PastBehavior · Fieldwork No. 043 · Evidence cutoff 27 Aug 2026

When the Relationship Cannot Monetize Itself, Where Does the Money Move?

Fifteen documented cases where the layer that produced trust, habit or attention was billed somewhere else entirely.

Loyalty · Software · Financial services · Retail · Media · Marketplaces · Children's content · Regulation · AI

Plenty of products create enormous engagement or trust while collecting revenue somewhere else. The usual explanation is that the relationship itself becomes the asset and someone eventually pays for access to it.

History is more selective than that. Two companies can both own a beloved user relationship, monetize it indirectly, and end up in completely different economic positions. One of them prints cash for decades. The other spends twenty years dependent on a single payer whose obligations are being litigated by the Department of Justice.

The cases below look at what separated them.

The cohort

Fifteen cases where engagement and billing sat in different places

Selection favored cases with disclosed financials, regulator records or peer-reviewed measurement, and cases where the separation between relationship and payment was forced by something identifiable: regulation, trust, platform rules, conflict of interest, low willingness to pay, or the accounting treatment of the product itself.

CaseRelationshipWhy not billed directlyWhere the money movedScarce inputCaptured byConf.
Delta / SkyMilesStatus and a currency travelers organize aroundMiles are a liability, not revenueCo-brand remuneration from American ExpressThe miles themselvesDeltaFACT
Mozilla FirefoxFree browser with a privacy-first identityNobody pays for a browser; the mission forecloses ad surveillanceSearch default placementContested. Distribution without leverageGoogle, in effectFACT
Credit KarmaFree credit score, advisory postureCharging destroys the trust premiseLender referral commissionsUnderwriting-relevant profileCredit KarmaFACT
CostcoTrust that pricing is honestDeliberate. Merchandise margin capped by policyAnnual membership feeCredible restraintCostcoFACT / INFERENCE
WikipediaNear-universal reference use, high trustNo-advertising norm is constitutiveReader donationsNone retained. Corpus given awayDownstream platformsFACT
SpotifyInterface, habit, recommendation, subscriptionNot constrained. Spotify does bill usersTwo-thirds out to rights holdersThe catalogRecord labelsFACT
Apple App StoreDevelopers own the app and the affectionPlatform rules, not choice15–30% commissionIdentity, payment, distributionAppleFACT
ZillowFree consumer search and the ZestimateConsumers will not pay to browseAgent advertising, then a failed move into the transactionConsumer intent at scaleZillow, in the ad layer onlyFACT
YouTube kids contentVery large children's audiencesCOPPA removed personalized advertisingMerchandise and licensing for the largest channels onlyScale, for those who had itNobody, at the marginFACT
UK RDRAdviser relationships with retail investorsRegulator banned product commission in 2012Explicit adviser fees and D2C platformsAdvice for higher-asset clientsAdvisers and platformsFACT
Tobacco marketingBrand loyalty under advertising bansBroadcast, outdoor and merchandise restrictionsRetail price discounting and shelf paymentsPoint of saleManufacturers and retailersFACT
Overdraft / interchangeFree checking relationshipsCFPB rule capping fees, later repealedFees persisted; interchange fell separatelyThe deposit accountBanksFACT
Character.AIAmong the highest per-user engagement in consumer softwareLow willingness to pay against high serving costAn acquirer, not the usersModel and founding teamGoogleESTIMATE
RobinhoodCommission-free trading, retail-friendly framingNot constrained. Billing was concealed rather than absentPayment for order flowRetail order flowRobinhoodFACT
DuolingoFree, habit-forming daily productNot constrainedConsumer subscriptionThe habit itselfDuolingoFACT

Confidence refers to the monetization figures cited in the sections below, not to every cell. Registered users, downloads and time spent are treated as inputs throughout, never as evidence of monetization.

Finding 01

The money follows the scarce input

Four cases sit at the corners of this.

Delta owns the currency

A frequent flyer earning and redeeming miles inside the program generates no cash and creates a future obligation. The program becomes a business only when a bank is allowed to buy access to it. Delta recognized $8.2 billion of American Express remuneration in FY2025, roughly 14% of adjusted operating revenue, against $7.4 billion the year before, with a stated target of $10 billion. Fact

American Express pays because it cannot issue SkyMiles. The status tiers, the redemption inventory and the earn structure are Delta's, and no amount of cardholder relationship substitutes for them. American Airlines reported $6.2 billion from co-brand and partner agreements in 2025, roughly four times its adjusted operating income. Reported

Credit Karma owned the profile

Consumers will not pay for their own credit score at scale, and charging would have contradicted the product's advisory posture. Lenders paid instead, because a pre-qualified applicant is worth more than an anonymous one and the targeting could not be recreated from the outside. Credit Karma reported nearly $1 billion in unaudited 2019 revenue with more than 100 million members, and Intuit acquired it for about $7.1 billion. Fact

Spotify owns the relationship and not the leverage

Spotify originates the subscription, holds the habit, runs the recommendation system and controls the interface. Roughly 68% of 2025 revenue went to rights holders and related costs, with the company reporting more than $11 billion paid to the music industry that year. Premium gross margin ran near 33%. Fact

Listeners will not accept substitute recordings, so the catalog holder sets the terms. Owning every part of the experience except the content produced a low-margin toll operation on someone else's asset.

Wikipedia gives the input away

Trust and utility here are as high as anywhere in the cohort. The Wikimedia Foundation reported roughly $185 million in FY2024 support and revenue, $168 million of it donations. Reported Search engines and, later, model developers built on the corpus without paying proportionally, and the Foundation's own FY2026-2027 draft plan describes declining pageviews and reduced search referral traffic as a structural shift rather than a temporary one. Fact

Owning the relationship and owning the leverage are different things.
Relationship ownerScarce inputWho paysWho captures
Delta
SkyMiles currency and redemption inventory
American Express
Delta · $8.2bn FY2025
Credit Karma
Underwriting-relevant consumer profile
Lenders
Credit Karma · ~$1bn 2019 revenue / $7.1bn acquisition
Spotify
The recorded music catalog
Subscribers
Rights holders · ~68% of revenue
Apple
Identity, payment and distribution on iOS
Developers
Apple · 15–30% commission
Wikipedia
Given away by charter
Readers, voluntarily
Downstream platforms
Mozilla
Distribution, without exclusivity
One search provider
Google · Mozilla dependent

Reading the map: the capture column tracks the scarce input, not the relationship. The two weak rows are cases where the relationship owner cannot restrict access to the scarce input or lacks leverage over the counterparty.

Finding 02

Trust is monetizable when it produces access to someone billable

The intuitive version of this says trusted intermediaries capture more when charging directly would damage the trust. The cohort does not support that as stated.

Credit Karma's trust produced qualified financial demand and lenders paid to reach it. Costco's members pay explicitly for the restraint: merchandise gross margin ran around 11.0% in Q3 FY2026 while $1.373 billion of membership fee income drove $2.815 billion of operating income. Reported The member is buying the discipline, and the discipline is the product.

Wikipedia has trust at least as strong as either and no commercial counterparty who benefits in a way that generates an invoice. It collects donations averaging roughly eleven dollars.

Robinhood is the case that closes the argument. The SEC found misleading statements between 2015 and late 2018 about payment for order flow, and $34.1 million in inferior execution costs to customers, settled for $65 million. Fact By 2021, transaction-based revenues were over 77% of net revenue at $1.4 billion. Fact A documented breach of the trust premise, a nine-figure penalty, and no lasting commercial damage.

Trust becomes monetizable when it produces access to a scarce, qualified or expensive-to-reach counterparty.

That formulation holds across the fifteen cases and should not be pushed further. It says nothing about whether trust is worth having for other reasons.

Finding 03

Regulation relocates revenue when another payer exists

This is the strongest pattern in the cohort and it has a sharp boundary.

Cigarette advertising was banned from broadcast in 1971 and restricted on outdoor, transit and branded merchandise by the 1998 Master Settlement Agreement. Spending went up. The five largest manufacturers spent $6.73 billion in 1998, then $8.24 billion in 1999, a 22.3% increase and the highest total ever reported to the FTC at that point. Across 1995 to 2001 the industry's advertising and promotional spending rose 96%. Fact

It moved to the one channel left open. By 2002, price discounts paid to retailers and wholesalers were $7.87 billion, or 63.2% of total spend. By 2013, 93% of expenditure went to retailers and wholesalers. Fact Restricting persuasion did not reduce commercial pressure. It converted it into price at the point of sale.

The UK's Retail Distribution Review produced a quieter version. The regulator's own post-implementation review found product bias reduced, sales of formerly high-commission products down, product and platform costs broadly falling, adviser charges not falling, and consumers increasingly buying on a non-advised basis through direct-to-consumer platforms. Fact Payment became visible, moved up-market, and pushed the mass segment into self-service.

Overdraft showed how reversible the restriction itself can be. The CFPB finalized a rule in December 2024 capping most fees near $5. Congress repealed it under the Congressional Review Act in 2025, and in 2025 banks and credit unions still collected over $12 billion, with several large banks up 22% to 40% versus 2023. Fact

Then the exception.

Adjacent payer exists

Revenue relocates and can grow

Tobacco, 1998 to 2013. Advertising restricted, retail left open. Spend rose 96% across 1995 to 2001 and shifted to price discounting, reaching 93% of expenditure paid to retailers and wholesalers.

  • Retailers were willing and able to be paid
  • The mechanism changed, the pressure did not
  • Product volumes fell for unrelated reasons
No adjacent payer

The activity contracts

YouTube made-for-kids, 2020 onward. After the $170 million FTC settlement, personalization was switched off for child-directed content. A difference-in-differences study of 5,066 top American channels found supply down 18%, views down 20%, original content down 9% and manual captioning down 28%.

  • Effects were largest for mid-sized channels
  • The largest channels found merchandise and licensing
  • The smallest were hobbyists with nothing to lose

Both outcomes followed a restriction on how the audience could be monetized. The difference was whether an economically viable payer sat next door. Fact on both measured outcomes.

Regulation redirects money only when there is somewhere economically viable for it to go.

Scale determined who could rebuild an economic layer. That distribution matters later.

Finding 04

Deliberate separation is usually caused by a constraint

Costco is the only case in the cohort where separating the profit line from the transaction reads as a genuine strategy, and even there the member is persuaded and billed by the same company on the same day. The separation is between products.

Delta's arrangement was not designed as a strategy. Miles are an accounting liability, so the program had to find a counterparty. Robinhood put the monetization inside the relationship, concealed it, was penalized, and grew. Duolingo charges its own users and does very well doing it.

Zillow tested the opposite move and lost. Premier Agent revenue reached $1.4 billion in 2021. In the same year the company went after the transaction it had been renting out, and in Q3 2021 alone the Homes segment lost $422 million before taxes on $1.2 billion of revenue. Fact The unit closed with a reported $881 million loss for the year and about a quarter of staff cut, while the advertising layer it was meant to supersede continued intact.

Where monetization appears deliberately separated from the relationship, look for the constraint that created the separation.
Finding 05

Gates capture when they are unavoidable

Apple holds identity, payment and distribution on iOS, and has defended a 15 to 30% commission through five years of litigation. A 2021 anti-steering injunction produced a compliance plan with a 27% commission on linked-out purchases plus warning screens. In April 2025 the court found willful violation. In December 2025 the Ninth Circuit restored Apple's ability to charge some commission on external purchases, and the rate-setting fight returned to the district court in August 2026. Fact The rate has moved a little. The position has not.

Mozilla also distributes, to a large and loyal user base, and holds nothing comparable. Roughly 86% of Mozilla Corporation's 2024 revenue came from Google search royalties on total revenue near $680 million. Fact The company formally warned the DOJ that a ban on Google's search distribution payments would harm Mozilla, which is the position of an organization whose survival depends on the remedy phase of an antitrust case it is not party to. Fact

Distribution captures value only when the party being distributed cannot cheaply reach the user another way.
Relationship
Wikipedia Character.AI Mozilla
Content / IP
Record labels Kids channels with merch
Qualified data
Credit Karma Delta SkyMiles
Payment rail
Robinhood Banks, on overdraft
Platform gate
Apple Costco Nintendo, Amazon

Where the durable economics landed, by layer. Cases in the leftmost column own the relationship and captured the least. Bolded entries captured the largest share of the value their relationship created.

Finding 06

Engagement without economic architecture is fragile

Character.AI reached among the highest per-user engagement in consumer software, with third-party measurement putting peak monthly actives around 20 to 28 million at roughly 75 minutes a day. Estimate Against that, 2024 revenue was approximately $32.2 million from a single consumer subscription. Estimate In August 2024 Google executed a reported $2.7 billion non-exclusive technology licensing deal that returned the founders to Google while the company continued independently. Unresolved The split between license, talent and competitive value has never been filed publicly, so the multiple is directional rather than exact.

Wikipedia is the same shape at a longer duration. Enormous utility, decades of it, and a funding model that depends on human pageviews arriving at a page with a banner on it.

Both were strong for years before the weakness became visible, and in both cases it became visible when an external party changed the terms of distribution rather than when engagement declined.

Engagement can stay strong for years while economic leverage stays weak. The gap only shows up when somebody else changes the terms.
Boundaries

The counterexamples that survived

Duolingo, on charging users directly

As of Q2 2025, 10.9 million paid subscribers against 47.7 million daily active users, subscriptions at 83% of revenue and gross margin near 72%. Fact Roughly 23% of daily actives pay. Where willingness to pay exists and nothing forbids the charge, direct beats indirect by a wide margin. Nothing in this cohort argues for separating monetization when you do not have to.

Robinhood, on the durability of trust

Direct monetization survived being concealed, penalized and publicly documented. Treat trust as a commercial input with conditions rather than as a constraint that automatically binds.

Overdraft, on the durability of rules

A finalized federal rule was repealed within a year and the fee revenue rose. Regulation of a monetization mechanism is politically reversible in a way that market structure is not, and forecasts that assume a rule will hold should carry that as a stated assumption.

The answer

What the fifteen cases establish

The money follows the scarce input, not the relationship.

Regulation redirects monetization only when an adjacent payer or economic rail exists. Without one, the underlying activity shrinks.

A relationship produces economic leverage when it gives its owner control of something another party cannot cheaply reproduce: a currency, a qualified profile, a content catalog, an identity layer, a payment rail, or an exclusive distribution point. Delta, Credit Karma and Costco held one. Spotify, Wikipedia and Mozilla did not.

Where the relationship owner holds no such input, another layer captures most of the economics while the relationship owner keeps the attention, the trust and the operating cost. That arrangement can run for a decade or more before anything forces the issue.

The practical test. Name the counterparty who benefits from your user relationship. Then ask what they would have to build to reach that user without you, and how long it would take. If the honest answer is a marketing budget and eighteen months, you are in Mozilla's position regardless of how much your users love the product.

Application

Children's conversational AI as a live test

The current market is not evidence for the historical finding. No conversational children's character has disclosed usage, retention, conversion or revenue, so what follows is a forecast derived from the cohort rather than a reading of results.

What is established is the structure. A Hasbro-developed Peppa Pig voice experience has been live on Amazon Fire Kids tablets through Amazon Kids+ since December 2025. WildBrain acquired Personality AI, the company behind it, in August 2026 for about $11 million in cash plus stock, with up to $56 million more tied to revenue targets in 2027 through 2029. Hasbro launched Sixth Wall in June 2026 with a licensing category it calls behavioral licensing, covering how characters think, speak and interact in dynamic experiences, and stated that the studio is not currently developing AI products targeted at young children. Fact on all four. Commercial results: Unresolved.

The relevant constraint is that Washington's HB 2225, effective January 2027, bars AI companions directed at minors from soliciting gift-giving, in-app purchases or other expenditures to maintain the relationship. Fact That closes the most direct route from attachment to revenue before anyone has shown it works.

The question the cohort poses

If direct monetization inside a child's AI relationship is restricted, which scarce input commands the money? Four candidates, and the cohort ranks them unevenly.

The gate

Amazon holds the account, the hardware, the subscription, the age settings, the parental controls and the distribution. Neither the IP owner nor the model supplier reaches the child without it. On the historical record this is the strongest position, because the gate captured in almost every case where anyone captured durably. It is also the position Apple has defended in court for five years rather than concede.

Character IP

Behavioral licensing is an attempt to make character behavior a scarce, billable input while the model layer commoditizes. Whether children's attachment attaches to a specific licensed character or to the responsive object it inhabits is not settled by this research. Unresolved The Spotify case shows what happens when the catalog is genuinely irreplaceable. The Mozilla case shows what happens when it is not.

Model and interaction quality

Character.AI is the closest analogue and it did not capture. Extraordinary engagement, roughly $32 million of revenue, and value realized through an external licensing and talent transaction while the company continued to operate independently.

Relationship memory

Persistent personal history is the input with no historical precedent, because no prior product let a child carry accumulated relationship between competing services. It could create real switching costs. It has never been tested. Unresolved

Relocation or contraction

The tobacco and YouTube contrast is the useful frame here. Tobacco's spend relocated because retailers were willing and able to be paid. Made-for-kids content contracted because mid-sized creators had nowhere to go.

Children's conversational AI faces the same fork. Subscriptions, hardware bundling, IP licensing and platform fees are all candidate adjacent payers, and none has yet shown it can absorb the value that direct monetization was expected to carry. If they can, the category reprices. If they cannot, the historical precedent says fewer of these products get made rather than the same number getting made differently.

Predictions

Four claims that can be checked

Prediction 01

Direct monetization inside children's conversational relationships will be constrained by regulation before it is proven at scale.

Washington's HB 2225 takes effect in January 2027 and bars soliciting purchases to maintain an AI companion relationship with a minor. No company has yet published evidence that this monetization route works with children.

Resolves by Jan 2028. Supported if HB 2225 takes effect without amendment and no US company has published paid conversion, attach rate or renewal figures for a children's conversational character by then. Falsified if a major IP owner or platform discloses working in-relationship monetization economics before the restriction binds.

Prediction 02

The platform controlling identity, age verification, permissions and payment will capture a disproportionate share of the economics.

Across the cohort, the unavoidable gate captured in almost every case where durable capture occurred. Amazon Kids+ currently occupies that position for the only shipping licensed conversational character aimed at young children.

Resolves by 2029. Supported if disclosed licensing or revenue-share terms show the platform retaining the majority of gross consumer spend, or if the take rate becomes the subject of litigation or regulatory complaint. Falsified if IP owners or model providers negotiate majority economics, or if children's conversational characters reach scale outside platform-controlled distribution.

Prediction 03

If no adjacent payer develops, the supply of children's conversational products contracts rather than repricing.

This is the YouTube made-for-kids precedent applied directly. Supply fell 18% and views fell 20% when personalization was removed and mid-sized creators had no substitute payer. The largest players found merchandise and licensing. Everyone else stopped making the content.

Resolves by 2029. Supported if the count of licensed conversational children's products in market declines year over year while the largest IP owners continue. Falsified if subscription, hardware or licensing revenue visibly absorbs the restricted channel and product counts hold or grow.

Prediction 04

The structure of character licenses will reveal whether licensors themselves believe character IP is scarce.

A per-interaction or revenue-share royalty prices the character as irreplaceable. A flat or capped fee prices it as substitutable. Sixth Wall is currently running time-bound pilots and has disclosed no terms. Watch the contract shape rather than the announcement.

Resolves by 2028. Supported if disclosed AI character deals are predominantly flat or capped fees. Falsified if per-use royalties or revenue shares become standard. Note: this is a signal to watch rather than a settled question. Whether children substitute away from a specific character was not tested in this research.

Dashboard

Observable signals

The paying denominator
Whoever publishes paid conversion, attach rate, renewal or partner revenue is usually the party capturing. Delta reports Amex remuneration to the dollar. Character.AI's revenue had to be estimated by third parties.
Where the take rate is litigated
Capture shows up in court before it shows up in analysis. Five years of Epic against Apple identified the App Store as the capture layer more clearly than any market study.
Fee versus revenue share
Contract structure is the market's own verdict on whether an input is scarce. Flat fees price substitutability. Per-use royalties price irreplaceability.
Revenue concentration
One counterparty above roughly half of revenue is Mozilla's position. Track the share rather than the amount, and track what would happen to the payer under adverse regulation.
Where restricted spend reappears
After any marketing or monetization restriction, look for the equivalent of tobacco's shift to price discounting. In children's AI the likely destinations are subscription bundling and hardware, both of which sit outside a toy definition.
Product supply, not price
The YouTube study measured regulation's bite through an 18% fall in output. If products stop being made rather than being repriced, no adjacent payer was found.
Method

Evidence standards and known gaps

Claims carry a label where the distinction changes how much confidence the reader should place in them. Fact means a filing, regulator order, audited statement or peer-reviewed result. Reported means consistent secondary reporting without access to the primary. Estimate means an undisclosed method or a third-party measurement. Unresolved means public evidence cannot answer it.

Registered users, downloads, time spent and brand recognition appear here as inputs and never as proof of monetization.

Two weaknesses worth naming. Costco's membership fee income and operating income are reported quarterly figures, cited here through secondary coverage rather than the filing itself, and the merchandise gross margin is arithmetic performed on them rather than a disclosed segment line. The Character.AI deal terms have never been filed, so the engagement-to-value multiple is directional. Neither changes the direction of the finding, and both would change the precision of any number built on them.

A conversational children's character that monetizes at scale with no gate, no subscription and no adjacent payer would break the central pattern. Nothing in this cohort shows one. That is a statement about fifteen selected cases rather than about the wider universe, and Duolingo sits inside the same cohort as a reminder that direct monetization works well when nothing prevents the charge.

Download PDF