What we found
Recurring behavior creates economic value when it does one of three things.
- It improves or predicts the transaction you get paid on.
- It makes the next transaction materially easier than starting over.
- It protects or supports a separate monetized relationship.
What does not predict value
- Registrations
- Direct traffic
- Email permission
- App installs
- Engagement by itself
Every failed case we examined cleared the identity-capture stage comfortably, often at enormous scale. None of them converted it into durable economics. The hard part is not getting people to come back. It is what their returning changes.
The evidence in one view
Eight cases in one view
Each built something to bring customers back between transactions. What separated them was not the quality of the product.
| Company | What brought people back | What happened economically | Lesson |
|---|---|---|---|
| Credit Karma | Credit score and file monitoring | $2,263M segment revenue in Intuit FY2025, up 32%, with $835M operating income at a 37% margin1 | The return refreshes the data that prices the transaction |
| Mint | Budgeting and account aggregation | Shut down 23 March 20241 | Useful recurrence can still sit outside the money |
| Zillow | Saved searches and property alerts | Roughly 80% of traffic direct or organic; $2.6B FY2025 revenue2 | Monitor an object tied to future intent |
| TurboTax | Stored prior-year return | 79% online customer retention, FY20191 | Frequency is not always necessary |
| LendingTree | Registration plus credit monitoring | Sign-ups +17% to 25.8M while revenue attributed to members fell 47%, Q1 20233 | Identity is not relationship |
| Groupon | Daily deal email at scale | 48.1M active customers in 2015; revenue roughly 80% below its 2014 peak by 20224 | Permission is not demand |
| BuzzFeed | Direct and in-app visits | Direct share reached 63% while FY2025 revenue fell to $185.3M from $383.8M in 20215 | Owned traffic is not inherently valuable |
| New York Times | Games, Cooking, Audio | Bundle subscribers from 41% to 51% of base; digital ARPU $9.79, up 3.6%6 | Habit works when it protects recurring revenue |
Credit Karma
- Return
- Credit score and file monitoring
- Economics
- $2,263M segment revenue in Intuit FY2025, up 32%, with $835M operating income at a 37% margin
- Lesson
- The return refreshes the data that prices the transaction
Mint
- Return
- Budgeting and account aggregation
- Economics
- Shut down 23 March 2024
- Lesson
- Useful recurrence can still sit outside the money
Zillow
- Return
- Saved searches and property alerts
- Economics
- Roughly 80% of traffic direct or organic; $2.6B FY2025 revenue
- Lesson
- Monitor an object tied to future intent
TurboTax
- Return
- Stored prior-year return
- Economics
- 79% online customer retention, FY2019
- Lesson
- Frequency is not always necessary
LendingTree
- Return
- Registration plus credit monitoring
- Economics
- Sign-ups +17% to 25.8M while revenue attributed to members fell 47%, Q1 2023
- Lesson
- Identity is not relationship
Groupon
- Return
- Daily deal email at scale
- Economics
- 48.1M active customers in 2015; revenue roughly 80% below its 2014 peak by 2022
- Lesson
- Permission is not demand
BuzzFeed
- Return
- Direct and in-app visits
- Economics
- Direct share reached 63% while FY2025 revenue fell to $185.3M from $383.8M in 2021
- Lesson
- Owned traffic is not inherently valuable
New York Times
- Return
- Games, Cooking, Audio
- Economics
- Bundle subscribers from 41% to 51% of base; digital ARPU $9.79, up 3.6%
- Lesson
- Habit works when it protects recurring revenue
The natural experiment
Two free products, one owner, opposite outcomes
Intuit acquired Mint in 2009 and Credit Karma in 2020. Both were free. Both made money matching users to credit cards, loans and auto insurance. Both had real, returning users.
Credit Karma
- Reason to return
- Check your credit state.
- Connection to money
- The monitored state is the same object a lender uses to decide approval and to price the introduction. Every check refreshes it.
- Outcome
- Segment revenue of $2,263M in FY2025, up 32%. Operating income $835M, up 102%, at a 37% margin.1
Mint
- Reason to return
- Check your budget and tracked accounts.
- Connection to money
- Genuinely useful, arguably more useful, and it tells a lead buyer almost nothing about whether this person will apply for anything this week.
- Outcome
- Closed 23 March 2024. Migration carried the login, connected accounts and net worth history, but not budgets, categories, goals or bill reminders.1
People coming back was not the difference. What they came back for was.
Note that Intuit had already run a version of this test. Its own free product, Turbo, reached 22 million registered users by fiscal 2020 and was wound down after the Credit Karma acquisition. Two free consumer finance dashboards with real registered bases, both retired, in favor of the one built on credit data.
The revenue and shutdown figures are drawn from Intuit filings and earnings releases. The explanation for the divergence is our reading of what the two products monitored and monetized. Intuit has not publicly stated that Mint was closed because its recurring behavior was less transaction-adjacent than Credit Karma's.
Three mechanisms
How recurrence actually creates value
Where recurrence improved the economics, the return did one of three specific jobs. Where it did none of them, the result was the same regardless of how good the engagement looked.
Mechanism A
The return predicts or improves the transaction
The customer names an object. The company watches it. When its state changes, the customer has a reason to look, and that same change means they are nearer to transacting than they were yesterday.
Zillow is the clearest case. People move house every several years, which on conventional retention logic should preclude any direct relationship. Zillow reports roughly 80% of traffic arriving direct or organic and 221 to 250 million average monthly unique users through 2025.2 A price change on a saved property is both the reason to return and the signal that an agent connection is now worth paying for.
Mechanism B
Stored state makes the next transaction easier
No habit is required, and this is the most underrated finding in the study. TurboTax is used once a year. Nobody opens tax software in September.
Intuit disclosed TurboTax Online customer retention of 79% at its fiscal 2019 investor day, improved two points on the prior year after a similar improvement the year before.1 What brings people back is last year's return sitting in the account. The mechanism sends no notifications, so it creates no fatigue and no false positives, and its value compounds every year the customer stays.
Mechanism C
The return protects another monetized relationship
The New York Times moved bundle and multiproduct subscribers from 41% of its digital base at the end of 2023 to 51% by the third quarter of 2025, reaching 6.27 million of 11.76 million. Digital-only ARPU rose 3.6% to $9.79.6
Wordle does not monetize. Nothing commercial happens when somebody solves the puzzle. That is fine, because the transaction is a recurring subscription and the metric that matters is churn rather than conversion. Attach the identical product to an affiliate or lead-generation business and it becomes a cost center with an engagement chart, because there is no subscription for it to protect. This mechanism does not transfer automatically.
Failure patterns
Most recurrence strategies fail after identity capture, not before it
Getting the customer to sign up is rarely the hard part. All four of these companies did it at scale.
Groupon
Permission without a reason
One of the largest permissioned consumer lists ever assembled, 48.1 million active customers in 2015. By 2013 less than 45% of transactions came from email, and the company was investing in local landing pages to compete for Google traffic instead.4
LendingTree
Identity without economics
In one Q1 2023 filing: cumulative sign-ups of 25.8 million, up 17% year over year, with a million net new users added that quarter, and $19.6 million of quarterly revenue attributed to registered members, down 47%.3
Mint
Utility without transaction adjacency
Real weekly usage, years of stored financial history, connected bank accounts, and a monetization model identical to the sibling product that scaled to $2.3 billion. Closed in March 2024.1
BuzzFeed
Direct traffic without yield
Direct, internal and app traffic reached 63% of BuzzFeed.com by Q3 2025, up from 61%. FY2025 revenue was $185.3 million against $383.8 million in 2021, with a going concern warning attached.5
BuzzFeed is the most instructive of the four, because every metric the strategy would be judged on moved in the right direction. The audience came back and monetized at programmatic rates. Direct traffic is not intrinsically valuable. It is valuable when the returning visitor is close to a transaction that pays well.
The monitoring test
The strongest loops watch a specific object
A second investigation looked at whether this structure is generally available. It is not, and three cases explain the constraints.
Mortgage rates
Good mechanism, available whitespace
The only case we found where the link between the state change and the transaction is publicly established rather than assumed. The Mortgage Bankers Association publishes a weekly refinance application index against weekly rate data, with decades of history.
In the week ending 9 January 2026, the 30-year fixed rate fell to 6.18% and the Refinance Index rose 40% week over week and 128% year over year. In the week ending 22 May, after rates climbed 30 basis points to 6.65%, the index fell 18%.7 The addressable population is also growing mechanically: FHFA data shows 22.1% of outstanding loans above 6% in Q1 2026, up from a trough of 7.3% in mid-2022.8
- The market state is published and free to monitor
- The customer's own state is three numbers they already know
- The alert is arithmetic the customer can audit before acting
Insurance renewal
Good-looking mechanism, bad opportunity
We initially identified this as the strongest available loop. Deeper research reversed that conclusion, and nothing was wrong with the mechanism.
- Shopping already happens. LexisNexis found 47.1% of auto policies in force had been shopped at least once in the prior twelve months as of Q4 20259
- The carrier owns the trigger. Renewal notices arrive 45 to 60 days ahead and contain the actual new premium, which a third party cannot know
- Six-month terms are the default at the largest carriers, so the annual calendar the idea rested on does not exist
- A broker already owns the object and the transaction economics, re-shopping automatically every six months, free, in all fifty states
Medicare
Perfect trigger, weak behavior change
The strongest challenge in the record to the idea that timing creates action. Plans are federally required to send every beneficiary an Annual Notice of Change by 30 September stating exactly what is changing. The enrollment window is fixed, 15 October to 7 December. Roughly 68 million people are enrolled.
KFF's analysis of federal data finds 69% of beneficiaries did not compare their coverage against alternatives, and that voluntary switching among Medicare Advantage prescription drug enrollees ranged between 6% and 12% across thirteen consecutive enrollment periods, for coverage years 2008 through 2020.10 Over the same broad period, a Senate Finance Committee investigation found agent and broker commission spending among the insurers examined rising from $2.4 billion to $6.9 billion between 2018 and 2023.11
A clean trigger does not create demand.
Thirteen years, a legally mandated notice on a fixed date, and nearly seven billion dollars a year of commission-funded pressure aimed directly at the moment of decision. The switching rate remained persistently low. Set that against auto insurance, which has no coordinated trigger at all and where nearly half the market shops annually anyway. The two categories test the same proposition from opposite directions and both refuse it.
Occupancy
A good loop can still be a bad market
Before building around an object, establish who already monitors it. Credit score monitoring is the strongest mechanism in this entire study and a poor opportunity for a new entrant.
- Credit score and fileCredit Karma, Experian's own app, Credit Sesame, most bank apps. Free, at scale, cross-subsidized by lending revenue.
- Property searchZillow, with the saved-search relationship and the listing data.
- Fare trackingGoogle and Hopper.
- Insurance re-shoppingBrokers who keep the commission rather than selling the lead.
The questions to ask are specific. Who monitors it today. Is their product free. How large is their distribution. Do they hold data you cannot obtain. Do they own the underlying transaction, which would let them spend more to acquire the same customer than you ever can.
Trust
A fourth outcome that looks like success
A loop can be genuinely useful, genuinely productive, and still destroy the relationship it depends on. The archetype is the same product that anchors the positive case.
According to the FTC's complaint, from February 2018 to April 2021 Credit Karma told many consumers they were pre-approved for credit products when they were not, leading them to apply, incur hard inquiries and risk damaging their scores. The FTC alleged the company knew this from its own A/B testing, which showed that describing an offer as pre-approved outperformed language describing strong approval odds. The consent order was finalized in January 2023 with a $3 million payment.12
The mechanism transfers with uncomfortable precision. In any monitoring product, the strength of the claim in the alert sits closest to the click, which makes it the first thing an optimization process finds. What makes this case severe rather than ordinary is where the harm landed. A product whose purpose is protecting your credit score damaged it.
The same pattern appears at category scale in Medicare, where CMS conducted secret shopping and found that in more than 80% of calls reviewed, agents failed to provide required disclosures.11
A credible monitor must sometimes be able to say: do nothing.
If every alert economically needs to become an offer, the product is not monitoring. It is an acquisition channel with a notification schedule, and those decay. This matters strategically as well as ethically: in mortgage rate alerts the incumbents are lenders, and a lender's alert cannot credibly tell you to stay where you are. That inability is a competitive opening for anyone whose economics do not require every alert to convert.
The executive framework
Before you build the app, dashboard, alert, account or loyalty program
Most proposals fail on one of the first three questions, and the failure is visible before any money is spent.
What exactly are we monitoring?
Name the object in one noun phrase. Not the customer. Not their financial life. A mortgage rate, a saved property, a renewal date, a tariff, a points balance, a prior-year filing. If nobody can name it, that is not a detail to work out later. That is the finding.
What changes about it?
Identify the state change and ask whether the customer benefits from knowing, independent of whether they transact. If the honest answer is that they should return because you have something to sell, there is no customer-side loop.
Does that change predict or improve a transaction?
The economic test, and where Mint failed. A budget alert makes somebody more engaged. A credit score change makes them more valuable. Force the distinction in writing before anything is built.
Can we monitor it accurately at reasonable friction?
What must the customer give you? LendingTree disclosed that users who connected bank accounts showed markedly better retention and monetization, and that driving adoption of that connection proved harder than expected. Deeper data produced better signals and fewer users.
Who already owns this object?
Is their product free? How large is their distribution? Do they hold proprietary data? Do they own the underlying transaction? A good mechanism in a fully occupied market is still a bad strategy.
Can we credibly say do nothing?
Model the month where most alerts recommend no action. Does the product still function? Does the business model survive? If not, evaluate it as an acquisition channel, at acquisition-channel economics.
Do we need recurrence at all?
Compare honestly against stored state, better conversion on traffic you already have, higher revenue per transaction, owning more of the transaction, cross-sell and contractual recurrence. TurboTax reached 79% retention on an annual product with no engagement mechanic at all.
The decision rule
Build for recurrence when you can name a customer-specific object whose state changes before a valuable transaction, monitor it accurately at low friction, and send information the customer would value even when the right answer is to do nothing.
Otherwise, do not manufacture engagement. Look instead at stored state that makes the next transaction cheaper, lower friction in the transaction you already have, higher transaction economics, cross-sell, vertical integration, or owning more of the transaction you currently refer away.
A test you can run this week
Take the proposal currently on the table
A newsletter, an app, a loyalty program, a dashboard, an alert product, a membership, an account, a personalized tool. Answer five questions in writing.
- Name the object in one noun phrase.
If it takes a sentence, or if the answer is a category of person rather than a thing, stop here.
- Say what changes about it, and how often.
If the state changes continuously you have noise rather than a trigger. If it never changes you have a database entry rather than a loop.
- Say what the customer does when they hear, and whether you get paid.
Be specific about the action. If the most likely response is one you earn nothing on, that is disqualifying and easy to miss. In credit card annual fee monitoring, the most likely response is a call to the issuer's retention department, which keeps the customer where they are.
- Name who already monitors this object and what they charge.
If the answer is a company with free distribution at scale, or one that owns the transaction rather than referring it, mechanism quality no longer matters.
- Write the version of the alert that says do nothing.
Then ask whether your business model survives sending it in most months. If it does not, you have your answer, and it is better to have it now.
If the idea clears all five, run a small experiment with a metric tied to transaction behavior rather than registrations. If it fails one, do not improve the engagement layer. Reconsider the underlying mechanism.
The goal is not to give customers more reasons to open your product. It is to find the moments when a customer returning makes the relationship worth more, and build around those.