Four lessons Criteo, the $3.5bn ad-tech company, can teach us about machine learning business models

With the rise of ML/AI as the start-up flavor du jour, I’m seeing more and more companies that have software as their core, but business models that are not traditional SaaS business models. Instead, these companies seek to provide a direct benefit or ROI to customers without a user interface. In this post, I’ll discuss the business model of a successful company that employs this strategy, Criteo, as a way to think through the pros/cons and sources of long term competitive advantage for this type of business model.

First, let's define the category of companies I am describing, which are intrinsically different from SaaS. Traits include:

1) Software that runs in the background of other apps or as a managed service

2) Relatively short/painless implementation times (due to little need for user training)

3) Easily quantifiable performance (often a/b testable)

4) Some form of algorithm/machine learning is core to driving value

I don’t have a great name for business models like this. If you do, shoot me an email.

About Criteo:

Criteo is an ad-tech company that uses data collected from browsing activity on e-commerce websites to "retarget" users with ads from the same site. This is a fundamentally high-ROI activity due to the intent captured from a user's previous browsing history, and Criteo has it down to a science. Machine learning features heavily into the service, as Criteo dynamically recommends products, a/b tests creative and forecasts both a user's propensity to spend and the likely cost of an ad impression for trillions of ad impressions per year in real time. I highly recommend reading the company's investor presentation, as it is the most important internet company many haven’t heard of, running silently in the background at web scale.

Criteo is the epitome of the business model I described earlier. It is delivered as a managed service, and is evaluated solely on the benefit it provides to customers: sales in the form of post-click conversions for a given spend per click. If a competitor credibly claims it can do better, there will be an a/b test and the winner gets the business.

On the surface, these traits put together seem to be a challenge for building a sustainable business, and indeed for much of ad-tech they have been. Technology leads are hard to maintain over time as a form of competitive advantage without patents, and it is hard to earn meaningful profits when customers can switch so readily to other providers with minimal disruption. This is especially true when one considers that Criteo charges its customers ~40% of what it spends on media for the privilege of utilizing the service.

In reality, Criteo's financial results to-date show unqualified success. It is a behemoth, on track to nearly $1bn of revenue this year on $2.5bn of ad spend, with a $3.5bn market cap. It is growing at a ~30% clip, at a 30% Adj. EBITDA margin. It faces little independent competition for its core customers, which include many of the largest online retailers in the world (even some Amazon subsidiaries use the technology).

How does this square up with the challenges outlined above?

First, there are clear advantages to the model: Criteo can grow revenue frictionlessly, as its solution is “always on” and often run with an “uncapped budget” for a given level of ROI. As customer traffic and revenue grow, so does their ad spend on retargeting. As Criteo's algorithms improve, it can deliver more conversions at a target ROI that customers are willing to pay. The same goes for adding advertising inventory to the platform. Unlike traditional software models, Criteo needn't convince users to adopt a new module to broaden its deployment: as it provides more value, the dollars automatically flow in. There's an "easy come" to pair with the "easy go" of lower switching costs.

But how does Criteo maintain enough of a lead over other well-funded platforms to keep its clients at a 40% take rate? There are some specific market dynamics that have helped Criteo maintain a product advantage over time:

1) Network effects: Criteo's scale allows preferential access to some ad inventory. Criteo also has a data co-op with its customers that allows it to match users across devices using their pooled login data, a valuable feature other independent players don’t have. It both uses this data co-op to power its products and gives the data for free to its clients to help with their own marketing efforts: a sly move to build a tangible network-effect into the business.

2) Barriers to dual-sourcing: Customers generally don't run multiple retargeters on the same audience at once, for technical reasons including frequency capping and competitive bidding.

3) A really deep problem: Retargeting has huge opportunity for innovation and optimization, even today: dynamic creative, better product recommendations, etc. can all have measureable ROI impacts. Put simply, the product is far from perfect, meaning Criteo can maintain a lead over the competition more easily.

4) ROI is not the only consideration: Criteo customers are generally dual optimizing for overall revenue and ROI. To beat Criteo (given that dual sourcing is a challenge per #2), a competitor would need to demonstrate superior ROI and at least the same level of sales. Much of retargeting is arguably commoditized: "show a customer the product they just looked at, but didn't buy, x times," but Criteo still earns a 40% take rate on even this business because it is better at the hardest stuff than any of the competitors, and customers can't afford to cede the difference that makes. The calculated return per dollar of ad-spend might be higher, but few retailers will cede valuable incremental revenue just to maximize that metric.

Note: If dual-sourcing were feasible here, Criteo’s TAM would be much smaller, as a lower-cost competitor would likely do “basic” retargeting at a much lower take rate and leave only the technically challenging ad impressions (where the decision is borderline) to Criteo.

5) The giants can’t compete on a level playing field: Online advertising is dominated by giants: Google, Facebook and Amazon. Happily, none of them are equipped to compete head-on with Criteo. Facebook and Google are walled gardens, with limited access to each other’s inventory. Further major retailers are reticent to give full customer/catalog data to companies that they are already so beholden on. Amazon is a non-starter for obvious reasons. This provides an opening for an independent company to compete.

Lessons for other startups:

Going through the points above, I see four lessons for startups looking to provide services like Criteo:

1) The underlying problem must be deep enough that there is ample room for continual R&D progress for the foreseeable: if the rate of improvement plateaus that is a warning sign.

2) Hard problems are rarely huge markets on their own: instead they can be a spearpoint that provides access to a larger pool of value/spend (the way Criteo owns even obvious retargeting use-cases). Dual-sourcing can limit the value of a spearpoint.

3) The spearpoint should ideally derive from a durable competitive advantage, like an economy of scale or a network effect (in this case, Criteo’s data co-op is a plausible candidate).

4) There should be a reason for independence that justifies the company’s existence in the face of competition from the nimble giants.

The good news is that Criteo’s example has shown that it is possible to create a massive business built on machine learning algorithms with low switching costs and minimal user-interface.

978 views0 comments

Recent Posts

See All

A framework for modeling product development

This post is written to crystallize my own thoughts, and isn't particularly groundbreaking. I have no particular experience in product management and appreciate thoughts/insights from anyone who does.