Cardlytics: The Ad Platform with Purchasing Power
What this covers
This is a Business Breakdowns episode in which Clifford Sosin walks through how Cardlytics works and why he thinks its destination is clearer than its timing. It begins with the plumbing: advertisers design campaigns in an outside system, while an inside-the-firewall system identifies customers and returns only aggregate, non-PII results, so data stays with the bank. Targeting can be granular, and the Starbucks example segments by competitor shopping, wallet-share deciles, loyalty membership, and nearby spending. The episode also explains the accounting. Reported revenue is the 70 cents left after the consumer's 30-cent share, and the bank's cut is treated as cost of goods sold.
The middle of the conversation covers measurement and its contested claims. Clifford argues that multi-factor attribution from cookies and econometrics only shows association, so the returns Google and Facebook report are worse in reality than they look. Cardlytics' returns, by contrast, are measured against a holdout group. He also recounts losing advertisers when an internal champion left.
The economics come next. A bank's 35-cent share is put at 10 to 20 percent of its total benefit, with loyalty worth five to ten times the share. Clifford sees the take rate rising from about 7 percent toward 15, offer ordering and pricing as far from optimized, and U.S. Bancorp seeing 50 percent activation after a redesign. Advertisers are said to use about 1 percent of what the channel could absorb, implying a market above $100 billion. He compares the company to Visa and argues a bank building this itself would lose.
Cardlytics' edge comes from pairing bank transaction data, kept behind the bank firewall, with randomized-controlled-trial measurement, creating a channel that is underpriced and under-optimized. Its growth depends on advertisers trusting that measurement and on banks staying conservative enough to keep the channel closed.
- The bank-firewall architecture lets banks share spending-based targeting without releasing customer data, which is what won the bank relationships.
- RCT-based lift measurement is far more credible than attribution models, but advertisers do not yet trust or understand the difference, so buy-in is the main barrier.
- The revenue share captures only about 7% of the value created, and the offers are currently displayed and priced far from optimally, so the economics have room to improve.
Advertisers get measurable, high incremental returns that justify continued spending.
- When a channel has measurable returns, an advertiser should keep spending down to the last subset of consumers where the return still exceeds its margin, because stopping early leaves free money on the table, especially if other channels' returns are uncertain.
“not spending down to the last penny literally leaves free money on the table”
- A typical advertiser on the Cardlytics channel gets about $5 of incremental spend for every dollar spent, though this varies by category and campaign type.
“a typical advertiser will get about $5 of incremental spend per dollar spent in the channel”
- Google and Facebook's stated returns look better than Cardlytics' but their actual returns are far worse, whereas Cardlytics' measured returns are what it says, because platform measurement quality is far inferior to randomized measurement.
“the stated returns that you get from Google and Facebook are usually better than Cardlytics but the actual returns are usually far worse. Whereas with Cardlytics, the returns are actually what they say they are.”
- Early Google search advertising was extraordinarily cheap, with terms costing a penny or two, which produced returns that Cardlytics cannot match because its 30-cent consumer payment and bank-controlled advertiser base cap the return.
“even though the returns are great, they're not necessarily out of this world great the way the early returns might've been for Facebook or Google”
- Because Cardlytics sees spending across both the exposed group and a randomly assigned holdout group, it can run a randomized controlled trial and measure incremental spend from an offer to the penny, which is unusual in advertising.
“you can then measure the difference that occurs in spending between the two in a randomized controlled trial format, and that can give you, to the penny, measurement of the efficacy of the offer in creating incremental spend”
- Advertisers like Starbucks can segment bank customers by competitor shopping, wallet share deciles, loyalty membership, nearby spending, and offer responsiveness, and in theory can keep adding criteria with no practical limit.
“Starbucks can quite literally identify people who shop in the cafe category, say a Dunkin, but not at Starbucks, or could group them into deciles based on what percentage wallet share Starbucks has”
- Advertisers who measure channels by precise test-versus-control trials will tend to treat the Cardlytics channel as a high-return, precisely measurable one, which changes how they allocate spend across channels.
“When you have nearly perfect measurement, you use a channel differently than when you don't.”
- Multi-factor attribution models used by digital platforms are association-based, so they cannot establish causation and are at best indicative, which is why advertisers apply haircuts to their numbers.
“like any association type of analysis it's flawed, right? I mean, association is not causation”
- The Bridge acquisition gives Cardlytics access to retailer product-level data, enabling category- and SKU-level offers such as 15 percent off home decor or a dollar off a specific brand, which lets retailers target by margin and inventory goals.
“instead of saying, you're going to get 5% off at target, they can say, you're going to get 15% off of home decor at Target”
- Improving the UI to hold real copy would let advertisers like Chipotle highlight products and let small brands like Joe's Pizza build recognition, and Clifford argues this opens the roughly 70 percent of advertisers poorly served by the current tile.
“So having a new improved UI, you can have some ad copy that delivers some branding.”
- Cardlytics is a broken slot machine in the sense that every dollar put in returns more money, so not spending down to the last segment leaves free money on the table for an advertiser with measurable returns.
“not spending down to the last penny literally leaves free money on the table. Cardlytics, in some sense, the joke is that it's a broken slot machine.”
Data access, permissioned bank channels, and tech replication costs create a defensible moat.
- Network effects are a demand-side effect where one user's use improves value for others, and in Cardlytics, adding merchants can help consumers and other offers, but adding competing merchants in a small category can produce less-than-additive benefits, so the effect is complicated.
“In the first example, one plus one is more than two. And the second example of one plus one is less than two.”
- Adding a major brand such as Home Depot raises consumer engagement with the whole channel, which makes existing offers like Burger King's more effective, a clear demand-side network effect.
“as you increase the richness of the content available, consumers use Cardlytics more”
- Full transaction histories combined with a secure, bot-free, brand-safe channel to reach customers is unique data that banks are unwilling to share widely, making data a real moat for Cardlytics, though it is still very early in using that data.
“the idea of having basically all the person's transaction history, and also a way to talk to that person in a secure environment where there's no bots, and you know it's brand safe and everything else, is purely unique”
- Because the bank keeps its data behind its firewall and returns only non-PII summaries, Cardlytics is inherently safe from a PII perspective and well protected against regulatory or consumer backlash, since it simply helps banks use data they already hold.
“Cardlytics is inherently very safe from a PII perspective. A good and I think accurate framing is Cardlytics helps your bank use data of your bank already has to find offers for you.”
- Google's early model was permissionless with imperfect targeting, while Cardlytics is permissioned with perfect targeting but limited reach, so the two are roughly inverses, with the permissioned access being the constraint that is slowly changing.
“Permission versus permissionless access is a nice way to put it.”
- The growth of neobanks fragments banking, which Clifford argues helps Cardlytics because it can serve many banks, and a competitor like Square has little appeal because few merchants and users would benefit from its small offer network.
“the Neobank's fragmentation in the banking industry is good for Cardlytics”
- Customers rarely change banks because switching is painful, so a Google-branded banking skin layered on top of banks is unlikely to win, though it would be a risk to Cardlytics if it ever did evolve.
“One of the things about banking is that people really don't change banks very much.”
- Cardlytics' technology has steep learning-curve effects and an agency-facing self-service platform that is costly to replicate, requiring time, cost, and ongoing maintenance, which makes the advantage durable against a bank building it alone.
“There's a lot of learning curve effects in building that technology over time.”
- Bank customers are so reluctant to switch that people in one anecdote left dead fish in safe deposit boxes rather than close accounts, so retention inertia protects bank distribution.
“So they left dead fish in the safe deposit boxes because it was easier to do that. People just don't change banks that often.”
- A Google app layer that intermediates bank accounts is a real risk, but customers rarely switch banks, and Google's offers are not meaningfully better than banks' offers, so the threat is limited.
“One of the things about banking is that people really don't change banks very much.”
Bank control, a data-firewall architecture, and aligned incentives make the channel workable.
- Cardlytics sources offers from advertisers on behalf of banks and presents targeted offers to bank customers inside the bank's mobile app, matched to each customer's shopping preferences and the advertiser's goals.
“Those offers are powered by Cardlytics, and that fundamentally is the business.”
- Cardlytics splits its software into an outside system for designing campaigns and an inside-the-firewall system that identifies customers and returns only aggregate, non-PII summary data, so banks never release customer data, which was the key innovation that won bank partners.
“In this way, the bank literally never lets the data leave from behind its firewall.”
- Banks want control over which advertisers appear, favoring slow-moving large brands, and the tiny offer tile in bank apps lacks room for copy, so small or unrecognized brands like a local burger shop cannot use the channel effectively without a UI redesign.
“there's no room for real copy. And what that means is that if you are Joe's Burger, you may not have a logo that is widely recognized.”
- Banks want the Cardlytics offers to appear as their own, such as Bank of America or Chase offers, so Cardlytics can operate with little consumer brand awareness and banks can layer proprietary value on top.
“it's really important to the banks that this be Bank of America deals or Chase offers or whatnot”
- Around 2010 many ad-industry startups tried to get banks to send them customer data in exchange for promises of value, and banks refused, which is why the approach failed and Cardlytics needed a different architecture.
“And the banks thought about it and said, That's interesting, but no thank you.”
- Advertisers, consumers, and banks will eventually take the free money Cardlytics offers because the incentives align for all parties, so even if near-term obstacles persist, the channel will inevitably connect them.
“here's the thing, water goes downhill.”
- The Cardlytics model has four parties (banks, consumers, advertisers, and Cardlytics), and value must be understood for each before one can reason about expansion.
“there's really four players, right? There's the banks, there's the consumers, there's the advertisers, and of course, there's Cardlytics.”
- Ad agencies will only serve medium-sized advertisers through Cardlytics if they can run campaigns themselves on a self-service platform, because they need to put their own spin on each client's campaign.
“those are served through ad agencies and ad agencies will only do that if they have a self service platform where they can do it themselves”
- Adding richer content to the channel creates a true demand-side network effect, but in saturated categories like hardware, a new entrant mostly cannibalizes existing offers, so the total benefit grows sublinearly and the effect is not as algorithmic as it seems.
“the benefit of the second hardware store is smaller than the benefit of the first. And in part comes from cannibalizing the first”
Banks gain more from loyalty than revenue share, so building this in-house loses.
- The bank's largest benefit is not its 35-cent revenue share but increased customer loyalty, which reduces attrition, raises spending, and increases credit card revolving balances, with the value estimated at five to ten times the revenue share.
“which is about five to 10 times larger than the 35 cents they get is it increases customer loyalty to the bank”
- Advertisers face fixed costs to learn and manage an ad channel, so they value reach highly, and a larger channel is much easier for a sales force to sell than a smaller one, which makes the business hard for a single bank to replicate.
“advertisers really value reach. And so, a sales force will have more success selling a bigger channel than they will a smaller channel.”
- Disruption has a precise meaning from Clayton Christensen, a low-end toy product that incumbents ignore and that improves until it displaces them, but people now use it to mean anything new, which is a pet peeve.
“Anything new is disruptive.”
- Cardlytics can be analogized to Visa as an industrial utility sitting in the middle of the ecosystem, but it may be better positioned because every party benefits, unlike Visa's interchange, where merchants may feel they lose.
“We may be in some sense better, I think, than Visa because Visa, they have to charge the merchants. The merchants don't necessarily feel like they come out ahead.”
- If a large bank tried to build this capability itself, it would likely get 20 to 50 percent fewer ads in the channel and incur substantial costs, making in-house execution a losing proposition compared with outsourcing to Cardlytics.
“they'd get 20, 30, 40, 50% fewer ads in the channel. And they'd also incur a lot of costs. And combined, it would be a losing proposition.”
- Of each advertiser dollar, about 30 cents goes to the consumer as a discount, and the remaining 70 cents is split roughly evenly between Cardlytics and the bank, with each keeping about 35 cents.
“about 30 cents will go to the consumer, and the remaining 70 cents is roughly split between Cardlytics and the bank”
- A bank receiving a 35-cent revenue share likely also gets roughly $1.50 to $3.50 in other economic benefits, so the revenue share is only about 10 to 20 percent of the total value the bank receives from the program.
“it's likely that you're also getting something on the order of a $1.50 to $3.50 cents of other benefits”
- Cardlytics collects roughly 7 percent of the value created in the ecosystem, and its take rate could reasonably rise toward 15 percent over time as bank revenue-share terms improve or as it raises prices to advertisers.
“Cardlytics, They're collecting 7% of the value that's being created in this ecosystem.”
Better design, wider reach, and personalized pricing make the untapped addressable market very large.
- Cardlytics currently shows offers in a random order, with no good reason for the first four slots to be randomly assigned, and pricing is set to be advertiser-friendly rather than optimized, so there is large room to improve the channel.
“There's no reason why those four spots in the front page should be randomly assigned.”
- The best way to get advertisers to adopt the channel is for ad agencies to sell it, and younger industry entrants are more open-minded, so gradual change in buy-in is underway.
“the best way to get someone to believe something have them sell it, so get the ad agencies to start selling this stuff”
- The channel enables personalized pricing across the economy, such as a ski resort running a weather-triggered sale to nearby skiers or a movie theater filling empty seats with targeted offers, which Clifford describes as an unlimited set of possibilities.
“this is a channel that enables personalized pricing in the economy, for everybody”
- After U.S. Bancorp launched the new UI and UX, fully 50 percent of its mobile banking users activated an offer within the first three weeks, even though the offer set was still limited, suggesting that better UI can drive much higher engagement.
“in the first three weeks, since launch with the new UI, UX, fully 50% of all MAUs at Bancorp activated an offer”
- For the vast majority of advertisers, Cardlytics spending is about 1 percent of what the channel could absorb given their identified target segments, implying a large bottoms-up addressable market, possibly over $100 billion across the economy.
“for the vast majority of advertisers, they are spending on the order of 1% of what they can spend”
- Combining more monthly active users (250 to 300 million from roughly 170 million), higher engagement from UI and self-service, and more mobile adoption could expand the channel by a factor of five or more, which Clifford argues would make the revenue potential very large.
“So those things put together, I think, have the opportunity to increase by a factor of five or more.”
- Cardlytics' revenue opportunity is nested: enterprise sales are already multi-billion dollars, medium-sized businesses via ad agencies add about twice that given a self-service platform, and better UI and SKU or category offers expand the channel further.
“that is a multi-billion dollar opportunity just by itself. However, the next piece of that is medium-sized businesses.”
- Consumers see three to five offers on the first page of the banking app and must navigate to a see-all page for the rest of the 25 to 40 offers, so discovery is limited and the offers are then randomly sorted down the page.
“they'll usually be three or four offers, or five offers depending on your implementation.”
Early struggles from slow enterprise adoption were overcome through bank wins and mobile growth.
- Cardlytics' early reach through a few banks was too small to interest advertisers, which made it hard to attract content, consumers, and then banks, creating a chicken-and-egg trap that left the company struggling for years.
“and advertisers didn't care because it was just too small to be worth their time and attention”
- Cardlytics nearly ran out of money two or three times and had an IPO that nearly failed before it signed its turnaround deals.
“Finally, they nearly went broke, I think, twice or three times. They had an IPO that nearly failed.”
- Mobile banking users engage with offers far more than desktop users, producing roughly 80% of channel use while mobile banking is about 40% of banking activity, which helped the company's growth.
“circa 80% or something of their use comes from mobile banking, even though mobile banking represents circa 40% of banking”
- The summer 2018 announcement of Wells Fargo and Chase within quick succession gave Cardlytics a large jump in reach that started to change its problems, becoming a defining moment for the company.
“in the summer of 2018, they announced in quick succession the signing of Wells and Chase”
- Doubts about whether CEO Lynne had the bandwidth to build the advertising-tech side of the company were answered over the last two years, as the team grew into skill sets for bank deals and advertising, with the management team now complementary.
“Lynne, did she have that ability to sort of build that part of the organization? And over the last two years, they've made huge strides there, and I think that question has been answered.”
- Evaluating a management team is not a binary good-or-bad judgment; it is about whether the team's collective capabilities, norms, processes, and shared beliefs fit the specific problem the business faces.
“I tend to think management teams instead of a one dimensional way, like good and bad, I tend to think it's more the right fit for the problem”
- Cardlytics' founders came from banking, including Capital One, so they understood how to navigate large complex bank organizations, which helped them win early banks such as Regions and Bank of America.
“Scott and Lynn came out of banking. They came out of Capital One, in some sense spoke bank.”
- Winning Bank of America and regional banks did not produce broad reach, because consumers on desktop banking were not digitally engaged, and mobile banking was where usage actually happened.
“It turns out that even with Bank of America and regionals, they still had really very little reach.”
- Cardlytics has been slogging uphill for years since the beginning, systematically working to overcome advertiser measurement skepticism, and its near-bankruptcies and nearly failed IPO were part of that struggle.
“Finally, they nearly went broke, I think, twice or three times. They had an IPO that nearly failed.”
- Because banks wanted control over advertisers and prefer large recognizable brands, Cardlytics had to persuade slow-moving enterprises like Starbucks instead of opportunistically signing whoever wanted a keyword, which constrained early adoption.
“Cardlytics had to convince Starbucks to do this. And that's harder.”
- At Cardlytics' Investor Day, the management team's skill sets were strikingly distinct, with a product leader, a banking-savvy communicator, and a charismatic salesperson, which made the team's complementary fit visible.
“it being so profoundly obvious how different the skill sets were of the different people involved”
- Cardlytics' early advertiser sales struggled because the channel was new, unclear whose job it was to buy it, and not important enough for anyone to cut red tape, so momentum only came with larger reach.
“it wasn't clear whose job it was, and no one was going to cut the red tape for something that didn't matter”
Advertisers undervalue randomized measurement, the main barrier to Cardlytics' growth.
- The biggest barrier to Cardlytics' growth is that advertisers do not understand the difference in measurement quality, because attribution models built on cookies and econometrics are unreliable, so advertisers haircut Cardlytics' numbers even when the randomized data is stronger.
“I'd say the biggest barrier they face is advertisers understanding the difference in the quality of measurement here.”
- Cardlytics can run randomized store-level or SKU-level experiments, such as icing selected stores or selling iced lattes, so advertisers see the effect in their own sales data and do not have to trust a report or understand a randomized trial.
“you can create a sandbox where you can show people an effect in their numbers, and then they don't have to rely on you interpreting the randomized controlled trial data”
- Multi-factor attribution models built from cookie tracking and econometrics measure association, not causation, so they can only be indicative, and they fail when TV and brand-building effects are ignored.
“like any association type of analysis it's flawed, right? I mean, association is not causation”
- Advertisers tend to frame choices as Google return X versus Cardlytics return Y, but Clifford argues this framing is wrong and the right approach is to optimize the whole business first, then maximize each measurable channel.
“I don't necessarily think that's the right framing, and it's a framing people do”
- Advertising buyers often understand RCT measurement intellectually but forget the distinction when trading off channels, and some even argue randomized trials do not work, so the measurement advantage is hard to convey.
“some of them will intellectually understand, but then as soon as you start talking about trade-offs between channels, they forget all about it.”
Take rate and margins should improve as bargaining power and scale increase.
- Across bank renewals, the direction has consistently been toward a better revenue-share mix for Cardlytics, which suggests the take rate can rise over time as the company gains bargaining power.
“the direction of travel and all the bank renewals ever has been for a improvement in the revenue share mix towards Cardlytics”
- Advertiser spend is reported as non-GAAP billings, and the 30-cent consumer share is not counted as revenue, so reported revenue is the 70 cents, from which the bank share is treated as cost of goods sold.
“So, a dollar that an advertiser pays, they call it billings. It's a non-GAAP idea.”
- Because the incremental cost of adding enterprise and agency clients is close to zero, operating margins should converge toward gross margins over time, aside from some added technology and sales spending.
“So margins over time ought to start to converge. Operating income margins ought to start to converge to more towards gross margins.”
- Cardlytics' gross margin is a little under 50 percent, calculated after the bank revenue share is treated as cost of goods sold.
“A little under 50%.”
Tipping points and feedback loops make Cardlytics' revenue path hard to model.
- Clifford says he has no idea what Cardlytics' revenue will be in five years because the business has many tipping-point effects and positive feedback loops, so any model he built would likely be wildly wrong and skewed.
“I have no idea what Cardlytics' revenue is going to be in five years. I mean, there's so many tipping point effects here, so much positive feedback.”
- For a system, it is often much easier to reason about the endpoint, or steady state, as time goes to infinity than about the path to reach it, and this is especially true for Cardlytics where the destination is clearer than the timing.
“you can often think about the steady-state of the system as time goes to infinity. Where do you go to? ... that is almost always a much easier analysis than understanding the paths”
- A business's value can be roughly thought of as its competitive advantage multiplied by its total addressable market, and Cardlytics is the clearest example of this framing Clifford has encountered.
“a business's value is something like its competitive advantage times its TAM”
- Once a business reaches a golden-ring monopoly position, it is largely theirs to lose, and the first decision an entrepreneur makes, which business to enter, says a lot about their decision-making quality.
“the first decision an entrepreneur makes is which business to go into.”
Long-run value depends on the steady-state endpoint, which is easier to analyze than the slow path.
- The world is highly competitive, so when a path to a great business is obvious it is usually already taken or fails for a good reason, and most great businesses discover their monopoly-like advantage only in hindsight, not at founding.
“And so usually in these stories, there is some sort of, we started the business thinking A, and by the time we got to step Y we realized that actually we could be a monopoly.”
- In systems analysis, identifying the steady state as time goes to infinity is almost always easier than analyzing the paths the system takes to get there, so investors should focus on the endpoint for businesses with many obstacles.
“you can often think about the steady-state of the system as time goes to infinity. Where do you go to? What's sort of the endpoint?”
- Clifford believes the channel's full potential, including personalized offers and deeper data use, will happen over 10 to 15 years rather than two years, based on how far the platform is from realizing its attention and data nexus.
“this is definitely a what can happen in 10 years situation, not a what can happen in two years.”
- A business's value is roughly its competitive advantage multiplied by its total addressable market, and steady-state analysis, asking where the system ends up, is much easier than tracing the path to that endpoint.
“a business's value is something like its competitive advantage times its TAM.”
- Sosin expects the Cardlytics opportunity to play out over 10 to 15 years rather than two, and he says he has no idea what revenue will be in five years and expects to be wildly wrong about the trajectory.
“this is definitely a what can happen in 10 years situation, not a what can happen in two years.”
Champion dependence and past cash constraints shaped client retention and hiring.
- For much of its history Cardlytics was constantly short of cash, so it could not hire top-of-market talent like Netflix pays, and when the company could finally afford stronger hires it had to restructure leadership and make room for them.
“the organization was constantly starved for cash”
- Competitor Dosh, which had a more modern tech stack and won some neobanks, forced Cardlytics to buy it rather than lose its monopoly position, and the company now runs a lighter platform for neobanks and a separate platform for legacy banks.
“And so the company ended up buying Dosh because they had basically fallen behind.”
- Cardlytics has repeatedly lost advertising clients when the internal champion left, got fired, or was replaced by a new CMO who did not understand or believe in the randomized measurement, showing the channel depends on individual buy-in.
“when they've lost a client, it's always been like, well, there was a champion for it”