 Thank you so much, Chiara, and thank you to all the organizers for inviting me. I'm really excited to be here to present this paper on entry into two-sided markets shaped by platform-guided search. And just a fact, this is joint work with Kwakau Lee, who's in the audience, and he's just in the year below me at Princeton, which means he'll be a fantastic student on the job market next year that you guys might be interested in. So the other day, I needed to purchase a new water bottle. So I went to the usual suspects, the e-commerce apps on my phone. I typed in the words water bottle, and this is what I saw. Now notice on the left hand side how eBay actually tells you it has more than 280,000 results for water bottles available that you can peruse if you so choose. Now I don't know about you, but certainly even during a pandemic, I feel like I have better things to do with my time than to painstakingly evaluate 280,000 different water bottle options. And so it occurred to me that how these platforms choose to rank these search results, in general their search ranking and recommendation algorithms, they certainly seem to have an outsized influence over what product I eventually end up purchasing. Now I'm hardly the first person to have noticed this. In fact, in a report in 2018 to the Steakless Center at Chicago Booth, a group of economists noted that defaults can direct a consumer to the choice that is most profitable for the platform. And the US Senate investigation of competition and digital markets just a few years after that put a rather finer point on it, noting that Amazon specifically, Amazon can give itself favorable treatment relative to competing sellers. And it has done so through its control over the buy box, which is just another one of these search ranking and recommendation algorithms that I will go over in much more detail later in the talk. However, before we get there, note that a merchant selling on Amazon had a rather different take on the experience, noting that Amazon seemed to rotate the buy box, switching off sellers from time to time and hypothesizing that this might be to keep sellers from getting discouraged and leaving the site. So this is our first indication that entry interacts with this topic of search ranking and recommendation algorithms in a perhaps interesting way. Now I want to do two things today and that is I want to ask and answer two key questions in this context. The first question is what is the effect of search guidance? To answer this question, we will need to examine the interplay between Amazon's need to attract merchants on the one hand and its ability to guide consumer search on the other hand. The key trade-off facing the platform is that between creating competition between merchants thereby lowering prices and attracting consumers and making sure that these merchants are adequately incentivized to join the platform in the first place because a platform without product variety, a platform without many offers is certainly also not going to attract any consumers. What we will find is that empirically this entry channel matters in that if we ignore the entry channel, we estimate a static 51 billion dollar consumer surplus gain, a really staggering consumer surplus gain from the effect of the particular search guidance algorithm we will examine today. However, this major, major gain is actually offset down to a minor gain of only one billion dollars by a reduction in entry that the search guidance algorithm entails essentially because the search guidance makes consumers more price conscious, creates tougher competition on the platform and therefore makes the platform less attractive for entrance. Now the second question I want to ask today is what is the effect of self-preferencing? The regulatory concern here is that firms which participate on their own platforms as sellers make revenue not only from intermediation fees but also from their own sales and these firms may thus be tempted to self-preference that is they may be tempted to preferentially guide consumers towards their own offers. Several theories of harm have been advanced in this context but a key theory is that self-preferencing could act as a barrier to entry, effectively depriving consumers of choice in the long run as merchants opt not to participate in a platform that will just divert demand away from them. Empirically however our model suggests that when we estimate it we find no harm and instead we find a moderate to small five billion dollar consumer surplus gain from the self-preferencing and that is because consumers once we adequately account for the influence of recommendations on their behavior still prefer the platform's offers and that means that consumers would benefit from being shown the platform's offers more frequently or in more prominent spots which is exactly what our model suggests. Furthermore we evaluate this theory that self-preferencing could act as a barrier to entry and we find that actually there is a only a minuscule entry reduction that comes about from self-preferencing which we attribute to the fact that consumers have this strong preference already without the platform sort of influencing their demand and thus merchants already anticipate that they will not be able to outcompete on products that the platform was also selling. There is a major caveat to our analysis here however which is that we don't model Amazon's pricing decision because there's been a lot of controversy over whether the platform is really profit maximizing whether it has some more dynamic long run goal and so we don't want to take a stance on this profit maximization objective. Now if Amazon raises prices by more than 7.8% due to its self-preferencing then we find that our wealth of results flip and then the consumers are actually worse off just because of this Amazon raising its own prices effect not so much because of the effect that Amazon's behavior discourages third party merchants. In fact the third party merchants are more excited when Amazon raises prices they can also raise their prices and that increases their profit margins. Now the agenda for today's talk is going to proceed as follows. I'm going to start by introducing the setting in a bit more detail and providing some descriptive statistics. Then together we will build a novel structural model of intermediation power and by intermediation power I just mean the platform's ability to steer consumers which we will operationalize using the modeling technique of consideration sets at Aguri. In our model firms make an entry decision and then choose prices. The platform chooses some offers to recommend over others and finally consumers make their choice and just to be clear the key innovation of this model is this entry stage that allows us to assess for instance the trade-off that the platform faces between lowering prices by intensifying competition but also lowering entry by intensifying competition. We will take our model and we will estimate it on extensive proprietary data. We will estimate demand and recommendations using a maximum likelihood approach and fixed and unit costs using a simulated method of moments procedure and we will find that Amazon indeed has intermediation power as about 26% of consumers will only consider recommended offers. The platform uses this power to intensify price competition increasing the demand elasticity from around three to around four and to self-preference giving itself a recommendation advantage that's roughly equivalent to pricing 9% cheaper. Now just to be clear when I say self-preference it is a totally valid explanation of these results and I'm sure Amazon would say that they are just giving themselves an advantage because they do better on things that to me are unobservable right like for instance the number of times that when they ship a package there is actually a valid tracking code on it. Things like this are unobservable to me they are real quality differences and in fact this explanation is most consistent with my results which finds that consumers also have a strong preference for Amazon which might be exactly because of these to me unobservable factors. Finally we will use our model to evaluate the effect of self-preferencing and the effect of search guidance yielding the counterfactuals I described on the previous slide just a quick reminder self-preferencing seems to be good for consumers and search guidance is also good for consumers although our assessment of search guidance massively depends on modeling this entry stage. If there are no questions at this point I think it makes most sense for me to delve straight into the setting perfect so that we understand the context a bit more. So the setting I'm going to talk about today is Amazon Marketplace which is a website I think we all know but what we might not know is that third parties co-list here amongst Amazon's own offers and in fact 60% of sales were sales of co-listed offers and they generated some 25 billion dollars in third party profits in 2020 so these are not no small numbers this is a significant part of Amazon Marketplace. I showed you earlier this picture of my search for a water bottle and I do think that the search algorithm is perhaps the main algorithm that influences consumer choices on these platforms however the search algorithm also brings with it some challenges some econometric and these challenges are mostly summarized by the fact that the search algorithm conflates relevance and recommendation and it also has the issue that it gives you access to products that differ and that differ perhaps in ways that are hard for the econometrician to disentangle. So rather than focus on the search algorithm let us focus and set on this algorithm that has been repeatedly mentioned by for instance the US Senate antitrust proceedings to get to this algorithm we will click on one particular product here and when we click on this product we arrive on a product page it looks like this it has big add to cart and buy now buttons but it also has this little note down here that tells you there's actually seven different sellers who are willing to sell you this item now if you click on buy now or if you click on add to cart what you're going to get is amazon choosing the seller in fact amazon has already chosen the seller who owns these buttons on the website it looks like this there is a little box around it and I guess that saves it the moniker the buy box so that is the recommendation algorithm we will look at today just to be super clear what it does is it maps a list of offers for the very same product into the identity of a recommended author and I really like to think of it as econometrician's paradise because product characteristics here are going to cancel we'll still have to deal with offer characteristics but you're choosing between different offers on the exact same product this exact same UPC code level product now we will tackle the setting using novel high frequency high fidelity data that we sourced from a company repricing listings for third parties we have a panel of around 200 000 products that we observe over around 450 days and for these products we have prices recommendations sales ranks and for what I believe maybe the first time sales across a large variety of merchants although that data comes with its own challenges that we will discuss later today we will focus just on all the products we observe in the fashion category around 10 000 products to be perfectly honest the the estimation is very computationally challenging so we wanted to to limit our sample and we wanted to make sure it's sort of homogenous so we picked one category this is the category with the most sales observations so the first question you might have about the status certainly the first question I had is well what actually determines recommendation status and your intuition might be as it was for me that price should really matter a whole lot and so this is what I'm showing you on the left-hand side here is that your intuition is right on the y-axis you see the fraction of offers that are in the buy box the fraction of offers that are recommended and on the x-axis you see the price rank of the offer so the cheapest offer price rank number one is recommended the line share of the time around 56 but you still see that the second cheapest offer the third cheapest offer and so on these are all still recommended relatively frequently so it doesn't seem to be the case that it's just price that matters price certainly seems to have an outside influence but something else also matters what is that something else well here's one of them feedback count so feedback count you can think of as a noisy measure of seller experience every time you sell something there's a certain probability you get an additional review and I'm not looking at the quality of the review right now I'm just looking at the count now if you allow me a causal interpretation for now and I agree that there is some trouble here but we will discuss it later if you allow me a causal interpretation imagine you're starting off as a seller down here with a feedback count around 20 now you hustle over years and you build up to a feedback count of a hundred thousand so you really conduct a lot of sales well then this graph is showing you how much can you raise your price and still be recommended the same fraction of the time and what we can see is that you can raise your price by around two and a half percent so I think what I find interesting here is that these are sort of figures of a plausible magnitude right you would think that you know a 2.5 percent additional margin that is like quite a bit of additional margin but it's also not an unrealistic amount to incentivize people to become experienced sellers now we can look at the table where I've done this for various offer characteristics so for each characteristic moving from the first to the 99th percentile in the data I'm showing you by how much you can raise your price turns out if you use fulfillment by amazon you can raise your price a lot cynics might say this is because amazon makes more money of you but I would counter that it's also true that your fulfillment services have changed a lot your shipping is going to be faster your consumers are going to have a better experience now additionally somewhat metaphysically challenging but if you manage to become amazon you can raise your prices by an additional nine percent the feedback count we just talked about positive feedback also matters a little bit and what's really bad is if you take a long time to ship then you're going to have to cut your prices by a lot now with this basic understanding of what determines recommendation status I think at this point it's going to be most useful for us to delve into the model first before we go to estimation so we have sort of an understanding about how to structure the data and how to think about this in a structured way the model from a bird's eye perspective has four stages there's going to be an entry stage successful entrants have to make a pricing decision this pricing decision will determine recommendation status and jointly the pricing decision and the recommendation status will determine the consumer choice now we will start at the bottom of the model and work our way back up in this modeling section so before we go there however let me just emphasize that this model will play out in parallel across many markets and here a market is going to be a product page so we will assume we will get our all of our consistency by saying this plays out in parallel across these product pages so starting with consumer choice right here at the bottom of this model and we are going to use the most box standard choice model you can think of which is a loaded choice model I'll make fat in 1973 so given a market T consumers are going to face some alternatives call them J the alternatives always include an outside option and the alternatives have some observed characteristics XJT so think of these as things like shipping time or the amount of positive feedback that a seller got there is some unobserved XJT this is because each seller has a name and maybe you associate something with that name that IS an econometrician can't see and then of course there's a price PJT they combine in the usual linear fashion into a mean utility delta JT and then once we shock that mean utility with a consumer specific loaded shock we get our loaded choice model now here is where we depart a little bit from this box standard choice model by introducing consideration sets at a GURI and in more or less the same way as dynastine and al in 2018 did this although adapted to our setting of course so here what we'd note is that one alternative call it jr one alternative on each market T is recommended and the fraction a fraction row of consumers we will call sophisticated sophisticated consumers consider only sorry consider the entirety of the choices available to them so essentially they click on this little link that I showed you earlier to evaluate all offers fraction one minus row of consumers are naive and they will only consider the recommended offer as well as the outside option right and just to be clear I'm a naive consumer so this is not meant as as judgment once we write down this model it becomes clear that market shares are just going to take this very natural form of a convex combination of choice probabilities amongst sophisticated consumers those are standard loaded probabilities and amongst naive consumers also a standard loaded probability but not a naive consumer can only choose something that is being recommended or the outside option so naturally the next thing we have to explain is how do we actually get these choice probabilities here or sorry these recommendation probabilities here now the recommendations again are just going to be a discrete choice problem only this time the platform is solving this good choice problem right and in particular we will use a nested loger model here for each market T the platform is going to choose one offer to recommend note the platform can actually recommend the outside option that is something that does happen in the data and it's actually a motivating piece of evidence in my paper now because that happens we want a nested loger model because we want to exclude the out we want to basically emphasize that the inside options are better substitutes so we want nests that are just the outside option on the one hand and all inside options on the other hand basically from the platform's perspective if I recommend the outside option these naive consumers are never going to buy and so I only want to do that in extreme circumstances beyond there's a question from jack let me just read it so that will be fast I find it very strange to use a standard choice model for the choice of amazon vendors I certainly and I assume most people have no idea who the vendors apart from amazon are and what type of service they offer to put it in another way are there any sophisticated consumers yeah that's an excellent question I I find that there is quite a few sophisticated consumers but I also bias my estimation in the other direction so basically I make identifying assumptions to be super robust against people claiming oh Leon really all consumers are sophisticated because that was sort of the reaction I was expecting from people that you know all these platforms don't actually influence our choices so I you know I find even then that there is still naive consumers so your question goes in the opposite direction is it maybe is everyone a naive consumer now of course I could estimate that that would be a potential result my model could find that the fraction of naive consumers is one so in another way another way to say this is this fraction of role is just going to be estimated so that that should be in the model thank you great question okay so moving on to the pricing stage now we have to start making some more assumptions so we will assume that we have single alternative firms and that means that the subscript j doesn't just index alternatives it also indexes firms and that these firms have constant marginal costs which I would urge you to interpret as wholesale costs these firms are not generally producers they are firms that often you know the typical firm sort of buys stuff on Ali Express or one of the Chinese websites has a delivered either directly to an amazon warehouse or to their own warehouse and then resells it in the US so these firms are going to play a Nash Bertrand pricing game where the type of a firm is going to be given by its marginal cost its various characteristics right like how fast do I share and these unobserved characteristics the firms are going to simultaneously choose prices to solve a profit maximization problem right like in a standard Nash Bertrand setting with two twists the first one is that the platform gets a certain share of the revenue that's where there's five parameters there and in fact it's around 15 percent here but it varies a bit across categories and then the second twist is that these market shares include the effect of recommendations on sales so in particular if I'm a merchant I might say well should I lower my price well amongst sophisticated consumers it's not really going to make a difference but if I lower my price I'm more likely to get the recommendation and if I get the recommendation I am much more likely to make a sale amongst the naive consumers so it's still worth lowering my price that mechanism that entire mechanism is contained in these SJC. Final part of the model is the entry stage here we will assume that each market T has a set of potential entrance script NT and an associated fixed cost that doesn't vary across firms FT we will assume that markets are separable so that no entrant is considering entering on multiple markets at once which is going to be crucial for us to be able to treat these as separate decisions and to get consistency in our estimation. Now comes the major assumption of the paper that arguably I will have to defend a bit and this is an assumption on the information set in particular we will assume that each firm at the entry stage only knows their own marginal cost as well as the fixed cost associated with this market so that assumption really comes in two parts the first part is that firms don't know anything about other firms right so in particular I don't know that Chiara is currently selling these stocks for this price when I'm entering because if I knew that and a new demand I could invert it to her costs and I don't know these costs that sounds a bit crazy on the first read through but on the second read through keep in mind that what Chiara is doing might be irrelevant if she is not going to stick on this product right even though Chiara is currently selling these stocks there's a very high turnover in this data so is she still going to be selling these stocks in three months or on the relevant time horizon where I'm making my entry decisions that seems quite unlikely in this data set so I think this is actually quite a reasonable simplification and you can see why we needed to avoid this classic issue with entry being a game of strategic substitutes and multiple equilibria emerging so this part of the assumption I think yeah yeah there's a there's a question from Luis Cabral am I right that many of the sellers are we selling the same product they purchase from the same source so almost same costs for everybody that's how I interpret the question I think I think that is a good assumption I would focus on the almost the same costs so I think in general this is a good assumption yeah they are probably sourcing from either the same source of very similar sources quite frequently and so I guess an implication so a follow-up is from me so how can we assume that they don't know anything about other competitors costs well they will not not know anything they know that costs are how costs are distributed right so they just don't know the specific realization yeah perfect thank you and then a question from Justin Johnson if you look at websites that give advice to third-party sellers the suggestion is that sellers spend a lot of effort figuring out which markets are likely to be more profitable and have less competition I guess in answer to your assumption about the information that yeah no but that's a that's a great point and that is definitely in my model because that's basically amounts to the market size mattering right they want to know where's the market size high relative to the number of entrants at the moment and that's that's exactly what they do yeah so the second part of this assumption is I think a bit harder to defend and this is the part of the assumption where I say firms don't even know their own vertical characteristics right like how fast they will ship and so on so there's two things I can say here firstly the recommendation algorithm is a bit opaque so they might know their own characteristics they might just not know how it plays out in the recommendation algorithm the second thing I can say is that this is currently necessary for computational reasons to keep this computable and we're working on on relaxing this finally we will assume that firms play symmetric equilibrium and taking all these assumptions together gives us this very nice intuitive equilibrium where firms will enter if and only if their marginal cost draws sufficiently low where sufficiently low of course varies market by market right bigger markets can support more entrance crucially that means that the model allows for selection on cost why is that important well one of the counterfactuals we'd like to run is what if we make the recommendation algorithm more price elastic of course that makes the market more competitive and we will have less entrance but which entrance do we lose that's going to be really important for welfare right do we lose the cheapest most efficient entrance or do we lose a random draw across the cost distribution no we lose the least efficient entrance and this model is able to capture that now with the model out of the way let us go into the empirical application the empirical application I will discuss today I will discuss exclusively recommendations and choice because that's that's where the the interesting estimation happens wholesale and fixed costs are very naturally identified wholesale costs by prices fixed costs by the entry patterns but actually most of the work goes into estimating those because that's a massive simulated method on moments procedure but I think it's less interesting to talk about so recommendation on choice and we can talk about them together to start with the data and market definition are going to be as follows the data are recommendations and third-party sales in the amazon fashion category and a market is going to be a product page p on a day towel so that means that the alternatives jt just to emphasize again are going to be offers for the very same product and the first challenge we have is what I call pooling which is that usually you have like cross sections of the same of different geographical markets where the same products are on sale or time series of the same market over time but here we really have different markets with different products on sale and we need to pool the information across them to identify our coefficients and that is mostly fine the one place where it's not fine is the price coefficient people react differently to price of different markets and what we find in the data is that basically they care about block price which we emulate with with sort of an exogenous measure of price because we want them the price in game to still have a unique equilibrium and then we need to talk about partial observability this is the really a sort of exciting unique challenge of this data set so I'm really excited that we have demand data from amazon at all but we of course don't actually have demand data sourced from amazon we have it sourced from third party merchants and usually on each product on amazon that we have data on there is one third party merchant for which we have the sales data and then we have the price data for all the merchants so that means we have sales for exactly one offer in each market turns out that under this loget a structure we're still identified as long as we know the total market size and what we do for the total market size is the classic trick using the sales rank and shovel year and gold space observation that sales are essentially power law distributed that is that challenge the final challenge is going to be that this is high frequency disaggregate demand data so the berry 94 lock transformation is going to yield severe zero share buyers right on most days most merchants see zero sales so what could we do we could aggregate aggregation across products doesn't really make sense here aggregation over time is going to yield measurement error and attenuation bias because prices move over time essentially we'd be smoothing over exactly the variation you want to use to identify these coefficients so instead we're going to proceed by maximum likelihood maximum likelihood means that warning belts are probably going off because of endogeneity i'm going to have to do this slide real quick because i'm running out of time but basically what i do to deal with endogeneity is that i am going to assume that unobservable offer quality is time and variant so i'm going to allow endogeneity at the offer level as long as it's time and variant and then i verify that for the part of offer quality that i can observe the observable quality really is time and variant and the rest of the slide i'm going to skip my apologies so with that we will have enough time to do the counterfactuals which is really the exciting part right now we've eaten our vegetables now dessert is on the table so the first counterfactual that i want to do is the effect of self-preferencing how do outcomes change due to self-preferencing and i will do all my counterfactuals in three stages i will look at the short run where prices and entry decisions stay fixed at the medium run the prices are allowed to vary but entry is still fixed and in the long run prices and entry are both allowed to vary now two caveats the first one for both this and the next counterfactual you see some big numbers here that's because i'm cheating i multiplied the welfare numbers for the 10 000 markets where i compute welfare for and i scale them up to the entirety of amazon marketplace so you can get some feel for what the size of this numbers is please don't take this numbers to be exactly estimated for the entirety of amazon marketplace and the second caveat is that across both counterfactuals i'm going to keep amazon's behavior so when i say prices are allowed to vary i mean third party merchants prices are allowed to vary and the same for entry decisions because of this thing that i mentioned earlier where we don't know whether amazon is really maximizing profits in the short run okay so with all of that out of the way let's look at how outcomes change due to self-preferencing i can do that because i estimate in the model how much more likely amazon's offers are to be recommended controlling for the other things i observe and i can just make them as likely to be recommended as the other offers in the short run i see that self-preferencing raises consumer surplus exclusively amongst the naive consumers what what is going on here well consumers prefer amazon's offers sophisticated consumers they don't care about the recommendations so they don't care about the self-preferencing but naive consumers they have to be helped by these recommendations and basically what you want to do is you want to show them recommendations for offers that they're actually likely to purchase and so that makes them better off in the medium run you might worry that there is going to be a slight price effect because now you're emphasizing something in the recommendation system that's not priced and indeed we find this this rather minuscule effect here and then in the long run we see that there is some entry displacement but the entry displacement is really tiny there is around five to six entrants on these markets on average so basically what this is telling me is our our model is not failing the basics maltest that it predicts the right direction of the of the entry entry must go down but it barely goes down so it seems like there isn't really entry displacement due to self-preferencing then the second counterfactual i want to run is the effect of search guidance so here i evaluate the impact of the algorithm compared to a no recommendation baseline and what i'm going to say about a no recommendation baseline is i'm going to treat it exactly the same as if there was a random recommendation so this is a bit of a reduced form search model here in the background basically what i'm assuming is that these naive consumers they don't suddenly become sophisticated and gain the ability to evaluate all the offers they're just going to stick at their ability to evaluate one offer but now they have to evaluate a random offer so you might say that this is a bit extreme so take it with a little grade of salt but i do think it does well to illustrate sort of the underlying mechanisms and tradeoffs here so in the short run as prices and entry decisions stay fixed we see that the recommendation algorithm creates a vast gain of consumer surplus and note that this really is a trifecta of great outcomes for the platform the consumers and the producers what's going on well you should show consumers things that they want to buy if you show consumers other things they're just not going to buy anything and that's going to be bad for the consumer that's going to be bad for the platform and that's going to be bad for the merchants however in the medium run we get some inclination some idea of what it means to show consumers what they want to buy because it means showing consumers cheap offers we saw that the recommendation algorithm loads really heavily on price in fact the price elasticity of the recommendation algorithm is minus 19 so with a price elasticity of minus 19 we see that the prices are going to fall by around 14 in the medium run so this is really a rather large decrease in prices and of course the consumers love this but the merchants not so much and the platform not so much now in the long run you might have anticipated this already what if prices fall so much is there going to be an entry effect yes there's going to be a large entry effect the presence of the recommendation algorithm reduces the amount of entry on the platform to the tune of around two merchants on an average of five to six merchants that's a rather large reduction so what's going on here is you know if you make consumers a lot more price elastic to this recommendation algorithm then entry is going to be reduced and people usually get more excited about the self-preferencing part of the paper I think this is actually a part of the paper that really excites me because you can really see like a market design question here usually we don't think of consumers preferences as part of the policy that the market designer can influence but here we see how at least the part that matters to the pricing decision therefore the entry decision is something that the platform can influence in online settings so in conclusion then there are three main takeaways that I would like you to go home with today the first one is that we found that defaults matter and intermediation power is real so I skipped a bit too quickly to the estimation part so that we could get to the counterfactuals but what I find in the paper is that 26 percent of consumers will only consider the recommended offer and that is a finding after I stack the deck is sort of in against finding that naive consumers exist so if anything it should be higher when I talk to industry sources they say oh it should it should be quite a bit higher so we find that defaults matter intermediation power is real so that means careful monitoring of algorithms shaping market outcomes is that will remain crucial going forward if you're a business note that the short run and the long run impact when it comes to recommendations are extremely different right I showed you in the short run it's great you should just make them super price elastic everyone benefits but in the long run there's a real entry trade-off that you have to consider and even in the medium run these price elasticity's are redistributive so that means you should reduce your aligns on short run AB tests and try to supplement them with models or long run tests finally if you're a regulator note that self-preferencing can be welfare enhancing that's what we found here in this particular setting with this particular data I don't want to overemphasize that I'm certainly not saying it's always welfare enhancing what I'm saying is a much simpler point that is really rather obvious which is consumer preferences matter when consumers have the same preference for a platform's own office that the platform does then the platform's preference for this office and a recommendation algorithm raises consumer welfare and that means that this really this issue of self-preferencing that we're trying to sometimes address sort of by the stroke of law it's really an issue where we have to decide on a case by case basis using the data that is available and it's an empirical question whether banning will increase or decrease welfare in future research I'm extremely interested in looking at paying for recommendations what other people call sponsored search and I would call sponsored search guidance I think it adds an interesting complexity to the situation which I'm very grateful for the fact that with the buy box algorithm I didn't have to worry about this but in other settings I think that could yield even worse outcomes as merchants essentially compete not on price on lowering prices but on upping their bids for the consumer's attention thereby much decreasing the efficiency of the market thank you so much and I guess I'll turn it over to Mike for the discussion awesome thank you Mike take it away great thanks a lot for having me I'm really excited to let's see the work and presentation form and also discuss it as you can probably tell from the presentation or guess the paper is also very well written and so I think it's a really nice read for everyone it kind of shines through in two ways first it's extremely clear about the assumptions in the model kind of refreshingly so and very clear about which ones matter and which ones might be harder to digest and then second it really develops a theory of harm around entry I was admittedly a little bit skeptical to begin with my experience in platform markets online platform markets is a little bit more about ebay where you have a proliferation of products and it's kind of hard to believe that the marginal products are that valuable or the marginal listings I should say I think Leon and co-author pretty convincingly established in the paper that entry could be a big deal here they're not that many products to begin with and so it makes sense to investigate it so I think that that's really nice I'm also I mean this is just a central question not just in platform economics but probably the focus of a lot of antitrust economics at the moment so I think it's a really important question to answer and so I think a lot of people have been trying to answer this question and it's been really hard to get good data and so I think the data here is not perfect but Leon kind of highlighted where there are really improvements over what else you could find and in fact often thinking about how to evaluate these platforms and potentially regulate them in the future you know they own the data Leon's trying to estimate the recommendation model which is very much proprietary information and kind of the central part of their business and so trying to infer these things as hard and I kind of congratulate the authors on how much progress they're able to make I'm especially impressed by the considerations that model I'm often a bit skeptical of papers that are imposing pretty strong econometric assumptions to identify consideration sets as opposed to obscene them in the data Leon doesn't see them directly in the data but given the institutional environment and the data he does have I think it's quite excuse me convincing where they arrive I want to just in my few minutes bring up a few thoughts about on kind of directions they could take or or just how to think about this research you know in terms of thinking about what self-preferencing is I think you know other researchers have had a slightly different definition which is not whether Amazon or the platform gives a preference to their own products in the recommendation algorithm whether that preference exceeds kind of the demand difference condition on being shown both products and so I mean Leon makes the point that a reason why this is beneficial the self-preferencing is because it's steering people towards the products they seem to demand I'd be curious whether their steer is there more or less steering than demand might warrant and that might be a more kind of direct antitrust question given I mean given the results I don't think it'll change then the conclusions so I think it's a bit more of an academic exercise but in terms of interacting with the antitrust law that's where I would go excuse me I also wanted to you know bring up in the bigger counterfactual which is let's change the whole algorithm entry is going to matter a lot and so I think there are a few assumptions built into the model that's um that are working against this as Leon mentioned but there are also a few that are working maybe towards entry mattering a lot and we might think you know we want to investigate further um the first Leon was very upfront about but it was um the marginal entrance may not in an ex ante sense may not have lower quality because entrants don't know their qualities when they enter and so I would encourage and it sounds like the authors are working on this um that seems like a really valuable path to take in and if it requires having fewer products I would probably go for it and I appreciate that um you're not doing one product like I did in my paper but um I think you could do fewer products maybe to save on computation but add some more potential selection on quality um otherwise you know the marginal entrance is going to usually be a reasonably high quality type um this second thing is it was kind of related to the first question that Jacques asked which is the demand model so um it's a flat loge um so the you know the next product's probably pretty valuable and we in part infer that because we see kind of noisy choices in the data we see products that aren't in the buy box still have decent market share so we interpret that's okay first off they're a decent number of naive consumers or no sorry sophisticated consumers because they reach these products not in the buy box and second they value those choices and I think that second part's kind of a strong assumption about is this are people choosing things and really they want that different product or was it a pretty close call and I don't really have a better sense of how to infer it from the data but I do think interpreting the loge at choices in the welfare framework is you know driving the results a lot especially in the bigger counterfactual where you remove all of the search um the last thing I just wanted to bring up at the very end was I personally have found entry models are pretty challenging to fit across lots of different markets in terms of fit um and so you know the in data fit here is the in sample fit here um it's hard to pick up why does this market have 12 and why does this other market have two entrants um and so kind of the conometric predictions usually much more closer to the mean and that's that's what the authors are finding it's not obvious to me it's going to lead them to make incorrect conclusions about the change in entry in response to these policies I found that often kind of matching the cross sectional distribution of entry across markets doesn't always give give great fit on on treatment effects when you do see them um but uh kind of the lack of in sample fit is something I would think about a bit in terms of about how much does the model have to explain entry across markets um so with that I'm kind of curious what other people think in terms of the Q&A