4

The Economic Geography of Cyberspace

With Bruce Rogers

Every cheapening of the means of communication, every new facility for the free interchange of ideas between distant places alters the action of the forces which tend to localize industries.

—Alfred Marshall, Principles of Economics, 1895

In the late 1970s, a young economist by the name of Paul Krugman began publishing a series of articles about international trade. Krugman’s work was motivated by a puzzle. For more than a century-and-a-half, economists had explained trade largely through theories of comparative advantage. David Ricardo (1817) explained this idea through a parable about cloth and wine. Since the English climate is terrible for growing grapes, Ricardo argued, Portugal should end up with the vineyards and Britain should end up with the looms.

This was true, Ricardo said, even if Portugal was better at both weaving and winemaking: if relative advantage won out, trade would make both countries better off. Ricardo’s logic was compelling, as anyone who has actually tasted English wine can tell you. And with some modifications, it was the main framework for understanding international trade for a century-and-a-half.¹

But in the post–World War II era, it became increasingly clear that Ricardo’s theories explained less than they used to. As trade barriers fell, the bulk of international trade turned out to be between countries with similar economies—similar climates, similar natural resources, similar levels of industrialization. Even more puzzling, these countries were often engaged in intraindustry trade, or even trading exactly the same type of goods. This clearly could not be explained by comparative advantage. What was going on?

Krugman’s answer was a model with three central assumptions. First, the model assumed that firms had economies of scale—that larger firms were more efficient. Second, it was assumed that consumers had diverse preferences—that some buyers would prefer to drive a Volkswagen even if they could afford a Cadillac. Lastly, the model assumed that there were shipping costs.

With these key assumptions, Krugman was able to produce patterns of trade that looked much like those in the real world.² The model also produced other, initially surprising results that mirrored real world patterns. For example, it predicted that countries with the largest domestic markets would end up being net exporters—something that turned out to be true.

This chapter is about building simple economic models of online content production. The style of these models owes much to the so-called “increasing returns revolution” that began in the industrial organization literature,³ and then found expression in the “new trade theory” and in the “new economic geography” scholarship that Krugman’s work exemplified. This parallel is no accident. If economic geography studies the production and consumption of goods in space, much of this book is about understanding where content will be produced and consumed in cyberspace. There is a clear analogy between the industrial agglomeration we see in the real world, and the traffic concentration we see online.

The central building blocks of our simple web traffic model are threefold:

•  First, we assume that there are strong economies of scale online, both for content production and advertising revenue.

•  Second, we assume that users have at least modest preferences for diversity.

•  Third, we assume that it takes time and effort for users to seek out online content—that users face search costs or switching costs in navigating the web.

By this point in the book none of these assumptions should be controversial. But combined, they add up to a vision of the web that is starkly different from the one that continues to shape public policy: in this model, portal sites emerge spontaneously and have substantial market power, most content is produced and/or hosted at the largest sites, and niche dominance is the norm rather than the exception. It is an internet, in short, in which competition is deeply unequal.

Economic models have recently become a highly public front in the war over net neutrality and internet regulation. In one of his first high-profile speeches after being confirmed as FCC chairman, Ajit Pai denounced the supposed lack of economic analysis behind the FCC’s 2016 pro-net neutrality Open Internet Order.⁴ The notion that the Obama-era FCC ignored economic analysis, in the Open Internet Order or elsewhere, is questionable at best.⁵ But if economic modeling is to be the currency in which policy is argued, we should at least start with the type of increasing returns models that best capture the logic of the digital economy.

The statistician George Box’s famous aphorism—“all models are wrong, but some are useful”⁶—certainly applies to the models in this chapter. Economists often talk about these types of models as “fixed” or “static” models, but they are better thought of as sped-up or time-compressed dynamic models, in which things that happen fast are presumed to be nearly instant. As we will see, this acceleration helps show the often strange incentives of online content production. Understanding the incentives of both users and producers, in turn, helps explain otherwise baffling features of digital life.

Importantly, the results here are less sensitive to key assumptions than some past writings. Classic models from the broadcast era depended on strong assumptions about consumers’ media preferences, and much writing about the internet has assumed that diverse users’ preferences are the primary force fragmenting digital audiences.⁷ Our models here, though, produce similarly concentrated results for either strong or weak preferences. Weak preferences give no reason for users to spread out, but strong preferences just end up sending users to portals, aggregators, and search engines. If digital media diversity requires Goldilocks preferences—not too strong, but not too weak—then it is unlikely to be stable.

Yet before we delve into model building, several topics touched on earlier require a fuller treatment. We will start by discussing online aggregation and the economics of bundling. We will then lay out what we know about media preferences, and why larger sites now get more money per visitor than small local producers.

BUNDLING AND AGGREGATION

While chapter 2 and chapter 3 talked at length about the advantages that large sites possess over small sites, one crucial advantage has been deferred until now. Large sites have an edge, in part, because they construct large bundles of content. All of the web’s most successful sites are aggregators—bundlers—of one form or another.

Many types of firms offer their products in packages, a practice known as bundling.⁸ Early economic research on bundling suggested that the producers were simply taking advantage of economies of scale in production or distribution. As we saw earlier, these sorts of scale economies are common. Still, Adams and Yellen’s⁹ pioneering work showed that bundling made sense even without these sorts of scale efficiencies.

To see how, consider the case of Microsoft Office, one of the world’s best-selling bundles of software (table 4.1).¹⁰ Alice, Bill, and Chris all have different prices they are willing to pay for different parts of Office. Alice will pay $100 for Word, $45 for Excel, and nothing for Powerpoint. Bill will pay a $100 for a spreadsheet, $45 for a presentation program, and nothing for a word processor. Chris will pay $100 for PowerPoint, $45 for Word, and nothing for Excel.

In this simplified example, Microsoft can make more by selling Word, Excel, and PowerPoint as a package. Instead of selling one copy each of Word, Excel, and Powerpoint for $300 total, Microsoft sells three bundles for $435.

This model has been extended to media markets too. As Jay Hamilton’s research shows, the same logic applies to media goods.¹¹ Two subscribers of the Wall Street Journal may read the paper for very different reasons, but newspapers can make more subscription revenue by bundling these different types of news together. Media subscriptions capture not just variation in interest across media sections or stories, but also variation across time. Few people read the paper closely every single day. Some have argued that newspapers were the “original aggregators,”¹² combining disparate varieties of content into a single product. What’s true in print is also true for goods like cable television, which is almost always bought as a bundle or series of bundles.

Bundling works, in part, because it serves as a form of price discrimination. Companies would love to charge people different prices for the same exact product—ideally, as much as each consumer is willing and able to pay. Bundling takes advantage of the law of averages. It is easier to guess what a consumer is willing to pay for the package deal—for all of Microsoft Office, or all of the Wall Street Journal—than it is to guess what she might pay for the individual parts. Bundling thus provides “predictive value” for firms figuring out what price to charge.¹³ Willingness to pay for the bundle varies less than willingness to pay for the component parts.

Bundling is an especially powerful strategy with information goods, such as media or software. With real goods and services, extra items in a package deal cost the producer more. A two-for-one deal on laundry detergent means that Tide pays more in production, shipping, and packaging. But with information goods, the marginal cost approaches zero.¹⁴ Once the movie is released, or the software is finished, it costs little extra to provide somebody with a digital copy. If the software code for a website is well designed, it can cost almost nothing to serve an additional user.

Bundling has other advantages, too. Barry Nalebuff¹⁵ suggests that bundling discourages competition. Bundling forces an upstart firm to enter several markets at once—a much tougher proposition than just entering one. This is especially true in markets that have strong economies of scale: new competitors have to get big to be efficient, and getting big enough to compete in many markets simultaneously is a daunting challenge.

So far, so good. Still, research on bundling has focused on the amount of money firms can extract from consumers. But if bundling helps explain how much money firms can extract from consumers, a similar logic also helps explain their ability to extract time. Taking seriously the idea that attention is a scarce resource has important consequences in this case.

Most content consumed online is not paid for directly by those who view it, challenging the traditional explanation of bundling. For websites, bundling is a winning strategy because it maximizes the number of users who spend at least a little bit of time on their site. A larger user base leads directly to higher per-user revenue, thanks to better ad targeting (as we saw in chapter 3) and reduced audience overlap (more on this momentarily). More users also leads to lower average costs, thanks to economies of scale.

From an individual’s perspective, too, seeking out bundled content makes sense even when she does not pay for what she reads. Bundling provides predictive value for a site visitor spending his or her limited time budget. For an individual visiting an aggregator site, bundling reduces the risk that the trip has been in vain. If there is nothing interesting to read in the entertainment news, then perhaps sports or business news or the style section will have something worth reading. Sites that can provide users with their information needs, without requiring them to spend time and effort searching other sites, keep users engaged for longer. Recommendation systems (chapter 3) compound this advantage.

THE NEW ECONOMICS OF ADVERTISING

Understanding the basic logic of bundling is crucial to understanding the political economy of the web. But there is another key building block that needs to be acknowledged: bigger sites get more ad revenue per user than smaller sites. Despite efforts to push digital subscriptions (see chapter 7), digital advertising continues to dominate other online revenue streams. A site that doubles its readership will more than double its ad revenue—a game-changing break with traditional media.

The fact that local per-person advertising rates were higher than national per-person rates was arguably the defining feature of twentieth-century media. National broadcast advertising was expensive, but cheap per viewer reached. For example, General Motors’ success has been attributed partly to efficiencies in national advertising.¹⁶ This is the reason we see ads for national brands when watching the Super Bowl, but not ads for local restaurants. As Yoram Peles argued in the broadcast era, “It appears that there is an indivisibility in the advertising media, so that it is worthwhile only for big companies to advertise through some mass media.”¹⁷ By the same token, neighborhood newspapers traditionally had higher ad rates than metro papers, and local TV stations higher rates than network broadcasters. The expectation that this will remain true on the web inspired much of the hype around hyperlocal media (more on hyperlocal in media chapter 6).

Online advertising turns a century of advertising economics on its head. As we saw earlier, online revenue is hyper-concentrated on the largest players, with the Google-Facebook duopoly alone controlling more than 70 percent (and counting) of U.S. digital ad revenue.¹⁸ There are several reasons that the internet has reversed the revenue curve, some of which we discussed in chapters 2 and 3. Much of the local-advertising advantage came from the power of local newspaper monopolies. As Philip Meyer writes, “In their efforts to find one another, advertisers and their customers tend to gravitate toward the dominant medium in a market. One meeting place is enough. Neither wants to waste the time or the money exploring multiple information sources.”¹⁹ But it is now large web firms—not local newspapers—who enjoy the premium from reduced duplication of audience. Local newspapers no longer have the widest reach even in their own market.

In addition to lower audience duplication, big sites are also far better at measuring how well digital ads work. Even multimillion-dollar digital ad campaigns, with tens or hundreds of millions of impressions, are often unable to make reliable inferences about their own effectiveness.²⁰ Google has developed powerful statistical methods to help address this problem, but these techniques need data on how similar campaigns performed—information only Google has access to.²¹ And of course, as we saw in the last chapter, digital giants are better at ad targeting. The largest targeted ad campaigns, on the largest digital platforms, produce far more sales than simple untargeted ads.²²

The ultimate result is an enormous disparity in ad pricing. Advertising prices are often measured in CPM, or the cost per thousand impressions (with M being the Latin numeral for thousand). Hard numbers on ad prices are difficult to come by—a fact that by itself says much about the power of the largest digital platforms, which usually succeed in keeping most ad sales data private. But as of 2012, Google’s CPM for banner advertising was estimated at $95.²³ For newspaper sites, by contrast, CPM was only $12 to $15 for advertising sold directly. While this sounds discouraging for newspaper sites, the reality was worse: few such sites came close to selling the majority of their inventory. The balance of unsold advertising space was typically sold as “remainder” or “remnant” advertising on different ad networks, most often at a price between $1 to $2 per thousand impressions. A surprising number of smaller newspaper sites still offer only “run of site” banner advertising, in which the ad appears on all pages, or a random selection thereof. Further more, mobile or video ads are often crude or nonexistent on many smaller newspapers’ sites. The per user advantage for large sites is even greater than the per impression numbers above, because users tend to stay longer on big sites.

Both real-world data and theoretical work, then, show that large sites earn more per user, an assumption we will build into the models to follow. To be conservative, though, we will remain agnostic about exactly how large that advantage is—only that large sites earn at least a bit more per user.

WHAT We KNOW ABOUT PREFERENCES

The last, crucial building block for our models has to do with the character of media preferences.

Many continue to assume that media preferences are inherently diverse—and thus that the internet will create radical diffusion of audience attention. Assumptions about diverse preferences lead to the “Waiting for Godot” belief that audiences will eventually democratize (see chapter 1), or to Negroponte’s assertion that the internet would dissolve the mass media (see chapter 2), or to Andrew Sullivan’s claim that the internet allows for “the universe of permissible opinions to expand, unconstrained by the prejudices, tastes or interests of the old-media elite.”²⁴ If the economics of broadcasting produced bland, homogenized content,²⁵ the internet is supposed to produce the opposite. Chris Anderson’s writings on the “long tail” similarly assume that consumers have strong, diverse content preferences that will finally be satisfied in the digital era.²⁶

Yet the empirical case for diverse media preferences is surprisingly modest. Early work on the economics of media often assumed that people had strong genre preferences for radio or television shows.²⁷ Yet empirical work soon found that genre preferences were rather weak.²⁸ As one influential British study concluded, it is not true that “programs with similar content appeal to some particular subgroup of viewers.”²⁹ Viewers’ likes are not strongly influenced by genre—though their dislikes are often grouped.³⁰

Some might see the lack of strong genre preferences as an artifact of the broadcast era. Much the same pattern, though, emerges in the Netflix Prize data (see chapter 3). Teams expected most improvement to come from better matching users’ preferences with the characteristics of each movie, but these effects turned out to be remarkably modest. Some movies get higher average scores from nearly all users, and some Netflix users are more generous with awarding stars. These general biases that work across all users, or all of an individual’s ratings, accounted for three times as much variance in ratings as user-movie interaction effects.³¹

Still, there is good evidence that media preferences do affect politics more than in the recent past. As Markus Prior shows in Post-Broadcast Democracy, roughly 40 percent of the U.S. public prefers entertainment over news.³² These preferences mattered little in the broadcast era, when there was little choice of programming. The “switchers” who abandoned news for entertainment fare in the post-broadcast era show significant drops in political knowledge and their likelihood of voting.

Similar findings emerge in news readership data from the web, where preferences increasingly drive news consumption (or lack thereof). Pablo Boczkowski and Eugenia Mitchelstein document a “news gap” between the news stories that editors highlight as important, and stories that actually receive the most attention.³³ The size of this news gap is larger on some news sites than others, though, and it shrinks in election season.

In recent years the ideological preferences of users have been a major topic of research. Many scholars have found that a sizable minority of citizens prefer like-minded news content.³⁴ Though only a fraction of citizens prefer partisan news, those who do are among the most involved and partisan citizens—thus amplifying the impact of partisan media sorting.³⁵

Much remains uncertain about the character of media preferences. We still know far too little about their true origin. Some recent work, too, has begun to explore the extent to which preferences can be altered or cultivated.³⁶

In general, though, a few broad points are clear. Some citizens, some of the time, for some types of content, do differ significantly in their media preferences. There is little evidence for the radically divergent media preferences that underlie many claims about the democratizing power of digital media. Still, user preferences are clearly more important in today’s media environment than they were a generation ago. And continuing uncertainty means that we should show how different assumptions about media preferences change—or do not change—the conclusions of our models.

A SIMPLE MODEL OF DIGITAL CONTENT PRODUCTION

With these building blocks in place, it is time to start constructing our formal model. The full model, including mathematical detail, can be found in the Appendix. But the outline of the model, and its most important conclusions, can be understood without any math.

We will build the simplest model first, and then make the model progressively more realistic (and complicated) by relaxing some initial assumptions. As a general rule, though, increasing realism strengthens the position of the largest sites.

We will start with some number of websites, and some number of consumers. The consumers get utility—which is to say, enjoyment or satisfaction—from consuming site content.

Each site produces its own distinct variety of content. Our consumers have preferences for some varieties of content over others. For example, they might prefer serious versus escapist content, or perhaps liberal-leaning news instead of conservative-leaning news.

Users get more utility from consuming content closer to their ideal preferences. Conservative users would thus enjoy watching Fox News more than CNN, but they enjoy CNN more than liberal-leaning MSNBC. For simplicity we will consider content variety to be one-dimensional, but the model works with multidimensional preferences, too. The distance from the ideal could be measured in two or more dimensions: a liberal who prefers entertaining content might consider the Daily Show nearly perfect, but find Fox News doubly far from her ideal because it is both conservative and serious.

This content has to come from somewhere, so each site hires writers. The more writers they hire, the faster the content is produced: a site with twenty writers has a production rate twice as fast as a site with ten writers. Similarly, sites can pay different amounts to their writers. Writers paid higher wages create higher-quality content.

How much utility users can receive from each site is thus a function of three things: the quality of the site, the quantity of content the site produces, and how close the site’s specific content variety is to one’s ideal.

Consumers each face a budget constraint, which is the amount of time they have to enjoy the web. We will start by assuming that all users have the same time budget. Additionally, there is a fixed cost to the user’s attention—which we will call a switching cost—each time she navigates to a different site.

Switching costs are an important part of what makes this model work, so it is worth saying more about what we mean by them. The key is that it costs consumers to switch from one site to another. This might be search costs as traditionally understood in economics, the time and effort to find another interesting site (see chapter 3). Alternatively, we might conceive of these costs through the lens of cognitive psychology, which has found that it takes effort for people to switch tasks.³⁷ The “don’t make me think” school of web design,³⁸ or Jeff Bezos’s focus on reducing “cognitive overhead” in digital media,³⁹ similarly suggest that decision-making is costly for users. Any of these explanations are consistent with our model.

With perfect information, consuming digital content resembles the task of deciding which stores to shop at. An all-knowing consumer could visit numerous boutiques to purchase exactly the variety and quality of goods he wants, or instead save time and gas by going to a department store. A single trip to Walmart is more convenient, but requires compromises on quality, quantity, or variety.

If consumers are trying to maximize their enjoyment, our model websites want to maximize profit: revenue minus costs. In keeping with our findings in previous chapters, we will assume that revenue is an increasing function of the total content consumed at a site. Sites’ production costs have two parts: a fixed cost and labor costs, which are the number of workers times the wage rate. As we assumed earlier, content quantity is proportional to the number of employees, and quality is proportional to the wage rate.

The combination of quality, quantity, and variety determines how much of a site’s content users are willing to consume. As with the assumptions we have made thus far, users have a clear preference ordering for all websites. Rational users will consume sites in order, from their most to least favorite, paying the search cost at each transition until their time budget is exhausted.

That is the skeleton of our model. But though the assumptions that go into the model are straightforward, they already produce some surprising results. Let us consider some examples.

Example: Two Sites and One Consumer

For our first example, consider the simple case where there is just one consumer and two possible websites, which we will call Site A and Site B. Site A is closer to the consumer’s preferences. In fact, let us assume that Site A produces exactly the variety that our sole consumer most prefers.

The model shows that even perfect preference matching might not matter. If Site B produces high-quality content fast enough, it can earn all of the consumer’s limited attention. Sites that can make sufficient investments in quantity and quality can draw consumers past sites that are closer to their ideal preferences.

This itself is an important result. As we have already seen, debates about the future of media have often claimed that small online outlets will beat big media companies by producing niche varieties of content. Our model suggests, instead, that even modest search costs or quality/quantity advantages can make competition lopsided.

Our simple model also reproduces a central tenet of internet wisdom: the key role of fresh content. The importance of new, previously unseen content every time a user visits is a central component of site stickiness (more on this in chapter 7). Sites cannot be a destination unless they produce a steady stream of new stories. Even in this simple model the rate at which new content is generated is a key predictor of success.

Our model tells us, too, about the pressures on content quality. Somewhat surprisingly, it suggests that most websites have a strong incentive to drive down their content quality. Economies of scale are attached to volume, not quality. High-quality content costs more without necessarily producing any additional revenue.

Websites thus maximize profit when consumers read a lot of mediocre, inexpensive content. Quality needs to be high enough to prevent competitors from emerging—but no higher. Provided that readers have large time budgets, websites will seek to hire lots of cheap, low-wage writers. In short, the model has already produced the logic of the content mill.

Example: Two Sites and Many Individuals

For our second example, consider the situation where there are two websites and many consumers. Suppose that consumers’ preferences are spread evenly across the spectrum of possible content.

A couple of consequences are quickly clear. First, the strength of preferences becomes critical. Consider what happens if preferences are weak. Any user will still consume any variety of content, provided the quantity and quality are high enough. In the case of weak preferences, then, both sites will attempt to locate in the center of the preference space. This is a very old economics finding across many different kinds of markets, dating back to Harold Hotelling’s work nearly ninety years ago.⁴⁰ Locating in the center means that sites have minimized the distance to the average user’s ideal point.

Even more importantly, if revenue grows faster than production costs, more content production will mean larger profits (or at least smaller losses). Sites will strive to get bigger and bigger: hiring more writers always produces a greater payoff. If writers are all paid the same wage, each additional writer for a site provides more profit than the last.

Provided there are no upper limits on production, the highest possible revenue occurs when all users spend their entire attention budgets on a single site. If this level of production is profitable, then there is a stable monopoly: all consumers will spend their entire budgets on one site. This is the only site consumed; no other sites will be profitable. Our consumers even prefer this arrangement, because they do not have to waste their time budget changing sites.

Example: Strong Preferences for Variety

For our third example, let us see what happens when consumers’ preferences for variety become stronger. In the previous example, one dominant site could use its size and quality advantages to monopolize the audience.

But suppose that some individuals get zero utility from content varieties far from their own preferences. Consider a conservative news consumer who ideally prefers Fox News. Our conservative consumer might be coaxed to visit ABC News if quantity and quality are high enough. He will never visit MSNBC or DailyKos, though, because he gets no enjoyment at all from their content.

Our model can reproduce this result by giving consumers a limited preference window. While there is still a bonus for content closer to their own preferred variety, consumers will examine content only within a given distance of their own ideal. To keep things simple, we’ll also assume consumers’ preference windows are all of the same width.

As long as all users’ preference windows include the midpoint of the variety spectrum, a monopoly will still emerge. But as preference windows narrow further, it becomes impossible to capture everyone’s full attention. The narrower these windows become, the more opportunities there are for additional sites to enter the market—further fragmenting users’ attention.

In the real world, the extent to which consumers have strong variety preferences varies across different content categories. As we saw earlier, many scholars have found that a sizable minority of political news consumers prefer news from like-minded sources. But in this regard political news is unusual. Variety preferences are weaker in other news domains, such as weather or sports or entertainment. Entertainment content preferences show more overlap than disagreement. Democrats and Republicans alike watch their sports on ESPN. Many digital content categories show little or no evidence for diverse preferences.

Extending the Model: Bundling and Aggregation

We can also consider preferences for variety in a different way as well. Thus far we have limited sites to producing a single category of content. But what happens if some sites can produce more than one category of content? As discussed earlier, most traditional media are bundles of disparate content types: news, entertainment, sports, comics, cat videos, and so on. Bundling complicates the model, requiring a couple of additional assumptions to find an equilibrium, but in general it strengthens the hand of the largest sites.

Consider a simple case of bundling with three different sites:

•  A site that produces only news.

•  A site that produces only entertainment content.

•  A portal site that produces both news and entertainment.

For this extension, we will make preferences even simpler: users either get perfect preference matching from all content in a category, or they get no enjoyment at all from the category. We will assume, too, that users get slightly less utility from a content category as they consume more. The first entertainment article consumed provides more enjoyment than the tenth, and the tenth article provides more enjoyment than the hundredth.

As before, assumptions about the distribution of audience preferences are crucial. For starters, we will assume that a third of the audience likes only news, another third likes only entertainment, and the final third likes both news and entertainment.

Despite the even spread of preferences, the portal site captures all of the audience. With economies of scale, the portal site can produce more content and/or higher-quality content in both categories at once. Even the news-only and entertainment-only consumers end up at the portal site, though they get zero utility from half of the content.

Surprisingly, though, portal sites can dominate even if the overlap between user preferences is small. We can vary the portion of users who like both news and entertainment. We can even construct extreme cases in which a tiny amount of overlap allows portal sites to win out. Even if 49 percent of users like only news, 49 percent like only entertainment, and 2 percent like both, the portal site will still attract all of the audience. To be sure, the advantages in this case are smaller compared with more evenly spread preferences—and likely quite fragile since they are so close to the tipping point. Yet as this limiting case shows, even low levels of preference overlap can be an advantage for portal sites.

Overall, then, the models thus far suggest grim conclusions for smaller digital publishers. How solid are these findings? One measure of robustness is how much we need to alter the model for small sites to survive.

Varying the amount of time users have to surf the web is one change that might improve the fate of small publishers. Some users would be given little time online, others middling amounts of time, and still others an almost unlimited time budget. This assumptions of varied time budgets is broadly consistent with real-world web traffic, in which users with less time—such as those surfing at work—spend a greater proportion of their time on big sites. Big events such as elections similarly increase the market share of big news sites more than those of smaller niche outlets, such as political blogs.⁴¹

In this case, then, time-poor users with varied preferences will visit only portal sites. By contrast, users with near-boundless time will include niche sites in the tail end of their browsing.

Adding this complication can provide a narrow path for niche sites to survive—but it makes it hard for them to thrive. Even with favorable assumptions, these niche sites still have fewer users and lower potential profits than large portals.

Extending the Model: National, Local, and Hyperlocal

Bundling provides discouraging predictions about the economics of small-scale digital content. Even worse: local news is just a special case of bundling. The model predicts local publishers will struggle to survive, while national publishers hoard most of the attention and profits.

Consider two digital newspapers, in two different cities, which produce both local and national content. National content appeals to readers across the country, while local news appeals only to those in the same city. Users vary in their relative preference for local vs. national news, and in the amount of time they have to spend online.

Importantly, one of our cities—let’s call it Gotham—is larger than the other city, which we will call Midway. The Gotham Gazette is thus larger than Midway Meteor.

Start by making switching costs high—so high, in fact, that there is no cross readership at all between the two newspapers. High switching costs essentially return us to the print era, when local newspapers enjoyed regional monopolies. With no competition, both newspapers produce both local and national news. But because the Gotham Gazette is bigger, it earns more per reader, and produces more and higher-quality national news.

Then consider what happens as switching costs fall: Midway Meteor readers start switching to the Gotham Gazette for their national news. Readers with the highest time budgets, or those who have a strong preference for national vs. local news, make the switch first. If switching costs drop low enough, the Gotham Gazette will cannibalize most of the Meteor’s national news readers.

Real news markets over the past decade have mirrored the results of our model. In 2004 only about 12 percent of U.S. reporters lived in New York, Washington, D.C., or Los Angeles. By 2014, though, that proportion had jumped to about 20 percent and counting.⁴² More than 40 percent of digital journalism job listings are now located in the New York or D.C. metro areas.⁴³ Far from announcing the “death of distance,”⁴⁴ digital media has made news production more geographically concentrated.

The same forces that have diminished metro news apply even more strongly at the neighborhood and hyperlocal level. To see how, add a third level of news to the model above: national, local, and now hyperlocal. Imagine a hyperlocal news site in Midway, the Old Town Observer. If the model holds, this hyperlocal site will earn less per reader than the larger Meteor. As we will see in later chapters, that prediction is consistent with the chronic struggles of hyperlocal news.

Extending the Model: Search Engines

The model could be extended in another important way as well. It can accommodate a new, special kind of website: a search engine.

A search engine reduces switching costs. When surfing from one site to another, users can visit the search engine and pay only part of the normal switching penalty. In this context any site that helps users find new content is considered a search engine. Google and Bing are search engines, of course—but by this definition so too are platforms like Facebook or Twitter or Apple News. Crucially, search engines—just like other websites—are assumed to have economies of scale in revenue.

At first glance, search engines seem to expand users’ horizons. With the same time budget, users will range further afield in their search for content. Diversity of consumption increases as switching costs go down, with smaller sites in particular benefiting from this shift. This mirrors patterns seen in real-world data. Small sites get a far larger fraction of their audience from search engines than large sites do.⁴⁵ Similarly, when Google shut down the Spanish version of Google News in 2015, traffic to smaller news outlets fell precipitously, while the audience at the largest outlets was not significantly affected.⁴⁶

Yet search engine benefits to smaller outlets are not the whole story. In economic geography, a common pattern is that transportation hubs can end up becoming manufacturing centers. If the goods have to pass through the shipping hub to reach the consumer, it’s often cheapest to produce them in the transport hub in the first place. (We’ll see an example of this in the next chapter).

The model predicts a similar result online. All else being equal, the best place to host content is wherever the largest audience already is. Successful search engines thus face strong incentives to produce—or at least host—content themselves. Most search engines have taken this path, starting with the early efforts of Yahoo!, AOL, and MSNBC. Even Google, which initially stood out for resisting this trend, is now the largest host of digital content in the world through its ownership of YouTube.

The model helps explain the push by Facebook, Google, and Apple to directly host news content created by others. The spring 2015 introduction of Facebook’s Instant Articles was an important shift in the media landscape. Instead of Facebook just linking to articles on other sites, the full text of Instant Articles would appear directly in users’ news feeds. Facebook promised news partners more traffic, as well as a 30 percent cut of ads that Facebook sold against publishers’ articles. Google’s Accelerated Mobile Platform (AMP) and Apple News are also attempts by digital giants to directly host others’ content. By shifting content directly onto Google’s servers, news providers get faster load times and a privileged position in search results. Apple’s revamped News app, introduced in September 2015, similarly moved content from news partners to Apple servers.

Search engines, then, do not provide a solution to media concentration—they just re-create the original problem in a new form. Successful search engines face strong incentives to directly host content. Increasingly, all of the digital giants are acting to do just that.

INCREASING RETURNS IN THE DIGITAL ECONOMY

In 2008, in his Nobel prize lecture, Paul Krugman recounted the intellectual journey that had led him to the podium in Stockholm. Despite the influence of his work, Krugman acknowledged that his research had been partly superseded by events. Growth in international trade since 1990—such as growing trade between China and the West—was a return to the old comparative-advantage pattern. The economic advantages of industrial concentration, too, seemed to have waned. The U.S. auto industry, once clustered in a single region, has spread out to include most of the American South, and even much of Mexico. The “increasing returns revolution” in trade and geography, Krugman admitted, now described “forces that are waning rather than gathering strength.”⁴⁷

From a broader perspective, though, the increasing returns revolution is more important than ever. The internet economy now accounts for more than 5 percent of the total economy in developed countries.⁴⁸ The multitrillion-dollar digital economy is still growing at roughly 8 percent a year, far faster than the rest of the economy. The increasing returns models developed for economic geography help explain many of the apparent paradoxes of digital life. And these models show that widely held assumptions about the internet, taken together, add up to some surprising conclusions.

Our formal model provides some surprising insights. It suggests, first, that even perfect preference matching is not enough to ensure an audience. High quality, perfectly matched content can still end up ignored if other sites produce more content more quickly. This result follows irresistibly from strong online economies of scale, and it helps explain why a constant flow of fresh content is essential for audience building. (More on this in later chapters.)

Perhaps even more unexpectedly, the model shows that diverse content preferences can actually concentrate audiences rather than spreading them out. Many scholars and commentators still claim that strong, diverse preferences will eventually spread out digital audiences. Many still view the triumph of hyperlocal content as inevitable, despite failure after failure of real hyperlocal sites. Yet our models suggest that this thinking is backward. Portals sites and aggregators can get stronger, not weaker, when individuals prefer an extremely diverse media diet.

The model also casts light on the logic of content farms and aggregators. Websites maximize profit when they show a broad audience a mountain of cheap content. Low-quality content on a big site can earn more than the high-quality content on a small site. Journalists are regularly outraged when a big aggregator’s hasty rewrite of a story brings in more money than the original version. This model explains why this unfairness is both predictable and difficult to remedy.

The model also forces us to rethink the role of search engines, which prove to be a double-edged sword. On the one hand, they can push traffic out to smaller niche producers of content. At the same time, though, search engines and portal sites have powerful incentives to produce—or at least host—content of their own. The logic of industrial economics, in which shipping centers often become manufacturing centers, reappears in the digital realm.

The simple models of content production, then, are helpful in understanding the underlying logic of digital media production. Yet these models do have limits: by speeding up the time scale, this chapter has skipped over important details of how digital audiences evolve over time. The day-to-day dynamics of audience growth will be the subject of the next chapter.