Thursday, November 12, 2020

Even More On The Ad Bubble

I've been writing for some time about the hype around online advertising. There's a lot of evidence that it is ineffective. Recently, the UK's Information Commissioner's Office concluded an investigation into Cambridge Analytica's involvement in the 2016 US election and the Brexit referendum. At The Register, Shaun Nichols summarizes their conclusions in UK privacy watchdog wraps up probe into Cambridge Analytica and... it was all a little bit overblown, no?:
El Reg has heard on good authority from sources in British political circles that Cambridge Analytica's advertised powers of online suggestion were rather overblown and in fact mostly useless. In the end, it was skewered by its own hype, accused of tangibly influencing the Brexit and presidential votes on behalf of political parties and campaigners using its Facebook data. Yet, no evidence, according to the ICO, could be found supporting those specific claims.
Below the fold I look at this, a recent book on the topic, and other evidence that has emerged since I wrote Contextual vs. Behavioral Advertising.

The ICO's conclusions are summarized in a letter to the chair of the relevant Committee of Parliament:
SCL’s own marketing material claimed they had "Over 5,000 data points per individual on 230 million adult Americans." However, based on what we found it appears that this may have been an exaggeration.
...
while the models showed some success in correctly predicting attributes on individuals whose data was used in the training of the model, the real-world accuracy of these predictions –when used on new individuals whose data had not been used in the generating of the models –was likely much lower. Through the ICO’s analysis of internal company communications, the investigation identified there was a degree of scepticism within SCL as to the accuracy or reliability of the processing being undertaken. There appeared to be concern internally about the external messaging when set against the reality of their processing.
See also Thom Dunn's UK ICO report on Cambridge Analytica finds no illegal activity or Russian involvement and Izabella Kaminska's ICO's final report into Cambridge Analytica invites regulatory questions.

But these conclusions haven't stopped Cambridge Analytica's successors hyping their wares. Alex Pasternak's This data expert helped Trump win. Now he’s built a machine to take him down reports on a 2020 version:
The goal is to use microtargeted ads, follow-up surveys, and an unparalleled data set to win over key electorates in a few critical states: the low-education voters who unexpectedly came out in droves or stayed home last time, the voters who could decide another monumental election.
...
“We’ve been able to really understand how to communicate with folks who have lower levels of political knowledge, who tend to be ignored by the political process,” says James Barnes, a data and ads expert at the all-digital progressive nonprofit Acronym, who helped build Barometer. This is familiar territory: Barnes spent years on Facebook’s ads team, and in 2016 was the “embed” who helped the Trump campaign take Facebook by storm. Last year, he left Facebook and resolved to use his battle-tested tactics to take down his former client.
...
Acronym was first out of the gate, and is thought to be the Democrats’ most advanced digital advertising project. By the election it promises to have spent $75 million on Facebook, Google, Instagram, Snapchat, Hulu, Roku, Viacom, Pandora, and anywhere else valuable voters might be found.
These whizzo hi-tech schemes are attractive to people with money. All they need to do is to sign a few large checks. A few highly-paid consultants, their kind of people, wave their magic wands and the right kind of voters stream to the polls.

But the results of the recent election suggest that this doesn't really work. What really does work is the kind of grass-roots person-to-person organizing that Stacey Abrams used to flip Georgia. But that means signing lots of small checks and working with lots of awkward people, a much less attractive proposition. No-one in Stacy Abrams organization is raking in the big bucks, so there are no marketeers hyping their product.

Someone else who has noticed the gap between hype and reality in online advertising is Tim Hwang. In Ad Tech Could Be the Next Internet Bubble, Gilad Edelman reviews his new book Subprime Attention Crisis. Hwang:
lays out the case that the new ad business is built on a fiction. Microtargeting is far less accurate, and far less persuasive, than it’s made out to be, he says, and yet it remains the foundation of the modern internet: the source of wealth for some of the world’s biggest, most important companies, and the mechanism by which almost every “free” website or app makes money. If that shaky foundation ever were to crumble, there’s no telling how much of the wider economy would go down with it.
The argument of Hwang's book is that, because the online advertising marketplace was designed by economists such as Google's Hal Varian by analogy with the financial markets, they inherited pathologies similar to those that caused the 2008 financial crisis. And thus the market is likely to suffer a similar kind of meltdown, which would have a huge impact on the online environment:
Intense dysfunction in the online advertising markets would threaten to create a structural breakdown in the classic bargain at the core of the information economy; services can be provided for free online to consumers, insofar as they are subsidized by the revenue generated from advertising.
...
a sustained depression in the global programmatic advertising marketplace would pose some thorny questions not entirely unlike those faced by the government during the darkest days of the 2008 financial crisis. Are advertising-reliant services like social media platforms, search engines, and video streaming so important to the regular functioning of society and the economy that they need to be supported lest they take down other parts of the economy with them? Are they, in some sense, "too big to fail"?
Hwang's prologue features a talk to the Programmatic I/O conference by Nico Neumann, who:
begins by showing an analysis done by him and his collaborators auditing a sample of the third-party consumer data—also known as the record of everything you supposedly do online—that form the basis of online ad targeting. When compared with verified data about those same consumers, the accuracy was often extremely poor. The most accurate data sets still featured inaccuracies about 10 percent of consumers, with the worst having nearly 85 percent of the data about consumers wrong.
Hwang surveys the vast array of evidence of online advertising's excessive hype, including:
When they first launched in 1994, the first banner ads generated a remarkable click-through rate of 44 percent. ... One data set drawn from Google's ad network suggests that the average click-through rate for a comparable display ad in 2018 was 0.46 percent. ... Recent attempts to measure click-through rates on Facebook ads reveal similar rates of less than 1 percent. ... Even these sub-1-percent click-through rates may overstate the effectiveness of ads on some platforms. On mobile devices close to 50 percent of all click-throughs are not users signaling interest in an advertisement, but instead accidental "fat finger" clicks—users unintentionally clicking on content while using a touch-screen device. Ads may also drive a response among only a small segment of the population. In 2009, one study estimated that 8 percent of internet users were responsible for 85 percent of all advertisement click-throughs online.
And:
In 2013, a controlled experiment on more than a million customers to evaluate the causal effect of online ads concluded that a customer "between ages 20 and 40 experienced little or no effect from the advertising". This was in spite of this demographics's proportionally heavier usage of the internet. In contrast, the study found that customers older than sixty-five, despite constituting only 5 percent of the experimental group, were responsible for 40 percent of the total effect observed as a result of the advertising.
And:
In 2014, Google released a report suggesting that 56.1 percent of all ads displayed on the internet are never seen by a human. One 2017 report by Comscore found that this problem is particularly pronounced for as purchased through the programmatic ecosystem.
...
One study by Deloitte from 2017 suggests that fully three-quarters of North Americans engage in "at least one form of regular ad blocking". In 2016, 615 million around the world were actively blocking ads.
But the main part of the book is a detailed comparison of the online advertising market now to the market in subprime mortgage bonds leading up to the 2008 financial crisis. Hwang bases his argument on the idea that, just as mortgages were treated as commodities:
What is different about the present-day online advertising system is that it has enabled the bundling of a multitude of tiny moments of attention into discrete, luquid assets that can then be bought and sold frictionlessly in a global marketplace. Attention is commodified to an extent that it has not been in the past.
By being commodified, the differences between individual mortgages were obscured, and so are the differences between different episodes of attention. The market is opaque:
Opacity isn't dangerous only because it can cause errors in valuation. It also allows for the active inflation of a market despite the fundamental shakiness of the thing being bought and sold. This sometimes results from irrational levels of market confidence, a regular feature of financial crises going back hundreds of years.
...
This divergence between rosy outlooks and structural vulnerabilities is kindling for crises of confidence. When a hot, overpriced commodity is discovered to be effectively worthless, panic can set in, causing the market to implode.
Hwang lists a number of causes for market opacity:
The measurability of the online ad economy is an inch deep and a mile wide. As such, the tidal wave of data that has accompanied the development of online advertising provides only an illusion of greater transparency.
...
Modern online advertising remains deeply opaque on three fronts. First is the ever-iincreasing automation of the marketplace. Second is the creation of dark pools of liquidity where advertising inventory is bought and sold outside of the public eye. Third is the dominance of platforms, like Facebook and Google, that have frequently introduced new layers of opacity into the advertising marketplace.
Here is my take on these three:
  1. Among the problems the scale and the automation cause is the difficulty of ensuring "brand safety". Advertisers don't want their ads appearing on controversial content, so they have an analogous problem to the platform's "content moderation" problem. What content is "controversial" enough to warrant exclusion? The real-time ad auction system makes it almost impossible for an advertiser to know what content their ad appears on, and even if they knew Masnick's Impossibility Theorem means there are bound to be a lot of cases where, in hindsight, the placement was a mistake.
  2. I discussed dark pools in my last post, The Order Flow, and Hwang's analogy with the financial markets holds here. Because both buyers and sellers see advantages from avoiding the public markets, and because the owner of the dark pool can profit by abusing their trust in the dark, dark pools were bound to arise in the online ad markets:
    Platforms increasingly give select buyers access to private marketplaces (PMPs)—exclusive exchanges for ad inventory. PMPs allow selected advertisers who have negotiated a special deal with a publisher to bid for advertising space, usually of a higher quality and in a less crowded, and therefore less competitive, market. These arrangements are attractive because they offer better transparency to the participants and also allow buyers to keep targeting data and other valuable information away from the public markets.

    PMPs are a growing segment of the transactions taking place in online advertising. In 2018, 45 percent of all the money spent in real-time bidding auctions took place within the confines of a PMP.
  3. Google and Facebook have much better information about the ad market than the publishers and the advertisers, and they are devoted to keeping both sides in the dark as much as possible. Facebook, in particular, has a long history of lying to both sides. Tim Peterson made a list of 12 such lies in 2017 and Amy Gesenhues added another 10 in 2018.
As a description of the market, Hwang's analogy holds up very well. Clearly it is possible that the market will suddenly collapse, as the subprime mortgage market did, when participants realize that wht they ar ebuying is worthless. If it did the impact would be huge. But his suggestions about how to let the air out of the bubble without disaster strike me as futile:
Two pillars of faith give programmatic advertising an aura of invulnerability: measurability and effectiveness. The core proposition of programmatic advertising is that it gives advertisers an unprecedented depth of accurate data about consumers, which is able to produce uniquely effective outcomes for advertisers.
...
reducing confidence in the measurability and effectiveness of programmatice advertising will chip away at the willingness of ad buyers to pour money into the ecosystem.
...
Independent research may be a particularly powerful tool for shaping industry views of online advertising.
I think a collapse of confidence in the online ad market is unlikely in the medium term for two main reasons.

First, there is already a vast amount of research showing that the value of on-line ads is near zero, and advertisers know about it. Some of it comes from experiments run by major advertisers such as Proctor & Gamble, and major publishers such as the New York Times. Both sides just have very strong incentives to ignore it, as described from the advertisers side in entertaining detail by Jesse Frederik and Maurits Martijn in The new dot com bubble is here: it’s called online advertising. Here's the merest snippet:
It might sound crazy, but companies are not equipped to assess whether their ad spending actually makes money. It is in the best interest of a firm like eBay to know whether its campaigns are profitable, but not so for eBay’s marketing department.

Its own interest is in securing the largest possible budget, which is much easier if you can demonstrate that what you do actually works. Within the marketing department, TV, print and digital compete with each other to show who’s more important, a dynamic that hardly promotes honest reporting.
...
Marketers are often most successful at marketing their own marketing.
The marketeer's effective marketing works for everyone:
"Bad methodology makes everyone happy,” said David Reiley, who used to head Yahoo’s economics team and is now working for streaming service Pandora. "It will make the publisher happy. It will make the person who bought the media happy. It will make the boss of the person who bought the media happy. It will make the ad agency happy. Everybody can brag that they had a very successful campaign."
Nico Neumann agrees:
One experiment he presents shows that, under proper experimental conditions, the impact of an ad for auto insurance had a negative effect on sales, rather than the massively positive one suggested by popular statistical models used in the industry.

So why does Nico think these technologies are so popular in the online advertising space? Marketers, he says, "love machine-learning/AI campaigns because they look so great in ... analytics dashboards and attribution models." This cutting edge technology is favored—in other words—because it makes for great theater.
Publisher's incentive to devalue the product they are selling is obviously zero, especially since so many are struggling to survive decreasing income as more and more is swallowed by the platforms:
One study by The Guardian suggests that some 70 percent of the money spent by buyers is consumed by the ad tech platform, with the publisher retaining the remainder.
Second, as Michael Lewis explains in his must-read The Big Short; Inside the Doomsday Machine, key to the collapse of the subprime mortgage market was the ability of investors who spotted the bubble to buy credit default swaps on the packages of subprime mortgages. These swaps were the equivalent for the bond market of short-selling in the equity markets; a way to make money by betting on a fall in value. As far as I can see there's no way to short the attention market, so there's no-one incentivized to make advertisers and publishers skeptical of the value of online ads..

Anyone interested in this topic must read Maciej Cegłowski's wonderful post from 5 years ago What Happens Next Will Amaze You.

1 comment:

David. said...

It isn't just Stacy Abrams. In How the Navajo Nation helped Democrats win Arizona, Rachel Ramrez reports on the organizing effort in Arizona, to which I contributed. Note in particular these maps.