Review - Subprime Attention Crisis (Hwang)

23 Jul 2023 advertising · economics · internet

Rating: 4/5

What would happen if all the free services that you use on the Internet, which are powered by advertisements, stop being free one day? This is the premise of the Subprime Attention Crisis. If you ask yourself this, you might realize that many of the services that we think of as free are powered by advertising. For me, the most frequently used services that are ostensible free are Google Maps, WhatsApp and YouTube. I use at least one of them almost every day. If they were to become paid services, or the hurdles to using them without giving up too much data increased, it would be mildly annoying. This calculus will be very different for an Internet user who uses Gmail and uses it to receive important communication. Hwang’s argument is sensible and easy to understand. He starts from the basics of advertising on the Internet and builds up to his myriad theses: Programmatic advertising is very similar to the financial markets. Ad networks claims that targeted advertising on the Internet is better than “spray everyone” advertising on TV. This claim is a lie and that banner ads don’t really change consumer behavior. Commercial interruptions are blocked by users using ad blocking plugins or because users reliably skip ads on video-only platforms in under half a second. To back all of this, he presents a lot of industry research and anecdotal evidence. This was a convincing case for being aware that free services could stop being free any day, and there would be nothing really surprising about it. (Just the other day, YouTube took a step towards blocking ad blockers.)

Concrete examples are helpful to set the stage:

• Gmail is a free, e-mail service that launched in 2004. In the past 19 years, it has grown rapidly and now has 1.5 billion active users. Gmail.com has become the most common e-mail domain. Gmail is run by Google, a search engine company, which had a total revenue of $70 billion between January and March 2023. 78% of this revenue came from Google’s advertising network.¹
• Facebook (now Meta) is the company behind 3 of the most common messaging and social media apps today: Instagram, Messenger, and WhatsApp. Facebook was launched in 2004, WhatsApp in 2009 and Instagram in 2012. Both Facebook and Instagram are powered by ad revenue, while WhatsApp is a popular end-to-end encrypted, ad free messaging platform. 3 billion people used at least one of these apps in March 2023. Facebook’s total revenue between January and March 2023 was $28.6 billion. 98% of this revenue came from advertising.²

What is Programmatic Advertising?

Programmatic advertising is a marketplace with buyers and sellers. Publishers (like the New York Times news website) sell space on their websites, where ads can be displayed. Buyers (like a local business) buy this space to display an advertisement about the goods or services that they provide. The commodity that is being traded (in the abstract) is the attention of the person viewing the publisher’s web page. The process that decides which ad should show up on a webpage when a user is visiting it is called Real-Time Bidding (RTB). RTB is a sub-100ms auction where the buyers’ ad networks look at the data that is available about the user and decide how much they are willing to pay to show a given ad to them. This process is easy to imagine when a few transactions are happening in an hour. But the process happens at a much larger scale and is completely automated. The technology behind it is probably complex, and (as Hwang demonstrates) publishers often don’t even know which ad network served a given ad on their website.³

Standardizing the Marketplace

There is one hiccup to setting up this marketplace. When exactly can we say that an ad has been shown to the user? In other words, when exactly has the user’s attention been captured? This is clearly hard to define. In a long news article, if the ad was displayed after the 10th paragraph but the user stopped reading the article or clicked on another link after looking at the headline, the ad was not viewed and the publisher should not be paid for placing an ad there. These things are specific and explain what ads should look like and when they should be considered to be “delivered.” I looked through the website of the Interactive Advertising Bureau (IAB), which sets these standards to find 2 interesting examples:

For ads that are delivered to a desktop web browser, ad measurement happens when the ad content has been downloaded by the client and counted as delivered when it begins to render. The actual guideline goes into excruciating detail and tries to define everything:

An Ad Impression across all display marketing channels is the measurement of responses from an ad delivery system to an ad request from the user’s browser, which is filtered for invalid traffic and is recorded at a point as late as possible in the process of delivery of the creative material to the user’s browser. The ad must be loaded and at minimum begin to render in order to count it as a valid ad impression. Measurement of begin to render should include logical components necessary to display the ad, but does not necessarily include logical elements that are not essential (such as other tracking elements)

– https://www.iab.com/wp-content/uploads/2017/11/Desktop-Display-Impression-Measurement-Guidelines-US-MMTF-Final-v1.1.pdf (p.5)

For podcasts (an audio medium), there are two metrics. The first one measures whether the file containing the ad was delivered:

Ad Delivered: an ad that was delivered as determined by server logs that show either all bytes of the ad file were sent or the bytes representing the portion of the podcast file containing the ad file was downloaded.

– https://iabtechlab.com/wp-content/uploads/2021/03/PodcastMeasurement_v2.1.pdf (p.16)

While the second one measures whether the podcast client can confirm that the ad was played:

Client-Confirmed Ad Play: counts an ad that was able to prompt a tracking beacon from the client when the file was played. Whenever possible, metric should include information about how much of the ad was played using the markers: ad start, first quartile (25%), midpoint (50%), third quartile (75%), and complete (100%).

– https://iabtechlab.com/wp-content/uploads/2021/03/PodcastMeasurement_v2.1.pdf (p.16)

Like commodities trading, the attention that is being traded in these IAB standards is different from the underlying “thing.” It has been standardized and defined in an effort to attain an unattainable perfect commodification of the user’s attention.

How is Programmatic Advertising Different?

Advertising is not a new concept. Newspapers have been funded by classified sections and outright advertisements for at least a century. Hwang identifies 3 things that are different about programmatic advertising:

1. The speed at which advertisers decide whether or not to serve an ad to a user on a webpage is much faster than before. Before buying an ad on a television show, or a newspaper, ad buyers would generally do months of research about viewership/readership data. This data was not perfect either, but the time that these decisions took was much longer.
2. The volume of ads that are served to users per minute has increased. On TV, a 30 minute program has 22 minutes of content and 8 minutes of advertisements. Online, a single news article which one would read in about 5 minutes can have 2 banner advertisments. Video streaming platforms are worse with 30 second advertisements for videos which are less than 5 minutes long in some cases. There is no way to say how much exactly the volume has increased. But the lack of this data should not prevent us from noticing that the number of ads we see on a daily basis has skyrocketed.
3. The granularity with which ads can be placed within a piece of content has increased. A newspaper has a given section for Classifieds, magazines assign pages or the back cover for advertisements, and TV shows demarcate content and advertisements very clearly. (“Mute the commercials, will you?”) Online, where ads are placed can be controlled right down to the paragraphs between which the ad shows up or the section of a video in which the ad would be most effective.

Similarities to Finance

The executives who were building advertising on the early Internet and trying to figure out how to make these new services profitable were from the finance world. For e.g., Sheryl Sandberg, widely regarded as the person who was responsible for teaching small businesses to use programmatic advertising, had a background at the World Bank and the United States Treasury. The model that they chose to replicate for programmatic advertising was the one of high-frequency trading.

The most important similarity is market opacity. When the commodity that is being traded is not well understood or its value is misunderstood by market participants, a crisis of confidence is always looming around the corner. This is what Raghuram Rajan and Reinhart have both argued are the prerequisites for the creation of a bubble in any economy. The crisis of confidence will trigger the boom/bust cycle as in the financial markets, as marketers realize that the money they spend on programmatic advertising is not delivering the returns that they were promised and stop spending as much. One question here is whether the old markets of advertising (television, radio) were not just as opaque? Of course, they were opaque too. However, the marketing that was done on TV or radio did not claim to be targeted or very good at anything except raising brand awareness in the first place. A famous ad industry adage is “Half of the money I spend advertising my product is wasted. But I don’t know which half.” The opacity in the market was accepted as a condition. Programmatic advertising makes unattainable claims about being targeted and measurable; while simultaneously, publishers are not sure which ad network is delivering what ads to their users.

Another similarity that Hwang talks about is the “private market.” Companies like Facebook and Google have premium ad space which is sold on private markets. This limits the information available in the public market, and makes it hard for advertisers to figure out how much a given webpage is really worth paying for.

A Bubble in The Attention Economy

First, ads don’t work. When WIRED, one of the early websites, ran a banner ad, 40% of people who saw it clicked the ad. Today, even 1% is a pipe dream, with the average being about 0.2-0.4% of viewers engaging with ads. One way to evade this critique is by saying that advertisers are doing “brand advertising.” Brand advertising simply focuses on showing people brand names and products and does not aim for any change in behavior at all. This allows advertisers to make the duplicitous claim that programmatic advertising works at increasing brand awareness, without measuring anything or changing anybody’s behavior.

Second, ad blocking is on the rise. Google admitted in 2014 that 56% of ads that were displayed are never seen by a human. 800 million devices had ad blocking enabled in 2022. The demographic trend is also discouraging for advertisers: Younger users and wealthier users use ad blockers at a disproportionately high rate. In this section, Hwang reveals another stunning fact that makes total sense once you read it: Snapchat’s users skip commercial interruptions within less than a second. Less than a second.⁴ Snapchat’s users skew younger and are clearly very ad-averse.

Third, click farms inflate the value of advertisement in some domains. There are also the more technical problems of domain spoofing, where unreliable and obscure ad networks sell an add on the domain NYTemis.com, but brand it as one that is going to appear on NYTimes.com. Buyers and publishers are not able to detect this behavior, and the technical solutions which prevent such domain spoofing when visiting websites is still not available when buying ad space.

Fourth, middlemen. There are resellers and middlemen who add a markup for no value at all. The marketing agencies which used to be premium places for creating ads have lost ground in that area, but they are active in this market of procuring discounted ad space from the ad networks, adding a mark up, and reselling the space to some buyer. These middlemen get a commission on the ad space that is bought and are completely disconnected from any effect of whether the ad actually changes consumer behavior. (Does this sound vaguely familiar? In the movie The Big Short (2015), about the 2008 financial crisis, these middlemen are the mortgage brokers in Florida who were selling condos by giving subprime loans to anyone with a pulse, and pocketing the commissions on these loans.⁵ The middlemen brag about the loans as “NINJA loans: No Income No Job” and talk about how the bonuses on these loans started skyrocketing a few years ago. In fact, Steve Carrel’s character asks them the very question that Hwang complains that no one is asking in the programmatic advertising business: “Do people have any idea what they are buying?”)

The way that Hwang writes, the similarities between the financial markets and the advertising market, and the time before the crisis in the financial market and now in the advertising market, become quite clear. The whole book is obsessed with bringing these similarities out and showing what is going on.

What Next?

What is an advertisement supposed to do? Ignoring the amorphous category of “brand advertising,” all of advertising has the same goal: Change the viewer’s behavior in a predefined, intended way. The online attention economy is based on the confidence that ads on the Internet, bought through ad networks and sold through real-time bidding, are more effective than old-school, “spray everybody” advertising. A crisis of confidence will happen when buyers stop advertising through this method and realize that it does not affect their bottom line adversely at all. Buyers and sellers have already started doing this⁶ and they are not seeing their revenues drop.

The ways that we use the Internet are constrained to the mediums in which advertising can spread: text with engagement signals like “Like” and “Favorite”; quantifiable popularity metrics like follower and subscriber counts. A place that enables computers to connect to each other without commodifying everything that people do in that place does not necessarily have to look like our current Internet. The way that the Internet is funded is important to the way that the Internet is. The algorithms put inflammatory posts that make people mad at the top of their feeds because that is what sells advertisements and motivates people to take action. Hwang’s optimism is touching here: He says that the Internet had other routes; it could have relied on other sources of revenue. I believe him. People could have run their own servers if the technology had been simple or if the education ministries had kept up with the technology. Long-form blogs would have been the order of the day, rather than “outrage of the hour” sentences. People would have found it harder to target others because interaction was curated and disperse.

We should care about how the programmatic advertising market works because it is the main source of funding for many free services today. This article links to a few YouTube videos and news articles, which remain free because YouTube and publishers have ads on their website, which they believe is making them enough money that they don’t want to stop people from just sharing links around. If that wasn’t the case, how would you share video clips with everyone? Hosting is not free, and neither is bandwidth. The collapse of this market would change the Internet completely. And possibly, for the better.

I keep returning to Hwang’s premise for this book. What would happen if all the free services that you use on the Internet, most of which are powered by advertisements, stop being free one day?