about Tallybook is made up so I can show each step.

Activation is the action that predicts staying

The "aha moment" is the point where a new customer first gets the value the product exists for. In practice teams look for an activation metric: one early action that customers who stay tend to have done, and customers who leave tend not to have done.

Growth teams like to tell stories of a single magic number, a certain number of friends within a certain number of days, a certain number of messages sent. An earlier version of this article had a table of them. I could not trace any of those figures to a source I could check, so they are gone. They would not help you anyway. Your activation metric is in your own data, and it has to be found there.

The method is a cohort analysis. Take everyone who signed up in one period, a cohort. For each candidate action, split the cohort into those who did it in their first week and those who did not, and compare how many in each group are still customers later.

The comparison uses two rates and their difference:

$$\Delta = r_{\text{did}} - r_{\text{not}}$$

Here $r_{\text{did}}$ is the share of customers who did the action and are retained, $r_{\text{not}}$ the share of those who did not and are retained, and $\Delta$ the difference in percentage points. Both are simple rates, see Foundations: ratios and rates, and the difference between a percentage change and percentage points is in Foundations: percentages.

Worked example: Tallybook's candidate actions

In March, 2,000 tradespeople started a Tallybook trial. Sixty days later, 560 of them were paying customers, 28%. Tallybook looks at five things a person could do in the first seven days.

Action in the first week	Did it	Retained if done	Retained if not	Difference
Sent a first invoice	640	55.0%	15.3%	39.7 points
Connected a bank account	400	50.0%	22.5%	27.5 points
Added a logo	700	36.0%	23.7%	12.3 points
Completed the profile	1,200	32.0%	22.0%	10.0 points
Watched the product tour	900	29.0%	27.2%	1.8 points

Take the first row step by step. Of the 640 who sent an invoice, 352 were retained: $352 / 640 = 0.550$. Of the 1,360 who did not, 208 were retained: $208 / 1{,}360 = 0.153$. So $\Delta = 0.550 - 0.153 = 0.397$, about 40 points. The two groups add up to the whole cohort: $352 + 208 = 560$.

Is that more than noise? The two proportion z test answers it:

$$z = \frac{p_1 - p_2}{\sqrt{p(1-p)\left(\frac{1}{n_1} + \frac{1}{n_2}\right)}}$$

Here $p_1$ and $p_2$ are the retention rates of the two groups, $n_1$ and $n_2$ their sizes, and $p$ the pooled rate of both groups together. The test is explained in Foundations: hypothesis tests.

The pooled rate is $560 / 2{,}000 = 0.28$. The standard error is $\sqrt{0.28 \times 0.72 \times (1/640 + 1/1{,}360)} = 0.0215$. So $z = 0.397 / 0.0215 = 18.4$. Far beyond chance.

The product tour is the useful surprise. Watching it barely separates stayers from leavers. Sending a real invoice does.

A correlation is a candidate, not a cause

Here is the trap. Tradespeople who send an invoice in week one may simply be the ones who had a job to bill that week. They would have stayed anyway. The table shows which action goes with retention. It does not show that pushing people towards the action creates retention. Foundations: correlation explains why.

The way to find out is to randomise the push, not the action. Tallybook gives half of the April cohort, chosen at random, a prompt on the first screen: "Bill the job you just finished", with a draft invoice already started. The other half sees the usual dashboard. Each group has 1,000 people.

In the prompted group, 44% send an invoice in week one, against 32% in the other. Retention at 60 days is 31.6% against 28.0%. The pooled rate is 0.298, the standard error $\sqrt{0.298 \times 0.702 \times (1/1{,}000 + 1/1{,}000)} = 0.0205$, so $z = 0.036 / 0.0205 = 1.76$, a two sided p value of about 0.08. Promising, not proven.

How many people would the test need? For 28% against 31.6% at 5% significance and 80% power:

$$n = \frac{(z_{\alpha/2} + z_{\beta})^2 \left[p_1(1-p_1) + p_2(1-p_2)\right]}{(p_1 - p_2)^2}$$

Here $n$ is the number of people per group, $z_{\alpha/2}$ is 1.96 for 5% significance, $z_{\beta}$ is 0.84 for 80% power, and $p_1$ and $p_2$ are the two retention rates. Foundations: sample size goes through it.

With the numbers in: $n = 2.80^2 \times (0.2016 + 0.2161) / 0.036^2 = 7.84 \times 0.4177 / 0.001296 \approx 2{,}527$. So about 2,530 per group, roughly two and a half months of sign ups. The honest plan is to keep the test running, not to declare victory in April.

import numpy as np
import pandas as pd
from statsmodels.stats.proportion import proportions_ztest

# Tallybook, simulated: 2,000 trial starts, one row each
rng = np.random.default_rng(42)
n = 2000
users = pd.DataFrame({
    'sent_invoice': rng.random(n) < 0.32,
    'connected_bank': rng.random(n) < 0.20,
    'added_logo': rng.random(n) < 0.35,
    'completed_profile': rng.random(n) < 0.60,
    'watched_tour': rng.random(n) < 0.45,
})
# Invented: retention depends mostly on invoicing and the bank connection
score = (0.12 + 0.38 * users['sent_invoice'] + 0.15 * users['connected_bank']
         + 0.04 * users['added_logo'] + 0.01 * users['watched_tour'])
users['retained_60d'] = rng.random(n) < score.clip(0, 0.95)

rows = []
for action in ['sent_invoice', 'connected_bank', 'added_logo',
               'completed_profile', 'watched_tour']:
    did = users[users[action]]['retained_60d']
    did_not = users[~users[action]]['retained_60d']
    z, p = proportions_ztest([did.sum(), did_not.sum()], [len(did), len(did_not)])
    rows.append({'action': action, 'share_did': users[action].mean(),
                 'retained_if_done': did.mean(), 'retained_if_not': did_not.mean(),
                 'difference': did.mean() - did_not.mean(), 'p_value': p})

print(pd.DataFrame(rows).sort_values('difference', ascending=False).round(3).to_string(index=False))
print('These are associations. Randomise a prompt before you call any of them a cause.')

Time to first value is a curve worth moving

If an action predicts retention, the next question is how soon people reach it. Someone who sends a first invoice in the same session has felt the value before the day's distractions arrive. Someone who gets there in week two may already have half forgotten the product.

The retention of a whole cohort is a weighted average across these groups:

$$R = N \sum_i s_i \, r_i$$

Here $R$ is the number of retained customers, $N$ the number of sign ups, $s_i$ the share of sign ups whose first invoice falls in time band $i$, and $r_i$ the retention rate of that band. The weights $s_i$ add up to one. It is the same sum as an expected value, see Foundations: expected value.

Worked example: moving Tallybook's first invoice earlier

The retention rates by time band below are the example's assumptions.

The horizontal axis shows how long after sign up the first invoice was sent. The vertical axis shows the share still paying 60 days later. The takeaway is the slope: each band is worth noticeably less than the one before.

Today the shares across the five bands are 15%, 30%, 25%, 18% and 12%. So $R = 2{,}000 \times (0.15 \times 0.50 + 0.30 \times 0.38 + 0.25 \times 0.25 + 0.18 \times 0.14 + 0.12 \times 0.04)$. That is $2{,}000 \times (0.075 + 0.114 + 0.0625 + 0.0252 + 0.0048) = 2{,}000 \times 0.2815 = 563$ retained.

Suppose a new first screen moves the shares to 35%, 30%, 17%, 10% and 8%. The sum becomes $0.175 + 0.114 + 0.0425 + 0.014 + 0.0032 = 0.3487$, so 697 retained, 134 more per monthly cohort. At €12 a month that is about €1,600 more in monthly revenue from each cohort while they stay.

The assumption doing the work: that someone moved into an earlier band takes on that band's retention rate. That is the causal claim from the previous section again, and it holds only as far as the randomised test confirms it. Use the sum to decide whether a change is worth testing, never to report its result.

Remove the steps between sign up and first value

Once you know the action, clear the path to it. BJ Fogg's behaviour model says a behaviour happens when motivation, ability and a prompt come together at the same moment, and when it does not happen, at least one of the three is missing.¹ Motivation is hard to raise. Ability, how easy the step is, is entirely in your hands.

The practical moves follow from that. Defer anything that is not needed before the first value: the profile, the logo and the preferences can wait. Shorten the key action itself. Prompt at the moment the person can act.

There is one subtlety, and it comes from the IKEA effect. Michael Norton, Daniel Mochon and Dan Ariely found that people valued a storage box they had assembled themselves more than an identical box handed to them built: builders bid on average 78 US cents, non builders 48 cents. The effect disappeared when participants built something and then destroyed it, or failed to finish it.² So filling a product with sample data shows value quickly but gives the person nothing of their own. Asking them to build everything from scratch gives them effort that may never finish. The middle path is a small piece of real work that is certain to succeed.

Worked example: Tallybook's first screen, before and after

Before, a new Tallybook user had to:

1. Confirm their email address.
2. Fill in a business profile with eight fields.
3. Upload a logo.
4. Watch a two minute tour.
5. Open "New invoice", add a client, add line items, and send.

After the change, a new user:

1. Lands on "Bill the job you just finished", with their name and trade already filled in from sign up.
2. Types the client's name and email, and the price of the job.
3. Presses send.

The confirmation email, the profile and the logo move to a checklist that appears after the first invoice is out. The tour becomes optional. Three steps instead of five screens, and the one piece of effort is the person's own real job, finished. That is the IKEA effect used honestly: the effort ends in success, and in money owed to them.

The Hook model is a map, not evidence

Nir Eyal's Hook model, from his book Hooked, describes a loop in four phases: a trigger, an action, a variable reward and an investment that makes the product more valuable for the next pass.^{3 4} It is a useful way to sketch why someone would come back. It is a design framework from a book, not the result of a study, so I use it as a checklist, not as proof that something will work.

Each phase in one line. The trigger is the cue: external (an email, a notification) or internal (a feeling or situation the person has learned to pair with the product). The action is the simplest behaviour done in anticipation of a reward. The reward varies, which holds attention. The investment is something the person puts in, data, content or effort, that improves the next loop.

The variable reward is the part most often oversold. The laboratory research behind it is narrower than the slot machine comparison suggests, and a reward that is a gamble can build exactly the kind of habit people later regret. The Hook model article goes through what the studies measured and where the line sits between a habit and a hook in the bad sense.

Worked example: Tallybook's hook, written out

1. Trigger. External: a notification when a client pays. Internal, the one to earn: the Friday worry about who still owes money.
2. Action. Open Tallybook and look at the list of open invoices. One tap from the notification.
3. Variable reward. Which invoices have been paid this week genuinely varies. Nothing is manufactured. The variety is real life.
4. Investment. Every client and line item saved makes the next invoice faster to send.

Every phase serves the tradesperson. If one of them only served Tallybook's dashboard, that would be the phase to cut.

Habits take months, not 21 days

The most quoted figure in habit design is 66 days. It comes from a study by Phillippa Lally and colleagues: 96 volunteers chose an eating, drinking or activity behaviour and did it daily in the same context, such as after breakfast, for 12 weeks, rating each day how automatic it felt. 82 gave enough data. For 62 of them a curve fitted that rises quickly and then levels off. The time to reach 95% of each person's plateau ranged from 18 to 254 days, and missing a single day did not materially affect the process.⁵ The median was 66 days.⁶ Gardner, Lally and Wardle note that the popular 21 days seems to come from anecdotes about patients adjusting to plastic surgery, not from habit research.⁷

A 2024 systematic review by Ben Singh and colleagues gathered 20 studies with 2,601 participants. The four that reported a time to habit gave medians of 59 to 66 days and means of 106 to 154 days, with individuals ranging from 4 to 335 days. Most of the studies had a high risk of bias.⁶

Two cautions before you apply this to a product. These were daily health behaviours chosen by the participants. None of them measured an app, a purchase or a notification. And the curve is driven by repetitions, not by the calendar.

Worked example: how many repetitions does Tallybook get?

A volunteer in Lally's study, acting daily, gets 84 repetitions in 12 weeks. A Tallybook customer who invoices once a week gets 12 in the same 12 weeks, a seventh as many. If a habit needs anything like the repetitions the daily behaviours got, the invoicing habit will take far longer than 66 days in calendar time. That is my inference, not a finding.

So Tallybook measures habit over months, not over the trial. And it looks for a behaviour that repeats more often than invoicing: checking who has paid, which a busy tradesperson might do most days.

Habit stacking ties the product to a routine that exists

James Clear's habit stacking pairs a new habit with a current one, in the form "After [current habit], I will [new habit]". Clear calls it a special form of an implementation intention and credits the method to BJ Fogg's Tiny Habits programme.⁸ The existing routine becomes the cue, so the person does not have to remember.

The evidence underneath is on implementation intentions, plans of the form "when situation X arises, I will do Y". Peter Gollwitzer and Paschal Sheeran's meta analysis of 94 independent tests found a medium to large effect on reaching goals, $d = 0.65$.⁹ That is a standardised effect size, not a promise that follow through doubles. Gardner, Lally and Wardle make the same practical point: a cue inside an existing daily routine is a convenient, stable starting point for a new habit.⁷

For a product, that means not guessing when to send a reminder. Ask.

Worked example: Tallybook asks when the paperwork happens

During onboarding, after the first invoice, Tallybook asks one question: "When do you usually do your paperwork?" The options are "After the last job of the day", "Friday afternoon" and "Sunday evening", plus "I'd rather not get reminders".

A plumber in Limerick picks the first one. Her plan reads: "After I finish the last job of the day, I will send that job's invoice from my phone." Tallybook sends one reminder at the time she usually finishes, which she sets, and never more than one a day.

To know whether the question helps, Tallybook asks it of a random half of new customers and gives the other half a standard 6pm reminder. It compares how many weeks out of the first 12 each customer sent at least one invoice. The question costs nothing to ask and the test is easy to run.

Unfinished tasks pull people back, though not for the reason usually given

Onboarding checklists often cite the Zeigarnik effect: the claim, from Bluma Zeigarnik's experiments in 1927, that people remember interrupted tasks better than finished ones. Maria Ovsiankina, also working with Kurt Lewin, reported a related effect in 1928: people tend to go back and resume interrupted tasks when they get the chance.¹⁰

A 2025 meta analysis by Romain Ghibellini and Beat Meier pulled together 59 publications. It found no memory advantage for unfinished tasks, but a general tendency to resume them. Its conclusion: the resumption effect is a general tendency, the memory effect lacks universal validity.¹⁰

That changes the design brief slightly. A checklist does not work because an open item nags at memory. If it works, it is because people like to go back and finish what they started, when the unfinished thing is in front of them and finishing is easy. So the checklist has to be visible where they return, and each item has to be quick.

Worked example: does Tallybook's checklist bring people back?

Tallybook shows the remaining setup items (confirm email, add logo, connect bank) on the home screen. To measure resumption, it looks at customers who left setup unfinished, and randomises whether the checklist is shown.

Of 1,000 without the checklist, 240 returned to finish at least one item within seven days, 24.0%. Of 1,000 with it, 310 did, 31.0%. With the same z test as before, the pooled rate is 0.275, the standard error $\sqrt{0.275 \times 0.725 \times 0.002} = 0.0200$, and $z = 0.070 / 0.0200 = 3.51$, a p value below 0.001. The checklist brings people back. Whether the finished items then raise 60 day retention is the next question, measured the same way.

A head start makes the finish feel closer

Joseph Nunes and Xavier Drèze ran an experiment in a Los Angeles car wash in 2004. The business handed out 300 loyalty cards. One version needed eight stamps for a free wash. The other needed ten, with the first two stamped as a special promotion. The economics were identical. Only 19% of customers with the first card claimed the free wash, against 34% with the second. The second group also came back faster.¹¹

They called it the endowed progress effect. Their explanation matters, because an earlier version of this article had it wrong. The head start reframes the task as one that has been started and is incomplete, rather than one not yet begun, and the effect appears to depend on that perception of completion, not on a desire to avoid wasting the stamps already given.¹² So it is not sunk cost. The reason given for the head start and the units progress is counted in also changed the size of the effect.¹²

For onboarding, the honest version counts steps the person really has done. Creating the account and confirming the email are real steps. Shown as done, they turn "six things to do" into "four of six left". Invented progress, a bar that starts at 20% for nothing, is the version to avoid.

Worked example: Tallybook's checklist with a true head start

Tallybook tests two checklists, each on 1,000 new customers chosen at random:

1. A: four items, none ticked (logo, bank, first invoice, payment reminder).
2. B: six items, with "Create your account" and "Confirm your email" already ticked, because the person has done both.

Within 14 days, 220 of group A finish every item, 22.0%, and 290 of group B, 29.0%. The relative lift is $(0.29 - 0.22) / 0.22 = 0.318$, about 32%, see Foundations: percentages. The pooled rate is 0.255, the standard error $\sqrt{0.255 \times 0.745 \times 0.002} = 0.0195$, and $z = 0.070 / 0.0195 = 3.59$, a p value of about 0.0003.

Then the check that matters: does group B also retain better at 60 days, or did it just tick boxes? A checklist that rises while retention stays flat is decoration.

Effort speeds up as the goal gets closer

The goal gradient hypothesis was proposed by the behaviourist Clark Hull in 1932. In 1934 he found that rats in a straight alley ran faster as they got closer to the food. Ran Kivetz, Oleg Urminsky and Yuhuang Zheng tested it on people in 2006, with a real café loyalty card: customers bought coffee more often the closer they were to the free one, with the time between purchases falling by 20% on average. A card with two "bonus" stamps already filled in also made people finish faster, and customers who accelerated more were more likely to stay in the programme.¹³

Their model puts effort as a function of the share of the original distance still to go:

$$d = \frac{\text{steps remaining}}{\text{steps in total}}$$

Here $d$ is the goal distance, from 1 at the start to 0 at the finish. As $d$ falls, effort rises. It is a plain ratio, see Foundations: ratios and rates.

That is why "one step left" is stronger than a percentage. It names the small remaining distance. And it is why a long process works better broken into short stages, each with its own finish.

Worked example: a coffee cart card in Galway

Canal Cup (canalcup.example) is an invented coffee cart in Galway with a ten stamp card. At the first stamp, $d = 9 / 10 = 0.9$. At the eighth, $d = 2 / 10 = 0.2$.

Suppose a regular comes every 5.0 days at the start of the card and every 4.0 days near the end. The change is $(4.0 - 5.0) / 5.0 = -0.20$, 20% faster, the same size as the café average in the study. The cart owner can see this in her own till data: plot the gap between visits against the stamp number. If the gap shrinks near the end, the card works. If it does not, the card is a discount and nothing more.

For Tallybook the same idea is a line on the home screen: "One step left: connect your bank so payments match themselves." The step has to be real, and it has to be worth doing.

Onboarding emails and the Irish rules

Onboarding runs on email, so the law comes into it. The rules on marketing email are in regulation 13 of the ePrivacy Regulations 2011 (S.I. No. 336 of 2011): unsolicited marketing email to an individual needs their consent, with an exception for existing customers usually called the soft opt in.¹⁴ The regulator is the Data Protection Commission (DPC).

Not every onboarding email is marketing. The DPC describes direct marketing as an organisation attempting to promote a product or service, or to get you to ask for more information about one.¹⁵ A confirmation, a reminder or an update to the terms promotes nothing. The hard case is the mixed message, and I could not find DPC guidance on it. Britain's data protection regulator says that if a service message has a promotional element, even a minor one, the whole message counts as direct marketing, and that a message which actively promotes or encourages people to use a particular service, special offer or upgrade is likely to be direct marketing, while a neutral, informative tone points the other way.¹⁶ I read the Irish rule the same cautious way.

The soft opt in is where a trial product has to be careful. Regulation 13(11) covers email details you obtained from a customer in the context of a sale, used to market your own products or services of a kind similar to what you sold, with a clear and free chance to object when the details were collected and in every message after, and with the sale no more than 12 months back.¹⁴ The DPC adds that you must not send electronic marketing to a prospective customer who does not complete a purchase.¹⁷ A free trial is not plainly a sale, and the article on consent and email capture explains why that question is still open. So a trial product asks for consent at sign up, with an unticked box, and relies on the soft opt in only once the customer has paid. Irish law also has a rule for business addresses: an email to an address that reasonably appears to be used mainly for someone's business, about that business only, falls outside the consent requirement for individuals.¹⁴ Most tradespeople use one address for everything, so I would not build a plan on it.

Worked example: sorting Tallybook's five onboarding emails

1. "Confirm your email address." A service message.
2. "Your first invoice was delivered to the client." A service message, it reports on the account.
3. "Your trial ends in three days. Here is what happens next, and how to cancel." A service message if it is neutral, like the British regulator's example of a neutrally worded renewal notice.¹⁶ Nothing in it promotes anything.
4. "Upgrade now and get your first month at €9." Marketing. During the trial Tallybook sends it only to people who ticked the box for product news at sign up. Once someone has paid, and was offered an opt out on the payment screen, the soft opt in covers Tallybook's own similar products, and every email carries an unsubscribe link.
5. "Three ways to get paid faster with Tallybook Pro." Marketing, because it promotes an upgrade. Same conditions as the fourth.

The flow runs from the draft at the top. The first question sorts service messages from marketing. Only marketing goes on to the second question, about consent or the soft opt in, and every marketing email that is sent carries a way to opt out. The safe habit is to keep service messages plainly informative and put any promotion in a separate email that follows the marketing rules.

An onboarding audit, in order

Putting the sections together, the order of work is:

The flow runs top to bottom, from finding the action to measuring the result. The first two steps decide what to aim for. The middle steps clear the path and keep people coming back. The last two keep it lawful and honest about whether any of it worked.

Retention is designed, and then measured

The difference between products that become habits and products that are forgotten is rarely luck. It is the path from sign up to the first real value, and what brings people back after that.

Find your activation action in your own data, then test it. Remove the steps in front of it. Use the Hook model as a map. Expect habits to take months. Tie reminders to routines people already have. Show unfinished tasks where people return, count only true progress, and name the last step. Keep service emails neutral and marketing emails within the rules.

None of this is a trick if every part is true and serves the customer. It is how people actually behave, used in their favour. I have been applying these frameworks to SaaS onboarding and user retention for over a decade. The code in this article runs on invented data, so adapt it to your own before you trust any result.

Footnotes

1. BJ Fogg, Fogg Behavior Model, B=MAP, behaviormodel.org, https://www.behaviormodel.org/

2. Michael I. Norton, Daniel Mochon and Dan Ariely, The "IKEA effect": when labor leads to love, Harvard Business School working paper 11 091, 2011 (published in the Journal of Consumer Psychology, 2012), https://www.hbs.edu/ris/Publication%20Files/11-091.pdf

3. Nir Eyal with Ryan Hoover, Hooked: How to Build Habit Forming Products, Portfolio Penguin, https://www.penguin.co.uk/books/278201/hooked/9780241184837

4. Nir Eyal, The Hooked Model: How to Manufacture Desire in 4 Steps, nirandfar.com, https://www.nirandfar.com/how-to-manufacture-desire/

5. Phillippa Lally, Cornelia H. M. van Jaarsveld, Henry W. W. Potts and Jane Wardle, How are habits formed: modelling habit formation in the real world, European Journal of Social Psychology 40(6), 998 to 1009, 2010, doi 10.1002/ejsp.674, record and abstract at Crossref, https://api.crossref.org/works/10.1002/ejsp.674

6. Ben Singh, Andrew Murphy, Carol Maher and Ashleigh E. Smith, Time to form a habit: a systematic review and meta analysis of health behaviour habit formation and its determinants, Healthcare 12(23), 2488, 2024, https://pmc.ncbi.nlm.nih.gov/articles/PMC11641623/

7. Benjamin Gardner, Phillippa Lally and Jane Wardle, Making health habitual: the psychology of habit formation and general practice, British Journal of General Practice 62(605), 664 to 666, 2012, https://pmc.ncbi.nlm.nih.gov/articles/PMC3505409/

8. James Clear, How to build new habits by taking advantage of old ones (habit stacking), jamesclear.com, https://jamesclear.com/habit-stacking

9. Peter M. Gollwitzer and Paschal Sheeran, Implementation intentions and goal achievement: a meta analysis of effects and processes, Advances in Experimental Social Psychology 38, 69 to 119, 2006, University of Konstanz repository, https://kops.uni-konstanz.de/entities/publication/2e749bfb-8533-437c-8203-7e788c910c5f

10. Romain Ghibellini and Beat Meier, Interruption, recall and resumption: a meta analysis of the Zeigarnik and Ovsiankina effects, Humanities and Social Sciences Communications 12, 962, 2025, https://www.nature.com/articles/s41599-025-05000-w

11. Joseph C. Nunes and Xavier Drèze, Your Loyalty Program Is Betraying You, Harvard Business Review, April 2006 (the exhibit "The effect of a jump start"), https://hbr.org/2006/04/your-loyalty-program-is-betraying-you

12. Joseph C. Nunes and Xavier Drèze, The endowed progress effect: how artificial advancement increases effort, Journal of Consumer Research 32(4), 504 to 512, 2006, abstract on EconPapers, https://econpapers.repec.org/RePEc:oup:jconrs:v:32:y:2006:i:4:p:504-512

13. Ran Kivetz, Oleg Urminsky and Yuhuang Zheng, The goal gradient hypothesis resurrected: purchase acceleration, illusionary goal progress, and customer retention, Journal of Marketing Research 43, 39 to 58, 2006, https://business.columbia.edu/sites/default/files-efs/pubfiles/1200/goalgradient.pdf

14. Government of Ireland, S.I. No. 336 of 2011, European Communities (Electronic Communications Networks and Services) (Privacy and Electronic Communications) Regulations 2011, regulation 13, paragraphs (1), (2) and (11), Irish Statute Book, 2011, https://www.irishstatutebook.ie/eli/2011/si/336/made/en/print

15. Data Protection Commission, Direct marketing: what you need to know about direct marketing, blog, 18 September 2019, dataprotection.ie, https://www.dataprotection.ie/en/dpc-guidance/blogs/direct-marketing-what-you-need-know-about-direct-marketing

16. Information Commissioner's Office (United Kingdom), Direct marketing guidance: Identify direct marketing (What are service messages?), ICO, https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/direct-marketing-guidance/identify-direct-marketing/

17. Data Protection Commission, FAQ on Consent for Electronic Direct Marketing, April 2020, dataprotection.ie, https://www.dataprotection.ie/sites/default/files/uploads/2026-05/FAQ_Consent_Electronic_Direct_Marketing_April2020.pdf