 I really am especially honored to be here today. Jean-Jacques LeFont was really an inspiration for me, and he, reading his work and watching him present in the 1990s, showed me a great role model of how you can bring together a theory and empirical work. And at the time that I first saw his work on auctions, I was mostly a theorist. And so it was just kind of all inspiring for me to see how he could bring together different fields. And so even against the advice of most of my mentors, I tried to follow in his footsteps. So it really means a lot for me to be here today. So today I'm gonna talk about the internet and the news media. I got interested in this topic when I was working on the search engine at Microsoft, the Bing search engine, which I started working for in about 2007 when it was actually a pretty young search engine. One of the things that I recognized immediately at the time was how powerful technology platforms are in terms of determining the information that you get. I could see that the way that we ranked results, the way we presented them, changed what people read, what information they got, and where they bought things. And of course, as we all now have mobile phones with us everywhere we go, technology platforms have now gotten in between us and almost everything that we do. We are seeing, and of course for many reasons, which I won't talk about so much today, but those technology platforms tend to be very concentrated. One or two or three, maybe usually two leading platforms, say Android and iOS, are platforms that get in between the consumer and everything that they do. Those platforms then have amazing power for determining which credit cards we use, where we shop, which coffee shop we find when we look at our maps, as well as what information we receive. And so we're going to see that this power playing out in more and more arenas of our economy and of our lives and of our political lives as well. Of course, Toulouse was the birthplace of the study of platform markets. Another reason that it's inspiring for me to be here today to talk about platform markets. Now, I was very aware that these platforms would have a lot of power. I think I come today to talk about the internet in the news media, and of course, the big question in the room is, how did the internet elect Donald Trump? What did the internet do for Brexit? And I'll have some hints about that today, but I hope in research to come that will build directly on what I'll show you today, we will have maybe even more insight about how important that is. So just to start to think about, where do you get your news? Where do you get your information? If it comes through Facebook, if it comes through Twitter, the news that you read is determined by a very complicated set of forces. It's coming from how the search engines, aggregators, and social media rank the content that you have. Okay, so that's one fourth. There's gonna be a bunch of things your friends post, but you don't just see them in Facebook, you don't see them in the order that they're produced. Facebook decides what's important to you. It decides based on the content, the website, which of your friends liked it, and so on. And in fact, those algorithms that determine how things are ranked are changing all the time. In the last two weeks, there's been a debate about whether Facebook should demote news that's fake or false. But that's a choice that they make. And they can change those choices. Maybe if they changed them two months ago, we might change the outcome of an election. Then users are gonna choose from that, and of course, that's going to be determined by what they see. Of course, they can only see things that were created. So they might have been created by their friends, or they might be news articles that were produced by the news media. If the news media doesn't write the story, of course, you can't see it. Okay? And so we're going to see this very important feedback effect that if people start getting most of their news, say through aggregators or through Facebook, then the news outlets will produce stories that are likely to get ranked highly on social media or get shared by their friends. And so what we've already seen is that even for news organizations that don't explicitly reward their reporters for producing shared news, the reporters themselves care about their careers. And so the reporters study how to make their news shareable on Facebook. They study how to make headlines catchy. Some of my former students are working at Facebook now. They recently released a feature to demote clickbait. So they had ways of measuring. If you clicked on an article, looked at it and came straight back, they decided that was bad and they would demote it. Also, if they had other heuristics for determining whether the headline matched the story. But those are decisions that are made by the social media outlets. And of course, again, those can change over time. And these are profit maximizing firms. They care about the user, but they may also care about other considerations as well. The user's choices also will be influenced by their past experience. So we've learned that Google and Bing put the best things on top. We learned something about the quality of the advertisements. You mentioned in the introduction, the paper I wrote with Glenn Ellison, in that paper we modeled users rationally searching through advertisements and the consumers would learn that the ads were high quality and then search more often if the firms only displayed high quality ads. And the firms, the advertisements bids were a signal of quality. So that's also endogenous. You're gonna learn whether Facebook is showing you articles that are relevant to you or not and that's gonna determine how much you click on them, how much you trust what Facebook shows you versus how much you trust what your friends like and other things. So we have this big sort of equilibrium soup in terms of what we actually read and how informed a population is. But again, in a close election, especially when characterized by lots of misinformation or fake news stories, questionable facts, very extreme news outlets producing perhaps misleading information or propaganda which has become a real issue in the United States recently, a really growing issue, then how all of this soup gets mixed can have a real impact on what happens. So I think an ultimate research agenda would be able to understand this entire equilibrium. As you can imagine, it's very difficult to study all of those factors at once. So today I'm gonna pick across a few pieces of that. I'll talk a little bit about advertising markets because without money, the newspapers don't produce news. And of course, people are complaining, why don't the newspapers do more? But we forget that many of them are about to go out of business. They've fired half of their reporters and they're operating on a shoestring. So it's a little bit unfair of us to ask these news outlets to save our societies when they can barely pay the bills and keep the lights on. Then we also, so I'm also going to study the demand side or sort of the consumer choice. How changes in what the firms do change consumer choice among a set of alternatives. I think what is most interesting is how the content responds to that. How the newspapers write different stories. How different newspapers come and go from being in and out of business. I won't be able to answer that fully today because it takes time for Facebook or Google News to put all these companies out of business. They hang on for a while. But ultimately, if you understand how they affect consumer demand, we can at least have some predictions about what's going to happen to the supply of content. So one more theme I want to highlight just to put all of this in a bigger picture is that beyond, you say, the Googles and the Facebooks determining what you read, another trend for technology platforms is that they change the nature of competition in an industry. So think about Uber. Uber allows a student who has time between classes to become an entrepreneur, to monetize their time. So before Uber, it would be hard to make 30 euros between classes. Nobody would trust you if you tried to sell rides on the street. But Uber has now allowed small businesses, that is, people, students, to become businesses and trusted businesses. This two-sided platform enables a level of trust that allows a small business, an individual, to actually compete with an established taxi driver or an established limo driver. This is a big change in the nature of competition. Similarly, Airbnb allows someone with an extra room to become an entrepreneur, to become a small business, and sell their services, sell their unutilized resources. And the platform solves the trust problem, which then changes the nature of competition. Now hotels have new competitors. So there's actually an open theoretical question, I'll sort of throw out to the theorists in the room. Is it good or bad for welfare if you introduce a technology that allows more small firms to enter and compete with large firms? We usually think competition is always good, but of course, sometimes those large firms make investments that are important. They might innovate. They might provide higher quality. Maybe they provide more inspections. There are many things that a large firm might do. So I think that's why it's controversial to allow Airbnb to compete with hotels or to allow Uber to compete with existing firms. There could be a welfare trade-off in terms of allowing that competition to occur. So I'm gonna show you how that happens in the news industry, and I'm gonna show you how things like Google News and Facebook actually allow small news outlets to compete with big news outlets, whether that's good or bad, sort of depends on the level of investments. But the news industry is one where there's a lot of research and development. Every day you have to go out and create the news stories. And so we should be very attentive to the impact of changes in the competitive landscape from the purpose of creating news. Okay. So let me start with just some basic facts about search. And so this is a chart that I produced at the request of the European Commission. I produced this when the European Commission was investigating Google for antitrust violations. So this has been going on for years. I worked on it for a number of years on behalf of Microsoft. So they had uncovered a fair bit of evidence that Google manipulated search results. So Google put their own things above, say, European websites. And we found many, many examples where they did this. The question was, does this matter? And actually my lovely economist at the European Commission, especially a few years ago, they were very skeptical that this matters. Because we are all trained as economists to worry about search costs, but really, how big a search cost is there to make your eyeballs go up and down a page? So they said, I understand search costs of driving from one gas station to another, that I get. But I do not understand search costs of eyeballs going up and down a page. You can't really be serious that this matters. So I said, well, of course this matters. What do you think the entire billion dollars of research and development at Bing is trying to do? It's trying to rank results to get the best results at the top. If it didn't matter, you wouldn't need to do all that R&D. And for that matter, it would be no differentiation between Bing and Google. So in order to sort this out, at Bing we ran a randomized controlled experiment. So what I did is right before a user was going to see a page of search results, I took the thing from the first position and put it in the third position, took the thing from the third position and put it in the first position. In treatment group one, there's treatment group two where I switch the first and the fifth. And treatment group three, I switch the first and the tenth. So I took the thing at the top and put it all the way in the bottom of the page. And then there's a control group where you see what happens normally. So first of all, in the control group, these are the dark orange boxes. This is what the click-through rate is for the natural results. So in the control group in the first position, it gets clicked about 25% of the time. And by the way, these are non-navigational queries. So I have excluded queries where somebody types in the name of the firm that they want to go visit. Because it's actually pretty bad to manipulate in that case. We didn't want to make the Bing user so unhappy just to make the European Commission informed. And also we didn't think Google was actually doing that either. So we threw those out. In the third position, the thing that normally goes in the third position gets clicked 7% of the time, fifth position 4% of the time, 10th position about 2% of the time. But just as dark orange bars don't tell you about the causal effect of changing positions because the Bing engineers aren't stupid. They put good things on the top and worse things further down. So in order, and so you can think of that as a confounder, the quality of the link is unobserved. And so we can't tell correlation from causality just from the dark orange bars. The results from the randomized controlled experiment are illustrated in the entire orange area. So when I take the link from the first position and put it in the third position, it gets clicked 12% of the time. So they lose the dotted area. So they lose a little more than half their clicks. So you just take things from position one to position three, still visible on the webpage and you lose more than half your clicks. You can also tell the difference between correlation and causality here because the thing that usually is in the first position gets clicked 512th more than would be expected. So this 5% here is sort of the difference between correlation and causality. That is, this is the quality effect of the first link. But the main point I want you to take here, just from a randomized controlled experiment, is that you are really powerful when you run a search engine. And of course, every search engine knows this. Amazon knows this, eBay knows this, Google knows this. This is the whole point of what they do their R&D on is to develop algorithms for determining what order things go in. So we all knew this already, but it's important to realize the magnitudes. If I take something and put it down the screen, you can basically lose almost all your traffic. So you should think of these technology intermediaries as kind of puppeteers and you're the puppets. So then we should care a lot about these puppeteers and what impact they have. So now let me talk, I'm gonna start out by talking a little bit about some theory and this is actually done jointly with Emilio Calvano who was a PhD student to lose and then came to visit me at Harvard as a visiting student. And this paper tries to look at the effect of the internet on the ability of newspapers to make money, the advertising markets. So the paper was motivated by two facts. First of all, the internet makes it easier for consumers to switch among outlets. So when you go to Facebook, when you go to Google News, they're gonna show you results from lots and lots of different news outlets. That's very different from the old days when you would read a paper and maybe you'd read one or two papers. So you're reading a lot more different things when you use the internet to help you find information. And the second issue is that it's actually quite hard for advertisers to really know who is seeing their advertisements, especially when they advertise on different outlets. So if you teach a class on advertising, which I do, sort of one-on-one, the first one of the first things you talk to your students about is media planning. And that's basically what an advertiser does. They try to figure out how to buy advertising in different media, say different news websites, in order to reach a certain number of people. So in the old world, you can imagine if I bought advertising in every local newspaper, I could reach everybody who reads a newspaper. So it's sort of one advertisement for one user. On the other hand, if I advertise on the internet, I'm going to be seeing the same user multiple times. So I'm gonna have less than one user per advertisement. And actually one of the big themes of my class is that advertising measurement is completely broken and advertisers do a terrible job keeping track of people. So some former students of mine at Facebook did a study that showed that they mimicked an advertiser trying to find unique users and they had 10 different cookies per user. So the advertiser thought that this one person was 10 different people. And this is a common problem. Another way to illustrate this, I had ComScore run, pull some data for me. ComScore measures people's behavior on the internet by watching everything they do on their computer and mobile. They looked at 30 of their largest campaigns. So these are big advertisers trying to reach basically the whole country. And so what the advertisers wanna do, don't ask me why they wanna do this, but they believe that you would like, it's optimal to reach the consumers between three and eight times. If you reach them only once or twice, the consumers won't remember the advertisement. If you reach them more than 10 times, they've already seen the ad and each additional ad is wasted. What they found is that these, so these advertisers were trying using cookies to reach people between three and eight times and they spent a lot of money and used fancy technology to try to make sure they did that. But in fact, more than half of their ads were in the kind of wasted bucket. So they were too few or too many. Which just is another example that the best in class modern technology, cannot help the advertisers really know what they're getting. So that's kind of the motivation for the paper and we build a model that captures that. This is a model of two-sided markets of the type that was pioneered here in Toulouse by Jean Tereau and others. A two-sided market is one where something like a newspaper, you think newspapers are in the business of news, newspapers are actually a two-sided market that bring together users and advertisers. That's what they do. In the old world, a user might either read the Boston Globe or the Washington Post. So they're doing one or the other while an advertiser, a national advertiser, might advertise on both. And if they follow that strategy, they'll reach all the users in Boston and in Washington exactly once in a simplified model. And so the theory that came out of Toulouse is that this is a, in this type of world where the user goes to only one platform or the other and the sellers are willing to follow the advertisers to both, you have a competitive bottleneck. It's kind of amazing that we didn't understand this as economists until well after I got my PhD and the people in Toulouse wrote the first papers on this, but this is now kind of the story of many modern markets. There's a competitive bottleneck and you can charge monopoly prices to advertisers. So in fact, we had lots of antitrust cases 20 years ago about how the mean old newspapers were charging monopoly prices to advertisers. They shouldn't be allowed to merge because they were exercising so much market power in the advertising market and these newspapers were very profitable. So what happens in the new world? Well, in the new world, it's not that consumers always go to all newspapers. Consumers sometimes go to some and sometimes go to others. They're kind of unpredictable. So they're doing like a switching, but in an unpredictable way and they're certainly not all going to all outlets. So the question is, in that kind of a world, what do the advertisers do and what happens to advertising prices? So to understand this a little better, think about the problem an advertiser faces and this is kind of a stylized example of our model. Starbucks might advertise on outlet one in the morning and they might advertise on outlet two in the afternoon. So if they get a different consumer on in the morning and the afternoon, they're happy, but there might also be a consumer that goes to outlet two in the morning and outlet one in the afternoon and then they'll miss that consumer entirely. It might also be the case that the same consumer will see their ad twice once in the morning and once in the afternoon and because it's difficult to track the consumers across outlets from imperfect sharing of cookies and information, they're going to not be able to achieve their goal of reaching everybody once when the user is multi-helmed. So we build a formal model of this. Just for the theorists in the room, I put up what the model looks like. There's consumers, they have a fixed amount of attention. So one difference in our model is that if we think about multiple newspapers, it's not the case that each newspaper just adds to the user's attention, but rather there's a fixed amount of attention in the world and the newspaper is compete for it. The advertisers want to impress each consumer once, but some advertisers have a higher value consumer for consumer than others. So like say car companies versus somebody selling a small consumer product. And as I mentioned, there's a gap between what you buy which is an advertising impression, a space on the newspaper, and how many users you get. And so we look for a market equilibrium where there's a single market clearing price for advertising. And so what we find in this equilibrium is that the market is going to segment advertisers according to their value. That's not surprising, like high value advertisers will advertise because their values greater than the market price, low value advertisers will not. Okay, but what's different here is that there's also another decision the advertisers make which is how many newspapers to advertise on. And that's where we get kind of interesting results. So what we find is that we get what we call mixed homing. So high value advertisers like car companies will advertise everywhere. They'll get everybody but they'll get some people twice. So they're getting a small number of users per ad but they're high value so they're willing to buy it. On the other hand, low value advertisers can't afford to advertise everywhere because their value is not high enough if they're not getting a full person for every unit of ad. But if they only go to one outlet then they'll get one person per ad and so they can afford the advertising and so they'll buy it. And so what we see is that these low value advertisers are gonna miss out on some consumers in order to avoid wasting. And that's gonna lead to lower advertising prices which is one of the many reasons that advertising prices are low on the internet. Basically because it's inefficient and that inefficiency further causes the advertisers to sort of pull out of the market. There's also just a traditional effect that because they're competing for users an advertiser could find that user on one outlet or another. You're gonna have the standard Cournot types of effects where you're going to have this new competition you're no longer able to charge monopoly prices and prices fall for that reason as well. And so what are some of the implications? Well things like blogs that take up time without a lot of advertising can be good for the market because there's a scarce amount of attention and so that can raise advertising prices. But things like aggregators like Google News are gonna really increase consumer switching. And so we find the more consumer switching there is the more inefficient the market is and the lower the advertising prices are. And then finally we can think about a few other types of innovations in the news industry. So something like Facebook can keep track of everybody. So it's really good at knowing that you've advertised exactly to the number of times you want to a certain number of consumers. If you wanna reach 100 million people in the United States exactly once this week Facebook is the way to do it. And so Facebook actually gets a premium in the advertising market because it can manage this reach and frequency trade off. There's also going to be benefits to mergers or at least mergers to share information because news media outlets that can track better can provide more reach. And so the multi-homing advertisers would prefer to advertise on a big outlet than a small outlet. If they're only, I mean the single homing I'm sorry. So if you're only gonna pick one outlet and the only reason you're not advertising on the other one is because you're getting wasted users you'd like to pick the big outlet and so the bigger outlets get higher, get a premium in the advertising market so they make more money per user. So that's one of many reasons I should expect a lot of media mergers. People often ask me what's the future of the news industry? They're gonna consolidate. For the advertising market reasons because they can't cover their fixed costs for a whole bunch of reasons you're not gonna have very many newspapers left and that's got some pretty important social implications. And finally paywalls, this is one of the ways the newspapers can try to claw their way back and stay in business. Well, paywalls have this problem if they shrink your audience then you're small and so in addition to losing the revenue from the users you crowd out you also make less money per user because there's a benefit of scale in the advertising market of getting the single homing advertisers. So there's kind of a double whammy from paywalls which makes them less attractive. Okay, so that's basically just a quick overview of some of the theory and maybe it just gets you thinking there's actually lots of questions that are still unanswered in two-sided markets. A lot of the early models just made assumptions about how many outlets people did business with but in most of these real world markets you get some people go everywhere some people go to some like some advertisers go to both Google and Bing some only go to Google some apps are on both Android and iOS some are on only one. So these models of where scale is important being big is important for attracting the sellers are quite important and where the sellers actually make choices and they don't all do it. Okay, so now let me move into the empirical work and then the next project is going to look at the effect of things like Google News on the news industry. So again, here's some more just basic facts that news consumption is moving online. This is the ad revenue for the news media and this is digital and this is print. So we see the ad revenue has kind of fallen off a cliff and if you look at over a longer time period it's really quite pronounced. This is the number of newspapers per 100 million people since 1945, of course, there's been a huge decline and the newsroom employment is plummeting. So this is the number of reporters per million population and it's falling very fast. And I think just in the last couple of years it's declined really very significantly. So this is part of like why we're yelling at the news media for giving us Trump but we're not actually putting a lot of people out there to uncover information. So why did newspapers suffer? Well, one thing is that the internet took time away from paper news and actually Matt Genska and my colleague at Stanford wrote a paper showing that advertising revenue per hour on newspapers declined as a result of that. In addition, the internet has just made it much easier to share the content of news articles. Now actually if you think for a minute if a student came to you today and said, suppose you'd never heard of a newspaper, they said, I have this great new idea. I'm gonna create a newspaper. I'm gonna spend a lot of money, hire reporters, they're gonna go investigate things. I'm gonna post them on the internet and freedom of speech will not allow me to copyright that. Basically anybody can copy the content of my story in about two seconds. Does that sound like a great business? So it's kind of crazy this ever was a business in fact. I mean it was a business because of limitations of the technology, that it took a day to copy stuff. So in some of them we shouldn't really expect this to work at all and I think it's only really a concentrated industry that allows it to persist in the end. So aggregators, aggregators are particularly contentious. Pure aggregators such as Google News don't produce any original content. So they don't hire any reporters. They just have an algorithm that goes out and finds the best content from everywhere else and puts it on the webpage. And that webpage looks a lot like the home page of a newspaper. So it looks like the newspaper except that they don't have any reporters. So there's a big question, a big policy question around this, are aggregators substitutes? They look like newspapers. So it looks like they should be competing with your favorite news outlet. And indeed if you look at the data, users only click through to read the articles about half the time or even less. So the ones that aren't clicking through presumably just went to visit it just like they would a news home page. And so they're missing out on of course the newspaper who produced the snippet. If you saw that same snippet on the newspaper's own home page, the newspaper would sell an advertisement and they would make money. But if they read that on Google, they don't make any money. Another concern is what the industry calls commoditization. That is you see the headline and the snippet but you don't really see the brand of the news outlet. And so you don't get as much of a reputation for quality because all that matters is how Google ranks you not actually the outlet itself. So you're kind of reducing the brand effect of a news outlet. Now again in a traditional industry with traditional competition, we might think this is good like the newspapers are charging prices if they were some other product. We might think it's good if they're substitutable and Google news would be kind of like a price comparison engine that helps force them to compete aggressively. But this is an industry where they have to do research and development every day. They have to go out and build the stories every day. And so it's important that they have some rewards that they have some market power or else they have no incentive to create the story. Of course in drugs we worry about this and we have patents on drugs but we don't have patents on news because news is important for speech and democracy. On the other hand you might think that these are compliments because they help in search and discovery. So actually I read Google news almost every day and I'm very sure that Google news makes me read a lot more news. Every time I'm about to start a referee report up pops Google news and I find something very interesting to read. And once I've seen the New York Times run so I've kind of seen what they have for today so that only gets me out of the first five minutes of my referee report. For the second five minutes I can go to Google news and they'll constantly have fresh topics and if anything is happening anywhere it's gonna be on Google news right away and so I know I'll be kind of the first to know on Google news. And so it's basically reducing my transaction costs. And actually you know when I lived in Boston I was at Harvard I never went to the Boston globe. Nothing could be happening in Boston that was that important to me. I was a citizen of the world I read the New York Times but then Google news started showing local Boston content and sometimes there was something interesting there but for me most of the time the search costs were too high for me to go to the Boston globe but when I Google news occasionally would make me realize huh something important did happen in Boston so I would go and look at the Boston globe and sometimes I might go back. So maybe they're compliments, okay. So this has been a big topic recently in the EU. So first Germany passed an ancillary copyright law in August 1st, 2013. Newspapers could decide unilaterally how much to charge aggregators. But they didn't negotiate collectively. So each newspaper individually could decide how much to charge. So if any individual said hey Google you have to pay me for my content they said too bad for you, price is too high. There's another guy here who's giving me the content for zero and they have the same stories you do. So I'm just gonna take all of your traffic and give it to your competitor, okay. So Google news wouldn't lose anything but the individual newspaper would lose all of their traffic, okay. So nobody charged and it was kind of a bust. Then Spain said okay we're gonna be smarter better than those Germans. We're not going to allow the newspapers to opt out the newspapers must bargain collectively. So we're gonna force the newspapers to bargain collectively because we realize there's a prisoner's dilemma problem here where each individual newspaper would provide their content as long as some did Google was fine and the rest of the newspapers would be shut out. So there's no way that they could they could acting unilaterally get any money out of Google at all. So they all had to negotiate together, no free license. The next day Google news shut down, okay. So that was the end of Google news in Spain and now they're discussing having a similar EU wide ancillary copyright law for all of EU. So this is actually a special case of a broader phenomenon where there's a platform that has a lot of users and there are content providers that are substitutable. In that type of situation the platform has all the bargaining power. So another example right now is Apple Pay. In this case in the United States if you put your credit card into the Apple Pay wallet Apple will charge more than twice as much as Visa and MasterCard get for using the card through Apple Wallet and they will not allow the banks or credit card providers to use mobile payments in any other way on their phone. Okay so the only way to use mobile payments through the NFC radio is through Apple Wallet and if you use Apple Wallet, 0.15% of every transaction goes to Apple. You can imagine this is very popular especially in countries that don't have Apple as a tax-paying resident but the, and so what's the problem? Like in the United States people have lots of credit cards and there's lots of banks. So if one credit card company puts their card into Apple Wallet the consumer will just switch to using that credit card and they'd be, they're much happier to use their second or third credit card and keep their iPhone than they are to go buy a Samsung phone. You're not gonna go buy a Samsung phone just because you like your first credit card so much more than your second credit card. So when one of the banks pulls, tries to hold their credit card back from Apple nothing really happens and except they lose all their customers because the other credit cards are substitutes. So they have all the bargaining power so they can, they end up like getting more rents than the credit card networks. Okay, so this type of application I think is, we're gonna see lots and lots of antitrust cases about that so I just told you a theory, nobody's actually written down that theory. I mean I wrote it down verbally for the Australian Competition Authority but nobody's actually written the paper that kind of completely formalizes it although it's a pretty close application of the competitive bottleneck stuff but there's a few more details that still need to be written down for that type of situation. So it's an interesting example. Here the European Union is basically recognizing the prisoner's dilemma problem faced by the newspapers and it's saying maybe they need to bargain collectively. In Australia the credit card companies ask to bargain collectively against Apple and there's not, hasn't been a decision yet about whether that will be allowed or not. Of course you could also just regulate the fees until Apple they can't do it but that's still TBD and that's happening in Switzerland and Korea and in some countries where there's actually a single large credit card company they've actually managed to negotiate the fees down with Apple but in countries like the US with lots of them the fees are high. So we're gonna see a lot of cases like this so here's another theoretical problem to be done. Okay so what are we gonna do here empirically? We're gonna try to figure out what was the effect of the Google News shutdown in Spain. And so we're gonna do this using data from a couple of hundred thousand internet users in Spain and we're gonna try to understand the effect of the shutdown. Now it's actually not entirely clear how to do this because the shutdown was December 16th. People don't read a lot of news around Christmas so there was actually a big negative time trend right around Christmas to start with. So we need to somehow control for what the users would have done without the shutdown. So we can do this doing a kind of difference and different strategy with what's called a synthetic control group. So it turns out that only a small fraction of Spanish users even use Google News. So if you never use Google News, of course there's no effect of Google News shutting down on you. So the only users that are affected are the ones that use Google News to start with. And then I wanna control for the Christmas season by finding users that are like the Google News users but they don't use Google News. So the way we started out doing this is we said well let's look before the shutdown, find users that are similar and then track them afterwards. But then we realized that actually wasn't the smartest way to do it. Because before the shutdown, they actually had sort of different technologies and different search costs and what we really wanted were users who would act the same without Google News. So then we realized actually a smarter way to do this would be to take a period, a month after the shutdown, use that period when nobody has Google News to find users that are similar to each other. So we see if you don't have Google News, what do you read? And we wanna find users that read the same thing as you when you don't have Google News and then project backwards to before the shutdown. It turns out we get the same numbers either way actually. So it doesn't really matter. If you don't like it, don't worry about it. But this is what we actually did. And so for each treatment user who uses Google News, we find a control user who is very similar to the treatment user. And then we compare the behavior of the treatment and control users before the shutdown. So the idea is if we take Google News away from the treatment users, they'll behave like control users. So then we try to set up a little model that we use to organize our empirical work. The model is really stylized, that's we don't take it too seriously, but it just motivates the functional forms we use for the analysis. So in this model, the users maximize utility, their news consumption is the product of a user effect and the data effect. And that's gonna allow us for the treatment user to kind of have a control user be able to proxy for the time effect. So I'm gonna learn about the user's intensity of using news, say from the post shutdown period, and then I'll figure out how much news they should be reading from the control users. So the fact that I can decompose them just helps me do that. In addition, the user's gonna consume news of different types and shares that are fixed over time. So for example, breaking news or popular topics, we assume they have fundamental preferences for these things, even though their search technologies might change. And so aggregators are basically gonna change the efficiency of search for the different types of news, which is gonna determine how much you read. And they also provide direct consumption of news. So reading an aggregator is like reading a newspaper homepage. So we basically think of one unit of time translating into a set of articles, and the news found through direct navigation, search and social, all of these different things is going to be denoted in this way. So the total news consumption is the sum of the news in a particular category that come from directly navigating to the website, getting them from search, getting them from social, and from home pages. And then through the aggregators, they're gonna get them from Google News, either from reading articles through Google News or just from reading the Google News homepage by itself. And so we assume the aggregator to augmented technology helps you discover more news, but then we can define substitutes and compliments according to whether when you read the news, you use through the aggregator without Google News, plus the articles that you click through using Google News is that greater or less than the amount that you use to read. And so the idea is that the homepage views on Google News are gonna crowd out other stuff, but Google News is also gonna send you through to more articles which can make it a compliment. So whether the sign of this relationship is gonna determine whether your substitutes are compliments. So the model for the consumer demand, we can think of a user is gonna consume, according to this Cobb-Douglas utility and everything, we can just write down how much news they're gonna read. The user of the aggregator is going to read this additional amount through the Google News and the homepage. And the news that's directly consumed on the website can likewise be written like this. And this nice functional form is just gonna mean that when I do things like take ratios, I'm gonna be able to cancel out the day effect. And if I have treatment and control users who are similar, I'm gonna be able to cancel out the user effect. And then I'll be able to interpret what's left in terms of how the technology helps them search. So when I go to do this, matching just real briefly, we match people among a bunch of different dimensions. We have a few hundred thousand users to choose from so we can find users that are similar in a whole bunch of ways. Overall page views, news page views, news article page views, articles access through search, breaking news and we call it hard news that's news other than celebrity and sports. So we find people that are similar in all of those dimensions to our treatment users. And again, we're not very sensitive to those assumptions. So what does the data look like? So this is our matching period. So the treatment and control users look very similar in terms of their news consumption in the matching period. This is not a matching period, but they also track each other very well. Weekdays, weekends, good days, bad days. Right here, this is when the Princess Christina was accused of corruption. So one thing you learn when you use internet data is that events matter. So these things kind of jump around all over the place. This is a big spike. We can drop that from our analysis and it doesn't matter. But then we see here are users that when Google News wasn't available did everything exactly the same all the time. But when Google News was available, those users read a lot more news. So this is showing that Google News for these users helped them read more news. They read more. So we can look at the referral shares and the referral shares are basically what share of news did you get from different sources? So direct navigation is when you go directly to a news homepage and then you read articles there. Google News, here for the treatment users, in the beginning they were getting about 24% of their news from Google News and of course that goes to zero when they go away. And so then you can see well what did they substitute to? They substituted towards search and also towards direct navigation. So when you take Google News away, that's how they find their news afterwards. So as I mentioned, the functional forms we use were just used to motivate an empirical specification where we can look at the ratio of the news between the treatment users and the control users. And so if we do that analysis, we can look at the actual magnitudes of the effects. And so what we find is that in total, the presence of Google News increases news consumption for the treatment users by about 20%. Excluding what you read on Google News, so from the perspective of the newspapers, it increases news consumption by 11%. But that's total news, that's articles plus the landing pages. But the newspapers are not indifferent between whether you go to their homepage or an article page. All of the very expensive advertising that they sell is on the front pages. And the advertising on the article pages is often sold through Google DoubleClick Ad Exchange where Google DoubleClick takes about a third of the revenue and they're also sold in this very inefficient way with lots of duplication and so on. Okay, can I perform under pressure here? So what, and if you break things down to the landing pages, that's like the front pages of all the sections of the newspaper and the articles, you see that all of the benefits are accruing to the articles and they're actually losing views on the home pages of the newspapers where they make all their money. So from a financial perspective, it's a negative thing. And we can do the same thing with dwell time. You might think that people just go to Google News and look at the page quickly and bounce back, but we see sort of similar effects in terms of the amount of time that they spend on the pages. So the first result is Google News is a complement to overall news reading and articles, but it's a substitute for landing pages, which makes sense because you're going to Google News and you're just reading the headlines just like you would read the headlines on the newspapers. But I think the kind of more interesting part of this, and I think the part again that could inspire some more theoretical work and should also really be considered for policy is the question of how does it affect competition and how does it affect entry and market structure? So if you break things out to the top 20 news outlets and the bottom 20 news outlets, we see that the small outlets actually benefit across the board. They get a slight, it's insignificant, but at least a slightly positive effect of Google News on their home pages, and that's probably a little bit like me with the Boston Globe. I actually might go visit the Boston Globe after Google News told me there was something interesting happening in Boston, or I might click on another home page, another landing page of another section while I'm at the Boston site. For the top outlets, there's a very significant negative effect on the landing pages. So basically your big outlets, the ones that hire reporters and send them to Syria and do investigative journalism and all of those things, those are really hurt by Google News. And again, they make all their money here. They're still, though, getting an increase in articles. They're still getting people to come read their articles more as a result of having Google News. So generally the welfare effects of this I think are kind of ambiguous. So there's a literature, and actually in the communication schools in Europe in particular, there's a literature on media diversity. And so the media diversity people kind of focus mostly on the benefits of all of these small news websites. The fact that they provide unique perspectives. So if the small outlets are providing unique perspectives, then this chart looks like Google News is great. It may be good for welfare. On the other hand, if you think the small outlets are mostly copying other people's news, then it can be a negative effect for welfare. And we don't know that without more measurement, more empirical work. So just summarizing, Google News is neither a compliment nor a substitute to overall news reading for large publishers, but it changes the mix towards articles and away from landing pages while it's a strong compliment for small publishers and it raises total news browsing a fair bit. And so this is again why the big outlets, they might call this in business lingo, commoditization. They used to have this brand and people associated them with high quality reporting and now they're just shoved in there with all the small news outlets who might not have any reporters at all. Then the next thing we do is we try to understand why this is happening and so we try to look at the type of news. So we look at news characteristics, things like breaking news. We look at the publication time. We have something we call supply scarcity. So we actually create topics out of the news by looking at the text of the articles that these people were reading. We put them into topics and then we call it relatively scarce if for an individual user, if the user's favorite newspapers don't cover that. So like if you read Fox News, Benghazi is very well covered. In fact, if you talk to the editors of Fox News, they will say part of their brand identity is if anything happens about the Benghazi terror, attack, then it will be on Fox first, okay? On the other hand, if you read The New York Times, Benghazi is covered less than average, okay? So that's a difference between them. So if I read The New York Times, news about Benghazi is relatively scarce for me because my news outlet doesn't cover it as well as others. We also look at popular news in terms of the topics, like which topics are very popular and hard news that's not celebrity or sports. And so we find that Google News is an important source of user scarce top of news. That is for the Google News readers, the scarce news is coming much more from Google News than from other sources. So we do an exercise where we basically try to understand the impact of Google News on different types of news. And so we break things out by how many hours it is since the news story broke as well as whether it's non-scarce that's blue or scarce that's actual. And so we predict what the counterfactual drop would be if the Google News users just didn't read Google News and they didn't replace it with anything else. And so the counterfactuals are the light and the actuals are the dark. So for very recent news and especially for scarce news, the Google News readers would lose 60% of their reading if they didn't read it on Google News. And furthermore the actual and counterfactual are basically the same which means that the counterfactuals if they don't find any substitutes and the actual is they don't find any substitutes. So down here there's no substitution where for news that is non-scarce that's well covered by their own favorite home pages these bars are very different which is showing that if they find perfect substitutes if Google News goes away they still read about those topics. So it helps you understand how this competition works. Now if you're a newspaper this is kind of depressing because it's saying how does Google News compete with me? Well they have everything and they have it right away. And if you're in the business of hiring reporters it's hard to have everything everybody has written immediately because you actually have to write the stories. So it's very difficult for a newspaper to actually compete directly against Google News for this type of news reading. So just gonna skip a little bit. We try to do a decomposition to say how do we decompose the volume drop? And again we find relative scarcity being very important. Also breaking news and somewhat hard news are important things that are hard to replace. So we actually did this also in France. It was a harder natural experiment because what Google did was they actually allowed you to opt in to Google News in 2009. And so I collected, you can see how long I've been working on this project. I collected this data actually 18 months after it happened so I was grabbing the data as fast as I could before it got deleted. But what happened was that you could put in what city you lived in and then you would see prominent local content. Now here the problem is the users had to choose to get this and so what you worry about is those users were particularly interested in local news right at that moment. And so the way that we do that, we deal with that is we make each user the date that they adopted Google News, we find the local on Google News. We find other users who are also reading a similar amount of local news right on that very same day. So we do our matching in a much more sort of time specific way to find the appropriate control user for each treatment user. The one advantage of this natural experiment is we don't have like a Christmas effect because everybody adopts on a different date and so we're gonna be able to kind of smooth over any kind of holiday or seasonality effects. But you have to worry that we haven't fully controlled for the fact that these users were really excited about local news at a point in time. And so if we compare the treatment and control group in the day relative to adoption, we find that after users sign up to read, to have local news on their Google News homepage, they read a lot more local news on the local news outlets. So again, this is a complementarity story. It's not just that they read the articles at Google News but they go to the local news outlets and they read it there. And that's basically because local news outlets are small outlets. So we already saw that small outlets are helped by Google News, local news is also helped that way. This is kind of an interesting natural experiment because we had a whole block of news and it helps you understand kind of also what happens if they could negotiate collectively. Because we also find that having that local content made consumers like Google News a lot more. So if you took it away, people were actually gonna use Google News less. So it was actually important for Google to have that collective block of content. And that tells you a little bit about the benefits of collective bargaining like that is being forced. Hey, last thing, and I'm just gonna have a few, couple minutes to give you highlights from this. This is work in progress. We're looking at social media in news consumption using similar data sets. So I mentioned before that we put news into topics. We did, this study was done in the US and we did it just for a couple of weeks in a prototype analysis and we're extending it out until basically right up to the election for over a two year period. So we're just getting those results now. In the prototype, these were the topics that we had. So this is how using text, we're able to put the articles into groups. We do it using a network classification method for machine learning. So then we use this to figure out how social news changes what we read. We do not have a natural experiment here. So all I'm gonna show you now is sort of comparing what people read from social media like Twitter and Facebook to what they read when they navigate directly. And the only way, I think I'm gonna be able to do to adjust for the fact that different people read social media and go directly to newspapers. So I'm basically gonna have like user effects. And so to illustrate that, I'm gonna have these green bars in the middle. And the green bars are sort of counterfactuals. Those are the counterfactual consumption if the user mix was the same on social media as it is on direct navigation. So if the green bar and the blue bars are similar, that's saying that the difference between social media and direct navigation is accounted for by the user mix. And so when we look at broad topics like we put the topics into super topics, crime, politics, et cetera, we find that roughly once you account for user mix, people read the same stuff when they come from Facebook. And this is mostly Facebook. When they come from Facebook to when they go directly to newspapers. So then we try to say, okay, within a topic do reporters actually write different stories? And there we start to see much more difference. So for example, we find on social media that you see more of an individual perspective from articles, slightly more complicated articles that assume more background knowledge, a bit more in depth analysis, much more socially controversial, more controversial on one sided, more opinion articles, more politically controversial, also more caring supportive stories. These are the things that people read when they come from social media relative to when they go directly from news websites. Things targeted more at women, targeted at parents, kind of people tend to read stuff that's more tailored and targeted to their interests. Of course, the social media websites do that for you with their ranking algorithms. Okay, and some results on emotion, a little bit less in terms of neutral emotions of the articles. Okay, so let me now at the very end, I'm gonna show you results about political bias that these results I wanna caution, these are not probably gonna hold up exactly in magnitude because our sample is too small. In a typical two or three week period, most people don't read that many biased articles unless you take the last three weeks, then they read a lot. So in this time I studied, I just don't really have enough data to do this exercise, which is why we extended it to a two year analysis instead of a two week analysis, but this is really a lot of work to do these analyses and so it's taken us a while, we have to wait a couple of years and collect the data and then analyze it. But just to give you a sense of what's going on, first of all, one interesting thing is one of the biggest sources of bias of news websites is what topics they cover. So it's not that the Fox News covers a particular topic so differently than the New York Times, but they just cover different topics. So different and then different topics in turn have a lot of bias. So school shooting articles are liberal, articles about Benghazi are conservative and so on. And so when you actually go and look at what explains bias in what you read, if you control for the topic, the outlet effect kind of goes away in the US, at least among mainstream news outlets. So you can explain all of the bias of Fox just by the fact that they cover all scandals for the Democrats all the time. But then it's not so much that they're covering it and they also are covering it differently, but that's not the most important effect. So if you're reading these outlets, you think the only thing that's happening in the world right now is whatever latest scandal they're covering. Then we try to look at political bias. And so first of all, we look at all readers and we see how the bias of the articles varies from social media and direct navigation. And so we see less neutral stuff on social media. But this is of course, this is mixing together liberals and conservatives. So to really do this, we wanna look at the liberals read conservative into conservatives read liberal. And this is the part where I just don't really have enough data. So just take this with a bit of a grain of salt. But nonetheless, what I find is that people who read a conservative article in the last two weeks are very unlikely to read liberal news anywhere, but they're much more likely to read neutral news through direct navigation. So a liberal reader is gonna read neutral news, say on the New York Times, but not so much on Facebook. And the same for conservatives, conservatives aren't gonna read neutral news. They're not gonna read neutral news through Facebook, they're gonna read it through direct navigation. And so if people start getting all their news on Facebook, maybe you're gonna see this echo chamber. Now some of you might be familiar with some research by my colleague, Matt Ginskow, that was kind of published in the AR that said there wasn't a filter bubble. The main difference between my work and his work is that I'm looking at the individual articles that people read. And so I'm only looking at political stuff here. So people read about lots of stuff. They read about tornadoes, they read about sports events, they read about all sorts of things. And so here we're only focusing on news articles rather than which outlet you read. And so that's responsible for the big differences in our results. Okay, so I'll stop here and take questions. Questions, okay, does somebody else, should I, let me see, do you wanna bring a microphone? So thank you very much for such a stimulating talk. Just wanted to ask you a bit about your results about news aggregators and their effects on local news businesses. So basically I don't personally use Google News. What I usually use is Reddit. And I don't know how Google News works, but Reddit is much more customizable in terms of what the link actually says. And the users actually get to, for example, put tags on the articles to say what, from what news source it comes out of so that people get to actually know what is the source of the news that they're getting. Do you think that would affect the results, the positive results that you see for, say, the small businesses, the small news outlets? The different aggregators, you know, show differently the source of the content. Oh, is this working again? Yeah, good. So, but even on Google News, they show the name of the outlet, but the problem is I think when you see that, just like when you see it on Facebook, you just don't pay that much attention to it. So the brand, if you think about where do you navigate directly, the brand effect is everything. Cause you probably have three websites memorized, right? Four, you don't even know how to get to all these other websites. So you really go to only a small number. But when you use Reddit or Facebook or Google News, it's just much easier to go to lots of different things and you're gonna put a lot more weight on the headline than you are on the news outlet. So the brand effects are still there, but they're just small. This goes back to the thing about the difference between the big news outlets and the small ones. You seem to suggest it was bad for the big news outlets that the landing pages get fewer views, but that seems to presume that their advertising strategy doesn't change. Now, as they come to realize that, they presumably will tilt their advertising much more to the article pages and not to the landing pages, and it's not so obvious that you have such a big trade-off. That's right. Part of the reason that those landing pages are so valuable is that if an advertiser wants to reach a whole bunch of consumers like once, then consumers are gonna tend to go to that home page. So if the newspaper kind of wants to sell 100,000 views, the easiest way to sell it is on the home page, and then also the buyer, the advertiser, knows what they're getting, because they understand that the Sacramento paper has local lawmakers in California, that's the capital of California, so when they advertise there, they're gonna get lobbyists and other people in California. If you go to the article page and you realize that a lot of those people are coming in through Google News, you don't really know what you're getting. So it's not as valuable to the advertiser, and those people might be people that you've gotten other places already, possibly. So it's not quite so easy to do that. If they could do the tracking better, if they could keep track of the people, then they could just adjust their advertising strategy. But that's not, right now essentially, they hand sell the front page and keep all the revenue and a lot of the articles, they sell it on Google DoubleClick for like a tenth or even a hundredth of the price. So somehow they need to fix some of these other problems to be able to monetize it better. Thank you very much for this beautiful, really beautiful lecture. I was wondering because when you were studying media in the traditional press and TV, there was very often a monopoly of government, and there is a monopoly or very few press agencies in the world that usually newspaper use to make the article before internet. So these competition issues is already there in the traditional press industry. So, and then there is also the competition of TV that seems to absorb the best journalist. So, which is not in your analysis neither. So I was really wondering what is totally specific to this innovation and what policy implication you will suggest to improve the quality of what you are, well, the news we have. Thank you. Sure, I think the issues are actually very similar and across time. So in some sense the reach and frequency issues, if you looked at a media planning problem for an advertiser in 1970s, they worried about the same thing, right? They might buy lots of small TV shows or a few big ones. And the big shows were more dollars per user than the small TV shows just because of this overlap issue. So it's the same types of effects. I think the thing that's different about the internet is just the magnitudes. People switch so much more. You know, instead of seeing three shows or five shows or 10 shows, you know, now it's dozens of different websites. And then the speed of copying is also just so fast. So I think it's really just the magnitudes. But I think that these things like say the Associated Press, where you basically consolidate on the expensive reporting, I think you're gonna see more of that. It's gonna come back. And we'll just have very few outlets that really do kind of their own branded, independent, expensive reporting. And then you can get local reporting, perhaps more cheaply even, because you can get citizen reporters and people who for their hobby like to sit in the legislative sessions. And so you can kind of gather up some of the cheap local reporting inexpensively. But the thing you're gonna miss is like the expensive reporting that somehow not covered by these centralized agencies. Those are the things you'll kind of miss out on. Thank you. I kind of have the feeling that there is a change in the business model of large newspaper that towards subscription instead of advertising. How do you view this, would there be a kind of two different ways to make money for this industry either by subscription or by newspaper, by advertising? That's right, for some of the outlets, the subscriptions will work. I think the problem is that for some of the newspapers, like there's a profit maximizing subscription price and that's zero. And so they say, well, we can't pay the bills from advertising, so we're gonna try subscription, but we also can't pay the bills by subscription. And so ultimately they're gonna merge and consolidate. You have to be pretty big for both things. You have to be pretty big to make the economy's scale work on the advertising. You have to be pretty big to cover the fixed costs of having a website and just all the overhead. And you need to be big enough that people value your brand and care for subscriptions. And if you're big enough and you spend enough money, then you have unique content and then people will pay for that unique content, which is why I think you'll have consolidation, not to say consolidation is good. Just economic forces. Is that good? Great, well thanks very much.