What this covers
Hal Varian discusses how economic theory applies to digital markets and the internet economy, with particular focus on mechanism design, auctions, and price discrimination. The conversation moves across Google's origins as an academic spinoff, the mathematics behind ad auctions, and why textbook models often diverge from organizational reality. Varian draws on decades of work spanning fair division, demand analysis, and the design of algorithmic systems, connecting classical economic theory to concrete problems in search, pricing, and market structure.
The terrain ranges widely: how consumer search tools and competition erode the power of theoretical price discrimination; the fusion of computer science and economics in algorithmic mechanism design; why durable goods monopoly explains textbook pricing; and how revealed preference inequalities from his 1982 demand paper became the foundation for Google's auction model. Varian argues that real-world friction—organizational messiness, behavioral departures, asymmetric information between buyers and sellers—consistently tempers dystopian predictions theory might otherwise generate. He also reflects on the gap between textbook ideals and practice: how rapid firm response to price cuts can sustain high equilibria; why operating systems have become less dominant than network-effect theory predicted; how music streaming succeeded where piracy thrived by offering frictionless access; and how excitement and irrationality in auction houses depart from theoretical equivalence. The conversation includes practical advice for economist-technologists, the credibility revolution in empirical economics, and observations on envy, fairness, and institutional design.
Varian argues that economic theory—particularly auction design, mechanism design, and price discrimination models—remains highly applicable to real-world digital markets, but that competition, consumer search behavior, and messy organizational reality consistently temper the dystopian predictions theory might otherwise suggest.
- Consumer search tools empower buyers and break down feared perfect price discrimination
- Algorithmic mechanism design fuses computer science and economics for real applications like ad auctions
- Real-world decisions are messier than the idealized textbook cases economists model
Market failures persist when prices are hidden from those paying and incentives misalign.
- Electronic journals stay expensive despite near-zero marginal cost because the end users (faculty and students) can usually obtain papers freely via preprints and don't protest prices, while libraries—who actually pay—are motivated to provide information rather than resist prices, so no one faces or fights the true cost.
“The end users — the faculty members or the students — they can usually get what they want, so they don’t protest these high prices much. It’s the libraries that end up paying.”
- Internal corporate prediction markets fail because the most valuable predictions—about sensitive matters like acquisitions—would turn any participant who views the market into an insider, violating SEC insider-knowledge rules, so the company can only run markets on questions of interest that don't reveal financially critical information, which is hard to find.
“the things that we really wanted to get a probability assessment on were things that were so sensitive that we thought we would violate the SEC rules on insider knowledge”
- Textbooks are priced remarkably high due to a durable goods monopoly problem: as used copies of previous editions compete, each new edition must differ markedly from the last to sustain the pricing model, a task that is getting harder and harder.
“there’s a durable goods monopoly problem there. As you have more and more competition from previous editions, each of the new editions has to differ markedly from the old edition to support the pricing model.”
- An internal prediction market won't succeed if it merely duplicates a market that already exists, such as Ford running futures on gasoline prices; to be valuable it must add something beyond the usual external markets to attract insiders with relevant information.
“You’re not going to get anywhere if you’re just duplicating a market that already exists. You have to add something to it to make it attractive to insiders.”
Consumer technology and search capabilities counteract seller pricing power in practice.
- Even where a retailer like Amazon offers superior convenience such as one-click ordering, it cannot extract much surplus because the most it can capture is the difference in the consumer's valuation between one click and two clicks, which is minor.
“the most they can make is the difference between your valuation of one click, two clicks, whatever.”
- Fears of highly personalized pricing leaving consumers with no surplus are unfounded because the internet makes it so easy for consumers to search and find lower prices, meaning technology has actually empowered consumers rather than handing pricing power to sellers.
“That story seems, to me, to be ridiculous because it’s so easy for consumers to conduct this kind of search and try to find a lower price. So consumers have been empowered by this technology in my view.”
- We failed to foresee with Napster that having well-organized, high-quality, readily available complete access to music would be much more attractive than cobbling something together via back doors; it's now clear people willingly pay $10 a month for smooth access to the content they want, which is why few young people pirate music.
“having a well-organized, high-quality, readily available complete access to music was much more attractive than cobbling something together using some sort of back doors”
- Research on online shopping shows rural residents on Midwest farms love online shopping because they lack the physical-world shopping opportunities of Manhattan residents, so the relative value of internet access, shopping, and content is higher for rural than urban users.
“if you live on a farm in the Midwest, you love online shopping. If you’re living in Manhattan, you’ve got a lot of opportunities to go shopping in the physical world.”
- Algorithmic price collusion among AI sellers is hard to sustain because as more sellers enter (triopoly, quadropoly, etc.) coordination gets harder, and especially because there are AI agents on the consumer side actively searching for the lowest price—so active price-setting by sellers is offset by active search and choice by consumers.
“It gets harder and harder to coordinate all of those activities, especially if you’ve got an agent on the consumer side who’s looking for the lowest price all the time.”
Auction design and bidding behavior diverge from theory due to psychology and emotion.
- A practical bidding heuristic is to never attend the auction with the best inventory because everybody shows up there; instead go to the second-best auction, where you have a better chance of beating dealers and exploiting inefficiencies at small-scale auctions with nonprofessional buyers.
“you never go to the auction that has the best stuff because everybody’s there. You go to the auction that’s sort of second best because then you have a better chance of beating the dealer.”
- Auction houses like Sotheby's likely earn higher revenue from the ascending/descending-bid auction than from the theoretically equivalent Vickrey auction because excitement, irrationality, and madness of crowds cause bidders to get caught up in the bidding, departing from theoretical equivalence.
“there is some degree of excitement, irrationality, madness of crowds — whatever you want to say — that people get caught up in the bidding. And they probably do make higher revenue from the descending-bid auction than from the theoretically equivalent Vickrey auction.”
- The winner's curse is most relevant under asymmetry of information, and while naive players do exhibit it, the experimental literature shows they learn over time to shade their bids for a better outcome, indicating learning behavior rather than a permanent irrationality.
“naïve players do have a winner’s curse, but they learn over time to shade their bids in a way that gives them a better outcome. So there’s a learning behavior going on there.”
- An empirical analysis of historical Dow Jones prices found that price increases accelerated as the index approached century marks (like 1,000) in most periods—not a theory tested but a data-driven discovery of unusual behavior near psychological barriers.
“What we found out was that the price increases accelerated as they approached the century marks in most periods.”
- Although people describe financial trading volume as massive, it is usually a relatively small fraction of the total assets held—even on a very heavy trading day, only a few percent of the total volume of assets changes hands.
“Even on a very heavy trading day, that might be a few percent of the total volume of the assets. So it’s not so big.”
Competition and entry, not regulation or piracy deterrence, drive price improvements.
- Journal price reductions come from competitive entry rather than piracy, as evidenced by the American Economics Association launching four new prestigious, reasonably priced journals, with similar entry occurring across scholarly disciplines.
“where we’re going to get the price reductions is competition. And you’ve probably seen that in the last couple of years, the American Economics Association has offered four new journals.”
- If firms move faster than consumers—contrary to the textbook assumption that consumers respond first—then when one gas station cuts price the other immediately matches before consumers can shift, eliminating any benefit from price cutting and sustaining a super-competitive (high) price equilibrium.
“one cuts its price. The other immediately matches the price, so there’s been no benefit from price cutting, and the consumers haven’t had a chance to respond. There, we end up with an equilibrium model where you can sustain a super competitive price.”
- The existing US antitrust legal infrastructure is pretty good largely because the DOJ and FTC publish guidelines (merger guidelines, intellectual property guidelines), whereas Europe is a less mature system that doesn't provide such guidelines, leaving firms 'flying in the dark' to a larger degree.
“the DOJ and the FTC put out guidelines, like the merger guidelines, the intellectual property guidelines... If you look at the situation in Europe, it’s a less mature system. They don’t provide you with those sorts of guidelines”
- In France, retailers are only legally allowed to hold sales twice a year, and these must be legitimate price discounts—meaning the retailer had to be charging a higher price beforehand and then actually cut from that previous high.
“in France, you’re only allowed to have sales twice a year, and they have to be legitimate price discounts.”
Economic models work as ideals but must accommodate behavioral reality and human diversity.
- Textbooks carry more exactitude than real life requires; just as no battle plan survives encounter with the enemy, organizational decisions are far messier than textbook models, so economics should treat textbook results as an ideal case and study robust departures from it, including behavioral economics.
“when you look at how decisions are made in organizations, they’re often a lot messier than they are in the textbooks. We might think of the textbooks as kind of an ideal case, and what we want to allow is that there could be departures from that ideal case.”
- Working in the private sector reveals consumer behavioral distortions in saving, self-control, and handling complexity, so when doing something like pricing one should consider both the textbook view and a behavioral view, often ending up with a blend of the two.
“when you’re doing something like pricing, you want to think about the textbook view, and you want to think about a behavioral view, and maybe what you’re going to end up with is a blend of the two.”
- There is more financial trading than any reasonable Bayesian model would predict because people genuinely have differences of opinion and aren't fully Bayesian—they don't find others' opinions credible, so we never reach 'agreeing to disagree'; Varian's own work found it comes down to people having different models that they can't agree on, and disagreement on the model produces non-uniformity in beliefs.
“it really came down to people do have a different model. We can’t agree on the model. If we don’t agree on the model, then we won’t get uniformity.”
- Graduate students should wrestle with a research problem on their own before consulting the literature, because reading a fully worked-out solution captures you in that author's viewpoint, whereas floundering independently can lead you to discover a completely different phenomenon—though one must eventually look at the literature.
“if you look at the literature, you’ll see this completely worked-out problem, and you’ll be captured by that person’s viewpoint. Whereas, if you flounder around a little bit yourself, who knows? You might come across a completely different phenomenon.”
Algorithmic mechanism design merges computer science with economics to improve market efficiency.
- Varian's 1982 paper on the nonparametric approach to demand analysis directly inspired his model of Google's ad auction, because the auction equilibrium conditions are revealed-preference inequalities—'I'd rather be in my position paying my price than in another position'—that you manipulate to derive a formula for the price, just as in the demand paper.
“that very paper was the inspiration for the model I constructed of Google’s ad auction. Because in the equilibrium conditions, basically, I reveal preference conditions”
- Algorithmic mechanism design—combining computer science and economics—is an exciting theory area: it brings computational costs into the economic model (requiring an actual algorithm to solve the maximization) while building incentives into algorithms so that users of a shared protocol have the right incentives for efficient use, with strong applications from telecommunications to AdWords auctions.
“you take the economic model, and you bring in computational costs, or show me an algorithm that actually solves that maximization problem. Then on the other side, the computer side, you build incentives into the algorithms.”
- Fresh-out-of-school candidates good for Google must have better-than-average computer skills, not just statistical analysis ability but also computer operation skills, because they must work with Google's powerful but complex infrastructure—a strong requirement that distinguishes a good Google hire from a good academic hire.
“they should have better-than-average computer skills because they’re going to have to deal with the infrastructure we have there... Not only having the skills in, let’s say, statistical analysis, but also the skills in terms of computer operation.”
The economics profession should prioritize faster publication and empirical research over slowing review.
- The 1980 economics-of-sales model maps neatly onto internet shopping because the population splits into two segments: people who use tools to search for the lowest price and people who buy on an as-needed basis without searching.
“Some people search for the lowest price, and they have all these tools to find it. Some people just buy on an as-needed basis. So you’ve got these two segments of the population that fit the model that I wrote in 1980 really quite nicely.”
- Theories of fair division work out cleanly when dividing a fixed pot of goods both equitably and efficiently—such as an estate among inheritors—but break down once production is introduced, because people contribute different amounts and there is a great variety of views worldwide on how contributions should be compensated.
“The difficulty came when you brought production into it because people could contribute differently — different amounts to the production. Then, how much would they be compensated for their contributions?”
- The single reform Varian would make to the economics profession is shorter review periods so work gets into print faster, since the journal enterprise has become really bogged down, particularly compared to fields like computer science.
“I would like to see shorter review periods for journals. I’d like to see things get into print more quickly. I think the whole journal enterprise now has been really bogged down”
- Accounting is not very complex as a computer system but is dominated by gray areas—such as what counts as a good versus a service—where there is no ground truth out there; outcomes are choices made to resolve particular situations and reconciled by agreement that a procedure is reasonable, with commercial systems mainly tracking evolving law and regulation.
“I don’t think there is ground truth out there anywhere. I think there are lots of choices that are made to resolve some particular situation that’s come up.”
- The content of a graduate economics textbook should be determined by what one needs to understand the research literature—surveying topics in journals like the American Economic Review—which is why tools like Shephard's lemma and duality remain warranted since they still appear in current research, including the live debate over how margins have changed.
“what do you need to understand the journals?... that’s what you need to know to read the journals, and that’s what the textbook is supposed to teach you.”
- Since the mid-1990s, economics has seen a great flourishing of empirical work driven by the credibility revolution and better econometric models, rather than a dearth of theory being a problem—and econometrics work has been central to building coherent statistical reporting within an organization like Google.
“What’s happened since the mid ’90s is, we’ve seen this great flourishing of empirical work with a lot of the credibility revolution in terms of coming up with better econometric models.”
- Advice for an aspiring economist-technologist: get the basic skills down (coding, design, economics), then exercise creativity by stepping back from the first or conventional way of looking at something to see the bigger picture—most attempts will flop, but occasionally you hit something new, exciting, and novel.
“don’t just take the first way of looking at something or the conventional way of looking at something, but try to step back and see what the bigger picture is. Now, most of the time that’ll be a big flop. But every now and then, you’ll hit something that’s new and exciting and novel”
- Robert Nozick's entitlement theory of justice was somewhat speculative and likely wouldn't work out as well as Nozick argued, leaving a lot of problems with his theory.
“I would say it was somewhat speculative. I don’t think it would necessarily work out as well as Nozick argued it would work out. There would still be a lot of problems with his theory, in my view.”
Internet infrastructure fragmentation and digital identity verification pose emerging economic challenges.
- The internet's governance is contested between the US, which created and oversaw it for global benefit, and countries demanding more control, and with examples of nations cutting off outside access, there will clearly be fragmentation of the internet in the next several years—likely into about half a dozen pieces.
“I think it’s quite clear there’s going to be some fragmentation of the internet in the next several years.”
- General cryptography is a big deal with huge demand—needed for things like cryptographically signing collections of bits to combat deepfakes and enable secure proof of identity—but blockchain itself is inefficient by nature, with proof of work, a single constantly-updated version, and slowness that an economist dislikes; these flaws are fixable and likely will be fixed.
“Blockchain seems to be, by its nature, relatively inefficient. As an economist, I don’t like this proof of work that this is... crypto in general — big deal. Blockchain — not so much.”
- Tenure will get more difficult to sustain because the push to enlarge educational access requires more university capacity to handle more students, producing a variety of roles beyond the tenured/nontenured professor binary, including intermediary roles like tutors.
“there’s now this big push for enlarging educational access... you have to have more capacity in the universities... there’ll be a variety of roles, not just professor tenured and professor nontenured.”
Information asymmetries, transparency, and user understanding shape tech company social perception.
- Google switched to a first-price auction on display ads to put all ad sellers on a level playing field, addressing concern—which Varian believes unfounded—about a 'last look' advantage where a party in a series of auctions sees the last price and decides whether to beat it.
“Google was switching to a first-price auction on display ads. And the reason they did that is because they wanted to put all of the ad sellers on a level playing field.”
- Facebook is more hated than Google likely because people don't understand the limits of what can be done at Facebook—its data use is amorphous—whereas with Google it's pretty clear the information is used to show targeted ads, giving Google a specific, legible application.
“one of the reasons people are most worried about Facebook is they don’t really understand the limits of what can be done at Facebook. Whereas at Google, I think we’re pretty clear that we’re showing you ads”
- A lesser-known aspect of Google's culture is a paternalism around its free food: a recent employee complaint was that too much kale was being served, reflecting tension over whether providing healthy snacks should be imposed with a light or a heavy hand.
“one of the recent complaints was that there was too much kale being served... there’s a bit of paternalism that goes on. They want to give you snacks that are good for you”
Tech companies emerge from academic research and benefit from governance clarity.
- Google began as an academic spin-off: the page rank search engine came out of a National Science Foundation digital libraries grant, an initiative that spawned three search engines—Inktomi, Lycos, and Google—and the early company of about 300 people had little management intelligence, with Eric Schmidt describing his role as 'adult supervision.'
“the Google search engine page rank came out of a National Science Foundation digital libraries grant... It spawned three search engines: Inktomi, Lycos, and Google.”
- Brin and Page tried to sell their page rank algorithm for about a million dollars so they could return to graduate-student life, and the most definitive account points to Excite rather than Yahoo as the buyer; the deal failed not over price but over disagreement on what search technology would be used.
“they couldn’t really reach an agreement with Excite. Not so much on the price, but on what kind of technology would be used in the search engine.”
Market participation and valuation reflect behavioral factors beyond rational economic theory.
- The most useful model for Bitcoin's price is capital controls: in countries where capital controls bite and people are anxious to get money out due to political instability, demand for Bitcoin rises, and this is really what drives recent Bitcoin price changes—implying that if every country were stable like Denmark, Bitcoin would be worth much less.
“I would say capital controls is what’s going on. You have countries around the world where capital controls have been biting, and there are several places where people are very anxious about getting their money out of the country”
- The typical American envies the next-door neighbor more than the billionaire; billionaires function as a form of royalty that people watch with curiosity rather than envy, and being a billionaire carries large costs—loss of public anonymity, need for bodyguards, logistical burdens—making it better to be a half-billionaire.
“the billionaires are like our instance of royalty. The people want to see what they’re doing... But I don’t think it’s actually envy.”
- The Information Rules prediction of a tendency toward a single operating system due to indirect network effects was partly right, but operating systems have retreated into the background: people now use five or six OSes daily (iOS, macOS, Linux, Windows, Chrome OS, Android) that interoperate reasonably well with standardized user interfaces, diminishing the OS's importance versus 20 years ago.
“the importance of the operating system has retreated into the background compared to what it looked like 20 years ago.”
- Voting systems already capture some preference intensity informally: people who feel strongly devote their own time, energy, and volunteerism, and recruiting another voter to your side effectively doubles your vote, so intensity gets expressed through effort even without formal mechanisms like quadratic voting.
“If I can recruit another voter to my side, then it’s like I get two votes, my own vote plus the person I recruited. So, there’s a little intensity going on there as well.”
Persistent market gaps remain where business models have not yet cracked sustainable solutions.
- Health is the most promising industry for a young entrepreneur to improve with mechanism design tools because it has many screwed-up incentives that could be improved through new marketplaces and market mechanisms, especially given an aging population driving increased expenditures and demands for efficiency and cost reduction.
“There are all sorts of screwed-up incentives. If you start thinking about it from a mechanism design point of view, you can imagine situations where you could improve those incentives by various kinds of marketplaces, market mechanisms.”
- The survival of small upper-Midwest towns hinges on whether they have a hospital: if a town has a hospital it will probably survive, and if it doesn't, it's in big trouble.
“if they have a hospital, they’ll probably survive. If they don’t have a hospital, they’re in big trouble.”
- Despite clear demand for local news—high school sports, supermarket offers—nobody has yet cracked a sustainable business model for it other than wealthy ownership; people are experimenting with many services and something will likely catch on, but it is surprising it has taken so long.
“this is the thing that surprises me — that nobody seems to be able to crack that nut yet. There’s clearly a demand for local news”