Programming Note: The Diff will be off Monday, July 5th for Independence Day. Back Tuesday.
Chumboxes or Trust-Busters? Taboola and Outbrain
A common sight on news articles is the unappetizingly-named chumbox, a block of recommended content that ranges from clickbait news to extreme clickbait news to ads for products of rather variable quality. That, at least, was what the chumbox was when the term was coined; we're stuck with the term, but the content is getting better. Two leading companies in the space happen to be going public at about the same time: Taboola has just merged with a SPAC (detailed proxy here) and Outbrain filed an S-1 Tuesday. Conveniently, the models and financial statements are similar—the companies were actually negotiating a merger, which broke off at the end of 2020. Very conveniently, Taboola's route to the public markets lets them talk openly about long-term projections and give lots of interviews, so there's a fair amount of material about how they think. Outbrain has had to be a little more careful, but has partly compensated by providing a nice manifesto defending the entire business.
The way these companies describe their business is that they partner with publishers to recommend content, and they intersperse those recommendations with ads. The business model is to share the revenue, with most going to publishers (Taboola collected 35.6% of the revenue its ads generated last year, Outbrain got 25.3%). In the early days of this business, the ads were exactly the sort of ads that would maximize clicks and revenue: gross-out cures, implausible financial promises, fake news—which Outbrain banned in 2014, a costly move at the time but a prescient one given how effectively real news competes with the fictional kind lately. When they recommended stories, the stories were also clickable: read an article about CEOs, and you might get a contextually-relevant article about another CEO's scandalous downfall; read anything remotely Kardashian-adjacent, and you're sucked into the Kardashian news vortex.
Welcome to Chumworld
Many small and scrappy companies start out with an anything-goes approach to content or monetization; better to get money in the door by whatever means necessary than to focus on the ten-year outlook when the runway is measured in months. It's almost a rite of passage for them to give up on this: when eBay announced the Paypal acquisition, they also announced that Paypal would be phasing out payment processing for online gambling; Tumblr restricted some of its most popular content after getting acquired; Reddit has slowly restricted access to, or banned, dodgier subreddits as it's grown; and just yesterday, Pinterest banned weight loss ads. Relying on direct-response advertisers has some benefits: they are more metrics-driven than brand advertisers, so they produce more clickthrough data, and they're certainly aware of their unit economics. If their ads are working, they’re willing to scale up fast. These businesses are constrained on brand-building though—if you had a successful company that sold dodgy diet pills or real estate seminars, would you want your brand to be a household name?
Because this business mixes editorial content and ads, it fits into the advertorial category, and many of the ads did (and some still do) follow this pattern, too: they look more like news stories than product pitches, but there's still a call-to-action. Advertorials in general pitch a wide range of products, all the way from supplements, low-value personal finance, and other direct response classics to ads for countries that ran in major magazines (The Economist eventually stopped accepting these.) They don't have a great reputation among journalists, because blurring the editorial/business distinction is always fraught. But they do have a good reputation among marketers, because they convert so well (perhaps for the same reason).1
Part of why advertorials can convert well is that they're longer than other ads, and long-form content can convert surprisingly well. Some people skim it, but for some readers there's a hypnotic effect to reading thousands of words of copy pitching a product. Content recommendation produces a similar kind of hypnosis on a different scale: once you've made one aimless click, you can make another, and another, and each one raises the odds of making a sale.
Escape from Chumworld
It's hard to start with a downscale brand and make it upscale, and it's often easy to grow a company temporarily by moving downmarket. But content recommenders can largely avoid this problem, because their brand is practically invisible. "Powered by Outbrain" does show up on some sites, but it's unobtrusive. Especially because everything next to this branding has been continuously optimized to be as clickable as possible. This has given the content recommenders an opportunity: they can unilaterally go upscale, albeit at the cost of missed revenue. Their customers are making an ROI decision, and will want to know about content quality, but as long as the content recommender can plausibly claim to keep the worst stuff out of a given ad unit, they're fairly safe. And they're well below Facebook/Google/Amazon scale; Outbrain has over 20,000 advertisers, and Taboola has over 13,000. Or, put another way, Outbrain spends about $4,000 per year in sales and marketing per advertiser, and Taboola over twice that; at that spending level, they're applying some amount of human vetting to everyone they deal with.
So they've upgraded. I chose a big news site at random from this list of Outbrain users, and the advertisers include a big furniture brand, a few (less prestigious but not terrible) news sites, a car dealer, and enterprise software seller Splunk. (And, to be fair, a bidet company, albeit a rather fancy one that's also the top advertiser in the category on Amazon.)
These companies work with increasingly classy advertisers because they have to (big publishers are more likely to stick around than small ones), but also because they can: content recommenders collect data, and much of it consists of first-party cookies rather than third-party; they're insulated from regulatory changes and Google policy changes that make it hard for advertisers to track people across sites. And they track a nontrivial number of people; Outbrain has 1 billion monthly active users, and Taboola has 500 million daily active users.
Taboola's CEO likes to talk about "the difference between Artificial Intelligence (AI), Deep Learning (DL), and Science Fiction (SF)" (although he clearly prefers a different phrasing for the last one). At a smaller sample size, like the amount of data an individual publisher uses, there may not be enough data to get meaningful signals on what people are interested in. One click could be due to underlying interest, a compelling headline, the photo, or a mistake. But content recommenders are using the entire set of data for all of their customers, so site A can use data sourced from sites B and C through Z. As advertising restrictions tighten up, this becomes much more important: a media company that runs ads, like Taboola or Outbrain, gets treated differently from a legal and technical standpoint than a company that purely bids on ads. The flexibility of the adtech stack has allowed companies to proliferate, specialize, and create sometimes-lucrative niches, but a shift away from third-party cookies and to first-party cookies benefits companies that can handle the full stack of publishing and advertising themselves. This applies to big media companies, but also to the content recommenders.
Having better-quality ads doesn't just land better customers; it can improve the business's long-term economics. Everyone says they care about user experience, but content recommenders are training people to either ignore or click on blocks of recommended content over time. If a click produces more future clicks, that produces future revenue, even if the first click doesn't extract as much value as it conceivably could.
The Outbrain S-1 has a manifesto of sorts about the downsides of a purely price-focused approach to choosing which ads to display. Readers, the piece notes, never say "This ad is so delightfully priced! I Have to spend more time on this site!" (Many other ad platforms do use clickthrough rates as a criterion for ranking ads, in part because it's a quality filter and in part because it's profit-maximizing.) But it does point to something important about the model: because Outbrain and Taboola are operating on specific sites with which they have a profit-sharing relationship, the P&L doesn't end when a click happens. They need to optimize each display to maximize the net present value of the viewer over time. This, too, is something Google and Facebook care a lot about; they do raise ad load over time, but carefully, and they're especially happy to do it when the quality of ads is similar to that of content. Instagram ads look like Instagram posts, and that's in Instagram's interest. The letter is worth reading in full; a good long-term KPI to shoot for is to be able to write a principled defense of your business in the S-1, and mean it.
And this shows up in their financials. In an interview, Taboola's CEO says their publisher-level dollar retention averages 110% to 120% annually (their investor deck has more; 2019 was lower at 100%), and Outbrain gives theirs as 108% in 2018, 95% in 2019 (or 104% excluding one customer they say they left willingly), 104% in 2020, and 123% in the first quarter of 2021. These are not the metrics of a business that's helping customers churn through readers to maximize profits while turning them off. (For one thing, it's hard to have dollar retention over 100% if most of your customers are desperate enough to run the scummiest ads.)
So it's gone from a low-quality product to an acceptable one. What kind of business is it?
Reverse Search Engines
Taboola's founder says his original inspiration was that he couldn't find something to watch on TV, and realized that TV channels should be looking for him, not the other way around. Content recommendation companies aim to solve this problem by guessing what people want to read next and showing it to them. The search engine comparison isn't specious, either: in both cases, there's a mix of paid and organic results, and the data from organic interactions can inform rankings for paid ones. What you click on when you search for "Chicago hotels" can help Google infer whether the ads you see should pitch you hotels still, or air travel, or car rental, or something else. When you choose a celebrity weight loss story from a Taboola recommendation, Taboola knows that you're more likely to click on diet-related ads. When you read a story about home prices, Taboola knows that’s an option to show you some very lucrative mortgage ads.
While the companies position themselves as a Facebook/Google alternative for publishers—a very smart decision given how media sites feel about those businesses!—the symmetries don't end there. Content recommendation sites are still aggregating content, and still influencing the flow of traffic between sites. The difference between them and the traditional aggregators is that Google and Facebook control this upfront, and Taboola and Outbrain do it after the fact. If Google and Facebook are first-click aggregators, Taboola and Outbrain are second-click-onwards aggregators.
The business is very different from Facebook's and Google's. Content recommenders have to win their audience by cutting deals with publishers, not just by providing a product users like. That makes it hard to grow incremental margins much over time. But the positive feedback loops are very similar to these big companies: as they work with more publishers, they get more data—and their data advantage relative to the publishers keeps going up. That lets them target ads better, and target content better, too. In fact, they can slowly reshape the entire media ecosystem, by recommending that users visit sites they're more likely to keep visiting in the future, thus generating more revenue. Meanwhile, the economic alignment keeps publishers happy. Everyone who gets traffic from Facebook knows that Facebook does not particularly care if there are losses and layoffs at any given outlet, as long as content is being produced—and Facebook wants a healthy chunk of that content to be created by users, not created offsite and just shared on the platform. Taboola and Outbrain end up acting as gardeners for the media industry, pulling the occasional weed, keeping the overall system free from pests, fertilizing the most promising sites a bit, and harvesting profits.
If this story is to be believed, the reason the merger was called off was because Taboola couldn't get financing, after making some disruptive changes to their model in response to Covid. They now have financing, and they've already gotten DOJ approval for their merger. Given their very similar models and fairly similar economics (Taboola collects a bit more revenue from publishers2, but Outbrain earns slightly higher margins overall) and similar size, a merger would not be too complicated to integrate. Data would compound faster if they were the same company, and while publishers might complain, there's a big gap between how they feel about Outbrain and Taboola and how they feel about Google and Facebook. The data advantage would compound faster if it were more comprehensive, and both companies would have more pricing power with fewer competitors. An eventual merger may not happen, but it's an interesting source of upside: as we've seen with Facebook and Apple on privacy, sometimes the most efficient solution to the problems of one monopoly comes from creating another monopoly with different interests.
The Wall Street Journal highlights the trend of hiring bonuses for low-wage jobs ($). The most striking datapoint: "Nearly 20% of all jobs posted on job search site ZipRecruiter in June offer a signing bonus, up from 2% of jobs advertised on the job search site in March." This could be just another tight-labor-market story, but there's something else to it: a hiring bonus is not just a way to raise wages, but a way to raise them temporarily. What businesses are assuming is that the market-clearing wage in mid-2021 is higher than it will be in 2022. They still want to get workers in the door, but they're not as worried about keeping them.
Stay tuned for much more on this in a few days, but Robinhood has filed to go public. One notable deal on the offering, rather than the company: they're setting aside 20% to 35% of the offering for retail investors. Since retail investors drive disproportionate IPO pops, this is smart for two reasons: first, of course, it's raising money where the demand is. But second, many of these retail investors are on Robinhood, and will be irritated if they buy stock in their favorite company and lose money. So Robinhood's incentive is to raise as much as possible in the IPO, in order to keep the price low and the pop modest. All that retail investment will make the stock a more reflexive one than most; an increase in their stock price will mean happier customers, more trades, and more fees. It will be especially interesting to see if this conditional probability is reflected in options markets, with a steeper-than-usual volatility smile reflecting the fact that when Robinhood's stock drops, its business is worth less as a consequence.
Incidentally, I'm always impressed with people who can tweet a link to the S-1 the second it drops. It took me two guesses.
Part of what makes the economics of search great is that searchers are a fragmented market, and destinations are also fragmented, but search itself has economies of scale. Being a natural monopoly that intermediates between roughly everyone in the world and roughly everyone else is a profitable position to be in. And one illustration of this is the search engines that deal with more consolidated suppliers: American Airlines is suing reservation search provider Sabre over misclassifying some of its seats, leading to unfavorable comparisons with other airlines. Every search product has to implicitly make judgments: a search takes the subset of all possible content that could be relevant to a given query, and then ranks it, which requires some implicit metric of goodness. For vertical search engines (a topic addressed in The Diff here ($)), that means having topic-specific metrics that can be used to sort and organize results. Trying to be objective doesn't always settle debates; sometimes it just changes their terms. This dispute will presumably get resolved, at a great cost in time and money; one contributor to Google's margins is that it's hard to sue them for rankings you don't like, and the algorithm is complicated enough that nobody can have an informed argument about what factors led to what exact outcome.
1-to-M-of-N Secret Sharing
Twitter is testing "personas" that can limit access to some tweets but not others. The default mode for Twitter is that each user broadcasts to everyone in the world. (This is very helpful for anyone who wants to prove that a given opinion is being expressed online, either as the default or as an example of just how dumb people are—any response you can dream up to a current event has probably been tweeted by somebody.) But as people get more worried about privacy, and about the long tail of consequences to Internet comments that take a few seconds to write and are forgotten within the hour, the best broadcast sites have to find ways to reach a narrower audience. Twitter's direct messaging feature already helps with this, but has the unfortunate downside that it's hard to advertise in. Having public, semi-public, and private Twitter personae will mean that users can broadcast some things, and share other stuff more narrowly, all while producing monetizable ad inventory.
Home sales are up in units, but actually peaked in December 2020 in terms of value sold, at a level below the non-inflation-adjusted housing bubble peak. Scarce materials and workers mean it's more economical to build smaller homes, or at least more feasible to finish them once they're in progress. A related phenomenon illustrates how hard it is to profit during a cyclical upswing: small contractors are missing out on the building boom because they don't have as many permanent workers, and can't get materials ($, WSJ), while larger ones have more full-time employees and long-standing supplier relationships. Part of managing a cyclical business is being able to wait out the bottom of the cycle, but another important part is expanding at the right pace and time. Many cyclical companies end up destroying shareholder value through top-of-cycle acquisitions, but smaller ones can miss out if the most constrained part of the supply chain is the one they’re buying from rather than selling into.
Full disclosure: if you squint, The Diff evolved from an advertorial. Before the product was subscriptions and ads, the product was me; I was writing to find a good job.
Which, technically, worked. ↩
Both companies disclose "Ex-TAC Gross Profit," a non-GAAP measure that just reports the cut of ad revenue they get from publishers, and excludes other costs of sales. It's a meaningful metric: in a footnote, Taboola says that the institutional investors they raised from alongside their SPAC deal didn't even need to see projections for gross profits according to standard accounting metrics, just this metric. ↩