Menu Close

Bezos’s ? Email Mechanism

If you want to know what a man’s like, take a good look at how he treats his inferiors, not his equals.

– Sirius Black, Harry Potter and the Goblet of Fire.

I vividly remember my first (of several) ? Jeff B. escalation emails. Sebastian Gunningham was still a “Section 16 officer”, and Amazon had only one CEO. I literally had no idea what I was supposed to do – surprisingly there was (and still is) no training manual for this kind of thing. After asking around, I figured out that I had 24 hours (the time to respond has since then been extended to 48 hours) to come up with a response that indicated a description of the issue, the root cause, any immediate fixes, and long-term actions that address the issue. The first Director in your chain of command emails Jeff your response and marks you and Amazon’s senior leadership team in the email. It can be unnerving but is quickly forgotten if the issue is addressed, or if there is a plan to fix.

Yes, this is a absolutely a REAL thing. I’m not kidding. Jeff literally forwards customer emails with a “?” to his senior leaders. Here are two examples (with the exact email redacted):

In a recent interview with Ken Hersh during the Forum on Leadership at the Bush Center, Jeff described this mechanism. While the entire interview (below) is a worthwhile watch, he describes the ? email mechanism from 22m:35s.

Surprisingly, the response to his disclosure about sending ? emails has been mixed. Several are in awe like this one:

Yet, a few like the one below have rightly pointed out that although glorifying executives who invoke fear in underlings is part and parcel of Corporate America folklore, it should be questioned, like this one:


So what is it then? Is this a mechanism that works, or is this a questionable mechanism that deserves its own ? email?

Well, like most things in life, the truth is somewhere in the middle. I am in favor of using anecdotes to augment data in order to ascertain if systems are running as designed or can be improved further. Allow me to demonstrate using an example.

This example is an automatic email sent out by Amazon’s vendor systems to encourage vendors to spend on enhanced marketing content to improve their product sales (and Amazon’s bottomline). And, it’s darn obvious that this particular email is worthless to the vendor – they would’ve spent $0.36 in just reading the email.

Now, the metrics for this email automation might look just fine. It is unlikely for a handful of low value emails (like $0.36) to affect the open rate, click-through rate, action rate, and opt-out rate, and as a result the product manager may not even notice the need for filters to filter out emails that do not add a significant improvement to vendor’s sales. These vendors are unlikely to opt out because the other emails they receive might be valuable.

In such a case, it is clear that a ? email anecdote can be really powerful in making a small change to the email algorithm – if the product sales improvement is less than a threshold dollar value, or less than a percentage change, do not send that vendor an email. Of course, the team should have caught this and built these filters before launching the automated email product, but sometimes when you are moving fast and launching products at scale, you may miss out edge cases, and anecdotes are more powerful than data in catching those.

Sure, this is only an anecdote and I have not provided any empirical data that backs up why Bezos’s ? email works. But that’s the point isn’t it? Sometimes the anecdotes are powerful enough for you to make a decision. Make no mistake, a not-so-great leader can use the ? email mechanism to belittle employees for their mistakes, or simply to ask wtf? In fact, one of my peers once sent a ? email to his manager as a response to a request from the manager. Let’s just say, the manager sent more than just a request in the next email – he clearly wasn’t thrilled. Taunting employees, or making them feel ashamed of a mistake is not the intent of Jeff’s ? email mechanism at Amazon. The intent is to seek product improvement opportunities that data sometimes fail to catch.

In the end, though, whether you forward such customer emails with just a ? or take a few more seconds to write, “Please look into this.”, or ask “What do you think about this?”, that’s a choice you have to make. I know what I would do. Do you?


Note: This blog post does not contain confidential Amazon information; these are my personal views and does not represent the views of Amazon or its management.

Diplomat Dara and Facebook’s Hell Week

Over the last couple of weeks, two big tech companies in Silicon Valley – Uber and Facebook – have faced major crises. Here’s a sampling of headlines on major news organizations about these crises.

On Uber’s self-driving car accident,

The Washington Post:

Self-driving Uber vehicle strikes and kills pedestrian.

The New York Times:

Self-Driving Uber Car Kills Pedestrian in Arizona, Where Robots Roam.

Fox News was less judge, jury, and executioner, instead attributing the statement to the police:

Self-driving Uber car kills Arizona pedestrian, police say.

Uber’s self-driving car | Image Source: Uber.com, ATG

And on Facebook’s Cambridge Analytica “data breach” controversy,

The Washington Post:

Everything you need to know about the Cambridge Analytica-Facebook debacle.

The New York Times:

Facebook’s Role in Data Misuse Sets Off Storms on Two Continents.

The Guardian:

Facebook’s week of shame: the Cambridge Analytica fallout.

The Economist:

The Facebook scandal could change politics as well as the internet.

The Economist’s Cover on Facebook | Image Source: (Ironic, but) The Economist’s Facebook Page

In both cases, if you don’t read past the fear mongering headlines (kills pedestrian, week of shame, data misuse, scandal), you would bet that both these corporations are pure evil. However, on closer examination, it will be clear that the crux of the Uber story is this: on a not-so-well-lit, multi-lane corridor where cars speed at 45 mph, a pedestrian crossing the road outside the crosswalk (walking her bicycle as she crossed) was fatally struck by an Uber self-driving vehicle with an in-vehicle operator (meant to take over the self-driving car during emergencies).

The Facebook story on the other hand is a little bit more complicated: In 2010, Facebook, pursing a platform strategy, provided access to their graph data to any developer willing to build software applications on top of Facebook (think Candy Crush!).1 Not only did they provide developers access to their graph data, they also announced this access publicly at their F8 conference and several news organizations (many of who are now gunning after Mark Zuckerberg and Facebook) reported exactly what data was being shared and why.  To access this data, a developer had to seek permission only from one user, and the developer would then get access to all the users’ friends’ info as well. This was originally intended to make your website experience better. For example, if a bookseller wanted to let you know that your Facebook friend likes a book, they had to know (1) who your Facebook friends are and, and (2) what books they like. And you have to give them access to do so (since your friends have already “told you” via their listing on their profiles what books they like. A question we should all ask is – who owns my data that friends of mine have access to (my phone number, my birthday, my likes, my dislikes), and can friends of mine share it with others for money or otherwise? Your answer to this question will determine whether you think there was an actual data scandal in this whole Facebook-Kogan-Cambridge Analytica episode. This post provides a well-reasoned, logical view of why this scandal is no scandal at all. Although Facebook prohibited developers from passing along or selling this information, there was and is no way to enforce it once the data leaves Facebook’s servers. I am unsure why Facebook users are alarmed now, in 2018, about the fact that developers had since 2010 (1) access to your information if your friends permitted them to get it, and (2) developers could easily share this information with others despite Facebook prohibiting them from doing so. Dick Costolo’s tweet perfectly sums this up.

Long story short, a researcher (Prof. Kogan) developed an app that asked users to share their info and who their friends are and what their interests are, and then passed this data along to Cambridge Analytica, who subsequently used this data to influence (some say manipulate) the US Presidential elections.2 One thing is very clear: every Facebook user (yours truly included) accepted the terms and conditions (who reads them anyway?) of using Facebook that clearly stated that Facebook stores info about my interests, my likes, my posts, who my friends are etc. and this info can be provided to 3rd parties.

Now, make no mistake, whether Facebook giving developers access to their data graph was a sound strategy is a completely different question. Posts from James Allworth and Ben Thompson are must reads about Facebook’s platform strategy. Facebook wanted to be a platform, and in order to attract developers to build on top of Facebook, they shared their graph data with them. It’s not a new strategy; in order to get developers on board, some companies throw money at them (hello there, Microsoft!) and some like Facebook, share(d) access to their data.

By now, it should be plainly obvious that I believe Uber and Facebook have gotten the short end of the stick from traditional media. In Uber’s case, a more truthful headline would have read – “Pedestrian in not-well-lit street hit by self-driving Uber” or “Pedestrian crossing outside crosswalk hit by self-driving Uber vehicle.”3 And in Facebook’s case, a more neutral headline would have been – “Facebook developer partner sells facebook graph data” or perhaps even “Facebook’s decision to open graph data to developers backfires”. One thing that’s clear from the contrast between my made-up headlines and the actual headlines is that people are warier about tech than ever before, so one of the easiest way to get your clicks and views is through click-bait headlines about big bad tech and polarizing political news. Clearly, my made-up headlines above are too bland to garner enough clicks and views, and would’ve have made the cut at any major news organization.

I digress. The point of this post isn’t to debate whether Facebook erred in sharing data with Prof. Kogan or Cambridge Analytica. This post is to highlight the difference between how Uber’s CEO Dara Khosrowshahi, and Facebook CEO, Mark Zuckerberg, managed their PR nightmares.

Uber could have very well argued that the pedestrian crossed outside of the crosswalk, or that the street wasn’t well lit, or that the operator wasn’t paying attention, or any of the numerous externalities that may have caused this incident. Yet their CEO, Dara Khosrowshahi and Uber’s communications team did not blame anyone else; they issued statements expressing their thoughts with the victim and preemptively grounded all their self-driving cars without waiting for a trial by the media and Twitterati, and well before any government regulation forces them to take these cars off the street. I was initially skeptical when Dara Khosrowshahi took over from a product-focused leader like Travis, but so far he’s shown that he’s a skilled diplomat: he’s shown that not only is he a master deal maker (more on that in another post), he also knows exactly when to be combative, and when to just lay low.

On the other hand, after the so-called Facebook-Cambridge Analytica scandal, Mark Zuckerberg did not show up for several days. And when he finally showed up, he went on a PR blitz through interviews with all major news organizations and newspaper ads (with an apology nevertheless). Yet, Facebook continued to play the combative blame game; for example, in their first post about this episode they blamed Prof. Kogan and Cambridge Analytica.

Aleksandr Kogan requested and gained access to information from users who chose to sign up to his app, and everyone involved gave their consent.

Although Kogan gained access to this information in a legitimate way and through the proper channels that governed all developers on Facebook at that time, he did not subsequently abide by our rules.

Cambridge Analytica, Kogan and Wylie all certified to us that they destroyed the data. Contrary to the certifications we were given, not all data was deleted.

They started the next post by playing victim right at the beginning:

What happened with Cambridge Analytica was a breach of Facebook’s trust.

After news reports of this episode indicated how people can check what data Facebook had about them, people started downloading their data and noticed their likes, their messages to people, and calls and texts (but not the actual content of these calls or texts) in Android devices, prompting Facebook to post one more update:

You may have seen some recent reports that Facebook has been logging people’s call and SMS (text) history without their permission. This is not the case.

In each update, they just made things worse by blaming others (even if that may be the truth). Exactly opposite of what Diplomat Dara did. He did not say Uber is to blame, but he also didn’t lay the blame on the pedestrian or anyone else.

M. G. Siegler puts it best when he calls Facebook, the “Foot-in-Mouthbook”.

To make matters worse, Facebook clearly has a bad case of foot-in-mouth disease. Whenever they try to respond to a situation, they just exacerbate the issue.

As an engineer and tech worker, I’ve always wondered why some people think with their hearts and not with their head. Why can’t they see the logic that a car driven by the human would have also likely hit the pedestrian in the Uber accident, or that the number of miles driven without an accident is several times higher for automated cars? Why can’t they see that if nearly 300,000 clicked “Ok, grant permissions” to Prof. Kogan’s app and shared their information and their friends brief information to the app developer, it’s not Facebook’s fault. Yet, sometimes, even when you’re not in the wrong, even when you wonder why others can’t be logical and see someone else is at fault, it takes a diplomat to skillfully manage public opinion by not blaming others, by not combating public mood.

When faced with crises that you know may not be your fault, it takes both skill and tact to stop yourself from fighting the criticism with logical reasoning and simply empathize with those most affected by the incident. That’s exactly what Dara Khosrowshahi has managed, and for that Mr. Khosrowshahi, hats off to you!4


  1. This graph data is Facebook’s competitive advantage compared to Google’s or Amazon’s data. Google and Amazon know primarily only about you and your searches or your product purchases. Facebook, on the other hand,  not only knows about yourself but also your relationship with other users who also like or dislike something. For example, through their graph data, Facebook can determine that you don’t usually watch action movies, but when you do, you only watch it with your buddies from high school (Facebook thanks you for checking in with your 5 high school buddies to all those action movies last year). So what about this data? Well, if I am selling action movies, I can now not only target regular action movie goers, but also influence you through the exact 5 high school buddies who are most likely to make you watch something. Ever seen those ads on Facebook which say a particular friend likes a page?

  2. Although in my opinion their influence or manipulation is still up for debate given how miserably they failed to help their first client, Senator Ted Cruz.

  3. Death is tragic and this made-up headline is by no means trying to downplay the tragic nature of the incident.

  4. I’m not suggesting Uber is not at all fault here. I’m only suggesting that even though there are other factors Uber could blame, they have chosen not to do so.

    Also, the two crises are fundamentally different in nature, and so not ideal to compare. But I make the comparison only to suggest that Facebook and Mark Zuckerberg could learn a thing or two from Uber’s new CEO regarding diplomacy and laying low during an onslaught.

The Illusion of a Profitable Amazon Prime Now + Prime Fresh

It works like magic! That’s what Stephenie Landry, the Amazon VP in charge of Prime Now, intended this service to be. Nicknamed “Amazon Magic” by her team, it indeed does work like magic for customers. Order one of the over 25K products in any of the several cities where it is available, and your order will be delivered in 2-hours for free, and in 1-hour for a flat fee of $5.99. The only caveat – you have to subscribe to the $99/year Amazon Prime membership which offers among others “free” 2-day shipping on hundreds of millions of products, millions of video titles, and several other benefits.

It also works like magic for Amazon shareholders, by pulling a Houdini on them and never returning their invested capital. To learn how, let’s first study Prime Now’s business model.

Prime Now – The Business Model

Like most of Amazon’s other businesses, Prime Now was a bet on scale – if your variable cost per unit is below your revenue per unit (i.e. if you are contribution profit positive), at the right scale, your fixed costs can get allocated well enough for you to turn an overall profit (or more importantly positive free cash flow). And similar to Amazon’s core Ecommerce business (that I described in my Amazon Marketplace Profitability Analysis post earlier) – Prime Now has two business models:

  1. A Retail model (denoted by store “Amazon” on the Prime Now app) – where Amazon takes ownership of the product from the manufacturer, stocks the product in urban Prime Now fulfillment centers (FC), sets the price to the end consumer, pays for inbound (from manufacturers to Amazon FCs) and outbound (from FCs to customers) shipping, and is the seller of record. As of Jan-2018, customers who place orders worth less than $35 pay $4.99 delivery fee for 2-hour delivery (no 1-hour delivery option), and those who place orders worth $35 and above get free 2-hour delivery (or pay an extra $4.99 to get 1-hour delivery).
  2. A Marketplace model (denoted by other stores such as “PCC”, “Bartell Drugs”, “New Seasons”, or “Restaurants”) – where Amazon allows third-party stores and restaurants to sell to Amazon customers through the Prime Now App, taking a cut of each sale (aka referral fee/rake that is rumored to be about 20-30% of total sales). Unlike the MFN model of Amazon’s core Ecommerce business (and similar to the FBA model), Amazon pays for outbound (from restaurants’ or sellers’ locations to the customer). As of Jan-2018, customers who place orders worth less than $60 (for some stores $50) pay $5.99 delivery fee for 2-hour delivery (no 1-hour delivery option), and those who place orders worth $60 and above get free 2-hour delivery (or pay an extra $5.99 to get 1-hour delivery).

Let’s forget all the 1-hour and 2-hour fee differences and complications, and perform a quick gut-check to comprehend the unit economics of Prime Now’s business.

Start with the variable costs first. If you’ve ever worked in physical retail, you know that the average contribution profit in the kind of products that Prime Now sells in the Retail model is about 10% of topline without delivery costs. In Prime Now’s case, the physical store still exists (the urban fulfillment centers in case of the Retail model above, and the actual physical store such as Bartell’s in case of the Marketplace model above); yet they incur additional variable costs – from fulfillment to delivery – that eat into this contribution profit.  I performed a very scientific study of asking every Prime Now delivery associate I saw how many orders they had delivered in the last hour. Across 25 associates, the average number of orders was only 3.  To recover and breakeven on fixed costs (such as software development, or cost of building Prime Now urban FCs), Amazon needs to make substantially higher number of minimum value orders per driver per hour. Simply scaling number of orders don’t help profitability; the orders have to be from the same apartment or neighborhood in order for the drivers to increase the number of deliveries per hour.

To be fair, Prime Now can be sustainable if the only sold profitable items (e.g. if they only sold $1000 smartphones) or if they increase the prices of products sold (e.g. if an apple from Prime Now is suddenly priced at $35), then sure, the contribution profit will not be 10%. Prime Now can also be profitable if they increase the minimum order value from $35 to let’s say $70. That’s exactly what Amazon has done with Prime Now’s Marketplace model. As of Jan-2018, they have increased the minimum order from $35 to $60 for marketplace orders (from Bartell or others).1 Customers who place orders worth less than $60 (for some stores $50) pay $5.99 delivery fee for 2-hour delivery (no 1-hour delivery option), and those who place orders worth $60 and above get free 2-hour delivery (or pay an extra $5.99 to get 1-hour delivery). Arguably some orders are much higher than the minimum orders, bringing down the number of orders needed to breakeven; however, remember, we haven’t even got to fixed costs component yet. Even if they are slightly contribution profit positive, Amazon is looking at a significant time to recover fixed costs.

While using the contractor-model of Amazon Flex delivery drivers has helped offload the fixed salary component of hired drivers to a variable component of on-demand drivers, Amazon still needs to be contribution profit positive for Prime Now to be sustainable. All this would be okay if you could bring down variable costs through economies of scale in the long-term. Yet, for that to happen, you need to make more number of orders per hour per driver – only possible in extremely dense urban locations such as New York or London.

The other important argument (that any Amazonian will definitely put forth) is the impact on the Prime Flywheel, i.e. downstream impact from new Prime customer acquisition and existing Prime customer retention. This is similar to the argument that one should view Prime Now through the CAC and LTV lens. Since Prime member data isn’t publicly available, it is unclear how much impact Prime Now has on the Prime Flywheel; but unless every dollar that Prime Now loses is made up by an additional dollar of profit from spend on the overall Amazon platform by Prime customers, Prime Now will continue to be unsustainable.

+ Prime Fresh

So, what about Amazon Fresh, the $180 per year grocery delivery service, then? Well, tack on the variable costs of specialized delivery bags, Amazon-owned fresh trucks, Amazon-employed delivery drivers, higher produce shrinkage and the additional fixed costs of refrigerated fresh fulfillment centers, and you get the picture. Even with the additional ~$15 per month from each customer and a minimum order of $40, Prime Fresh is unlikely to cover the additional costs incurred (compared to Prime Now). The bottom-line, if Prime Now is unsustainable, Prime Fresh is UNSUSTAINABLE in all caps.

+ Whole Foods

Amazon now has Whole Foods, that should change things for Prime Now right? Amazon has already announced Prime Now deliveries from Whole Foods, and in some ways yes, the fixed cost of Urban FCs and the fulfillment portion of the “nominal fulfillment and delivery” costs will now allocated to the physical store business’s P&L (which in Whole Food’s case is profitable). But that said, the variable costs of Prime Now’s delivery will still make it unsustainable at current prices. Again, unless every dollar that Prime Now loses is made up by an additional dollar of profit from spend by Prime customers on the overall Amazon platform – whether at Whole Foods or elsewhere on the platform, Prime Now will continue to be unsustainable.

The Refresh (Prime Now + Prime Fresh + Whole Foods)

Given that there are rumors of a consolidation between Prime Fresh and Prime now already underway, I wouldn’t be surprised if Amazon announces a complete shutdown of the Amazon Fresh business within the next one year. As for Prime Now, it will live to fight another day. That other day, though, will not look like today. In the current “delivered to the home” model, Prime Now + Whole Foods will need to make two major changes to their pricing – (1) increase the minimum order to $80 to increase the per order contribution profit, and (2) institute delivery fees of $10 (to cover for admin, sales & marketing, and fulfillment costs) for orders below $80. Increasing prices to these levels, however, will quell customer demand, and while I don’t have any data on price elasticity for Prime now customers, my intuition says that customer demand at these prices will be insufficient to sustain the Prime Now business.

For all these reasons above, in the long-term (perhaps 5 years or so from now), Prime Now will eventually go the way of several other experiments at the World’s Best Place to Fail.


Note: This blog post does not contain confidential Amazon information; these are my personal views and does not represent the views of Amazon or its management.

PS: Unexpectedly, this post briefly trended on the front page of Hacker news and I received some fair criticism and feedback about my analysis. The fair ones: (1) In an earlier version of this post, I used estimates that were are not backed by data. I can access this internal Amazon data on fulfillment costs and delivery costs if I want to but did not, and will not disclose it even if I did since it is confidential Amazon data. (2) I had incorrectly assumed SG&A as a variable cost; it is actually fixed costs, so per order SG&A should decrease with scale. (3) Prime subscription revenue isn’t baked in to this analysis. I only talked briefly about downstream impact from higher prime member spend because I believe that Prime subscription revenue is offset by higher Prime shipping costs. (4) Future tech and scale. If automated cars, drones and delivery robots become a thing (and it will eventually, just a question of when), yes, things will look very different for Prime Now. I also agree that scale (more number of orders per hour by the same driver) will improve profitability.

The point of this post wasn’t to get into the details of Prime Now’s financials; rather it was to suggest that Prime Now, like Shyp and other unprofitable on-demand businesses, will struggle to be sustainable in the long run due to their negative unit economics. Thanks for the feedback.


  1. When Prime Now first launched $20 was the minimum order.

Amazon Org Chart

A few years ago, Manu Cornet, a Google programmer published a now famous set of diagrams depicting the organizational charts of big tech companies that later made its way to the pages of New York Times and recently Microsoft CEO Satya Nadella’s new book Hit Refresh. When I came across the diagrams then, I found it to be spot on in its representation of all the major tech companies, particularly Amazon with its neatly laid out org structure.

Fast forward a few years, and having now worked in multiple Amazon teams, I’ve come to realize that there’s not one Amazon structure. Each org works differently and each team within those orgs works a little differently. In fact, “depends on the team” is the oft-repeated statement to describe anything and everything Amazon, both inside and outside the company. What’s funny is that each team thinks they are Amazon: AWS thinks Amazon = AWS (especially since they supposedly bring all the profit); Amazon Video thinks they embody Amazon’s leadership like none else; if you ask Retail, of course, they think they are Amazon, after all, that’s where it all began. The truth is that none of them truly represent Amazon. Here’s my best attempt to adapt the original Manu Cornet version to depict the org structures of different Amazon teams.

All right, all right, all right, I’m just kidding; it’s not exactly like that.

But joking aside, Amazon, in my view, is structured into several self-contained business units. While officially, Amazon has “only” three CEOs, I believe that each of the other businesses, such as Devices, can grow to have their own CEOs eventually. And businesses within can grow to have their own SVPs (for instance, SVP for Alexa, SVP for Physical Stores and so on). Theoretically the central Finance, HR and other functions may report directly to Jeff B. and support all business units; but for all practical purposes there are separate teams within these central functions to support the disparate business units. Here’s a view of what that looks like.

Goes without saying that there are large businesses within each of these, for e.g. EC2 in AWS, or FBA & Marketplace in Ecommerce. Similarly, each of these follow a single threaded leadership model with self-contained product, software, and s&m teams that allow them to, for the most part, move fast, experiment and build their businesses, eventually charting their own course and creating their own destiny.


Note: This blog post does not contain confidential Amazon information; these are my personal views and does not represent the views of Amazon or its management.

Amazon Marketplace – A Monopazari

Your margin is my opportunity.

Jeff Bezos, Amazon Founder and CEO. 

In 1958, Harvard Professor Malcolm McNair demonstrated that new retailers start typically by inventing a lower cost structure. They then pass on these cost savings to customers in the form of lower prices and attracts customers. As they grow, these new retailers drive volume away from competitors, gaining economies of scale (and increasing the competitors’ cost structure) and allowing them to expand even more. When they capture significant market share and drive competitors out of business, the goal shifts away from attracting new customers to generating profits through higher prices. As they raise prices, they become vulnerable to new lower-cost entrants, starting the cycle anew. Up until Amazon came along, this phenomenon that he coined the “Wheel of Retailing”, had been ably demonstrated by Walmart, who undercut then retailers with fat margins and passed those savings to the customer via lower prices.

So when I recently came across an HBR article where Harvard Professor James Heskett pondered whether Amazon could break the wheel of retailing theory, it got me thinking. When viewed as a retailer (as the article and most of the accompanying comments do), Amazon continues to offer the lowest prices to customers and hence is immune to low priced competitors, at least until another new entrant beats Amazon to an innovation that can offer even more lower prices (or equal prices with better convenience like Amazon did to Walmart). However, when viewed as a marketplace, it’s a completely different story – one that, in my opinion, makes Amazon vulnerable to the wheel of retailing.

Amazon’s Ecommerce Structure

Amazon’s Ecommerce business can be divided into two segments:

  1. Amazon Retail (denoted as shipped and sold by Amazon on Amazon’s website) – where Amazon takes ownership of the product from the manufacturer, typically stocks the product in fulfillment centers (FC) around the country, sets the price to the end consumer, pays for inbound (from manufacturers to Amazon FCs) and outbound (from FCs to customers) shipping, and is the seller of record.
  2. Amazon Marketplace – where Amazon allows third-party sellers or merchants to sell to Amazon customers through Amazon’s website, taking a cut of each sale (known as the referral fee in Amazon parlance, and coined rake by VC Bill Gurley). Within Marketplace, there are two models:
    1. Merchant Fulfilled Business aka MFN (denoted as shipped and sold by seller on Amazon’s website) – where the seller stocks the product in their own warehouses, sets the price to the end consumer, ships the product straight to the customer, pays for the cost of shipping, and is the seller of record. Amazon charges the seller a referral fee percentage (typically 15%) calculated on the total sales price, excluding any taxes, and including the item price and delivery or gift wrapping charges.
    2. Fulfillment by Amazon aka FBA (denoted as sold by seller, fulfilled by Amazon on Amazon’s website) – where the seller pays for “storage space and fulfillment services” from Amazon’s fulfilment centers, sets the price to the end consumer, pays for inbound shipping to Amazon FCs (while Amazon pays for outbound shipping from FCs to customers), and is the seller of record. In addition to the FBA fees for storage and fulfillment services, Amazon also charges the referral fee (just as described in #2a above).

These are two distinct business models – Amazon Retail follows what is traditionally called the Reseller or Wholesaler model (with control over pricing and inventory), while Amazon Marketplace is a two-sided marketplace platform (with no control over pricing and inventory). I wrote about platforms from a pricing lens earlier; within Amazon Marketplace, the MFN model falls into the Paid Cross-side Platform model while FBA in my view falls into the Hardware Cross-side Platform model.1

Monopazari Profit Theory

Cross-side platforms without competitors are neither a monopoly (from greek mónos “single”, and polein “to sell”) nor a monopsony (from greek mónos “single”, and opsōnia “to purchase”) since there are many sellers who sell, and there are many buyers who purchase. Instead, when only one or two marketplaces exist for sellers and buyers to interact, those marketplaces are either monopazaris (mónos “single”, and pazári “marketplace”) or oligopazaris (Uber-Lyft, AirBnB-HomeAway, Amazon Marketplace-Walmart Marketplace, Apple-Google appstores, Google-Yelp etc.) – one of the few places with the technology for sellers and buyers to connect, interact and transact.

Go back to Amazon’s Ecommerce structure for a second. Look carefully at Amazon Retail and reflect whether it could be a monopoly. The answer is complicated despite what many liberal news organizations may proclaim. Amazon Retail isn’t a monopoly in my view because (1) they have less than 23% of the US Ecommerce market segment,2 and (2) outside of certain categories such as books, customers have a variety of other retailers (online and offline) such as Target.com, Walmart.com, Costco.com, Macys.com, BestBuy.com and several others to buy from. I’m not saying that they do (although the numbers suggest that they do), but that they could. In fact, I am an Amazon employee and I routinely buy from those websites (primarily when products aren’t available from Amazon Retail such as this).3

In Amazon Marketplace’s case, however, there are very limited other marketplaces – Walmart, eBay, Newegg. These websites have not gained much traction with sellers (and consequently with customers) primarily because they follow almost the exact same referral fee model followed by Amazon. eBay did try a lower referral fee, but in my view, if not for their lack of single detail page structure (multiple listings for the same product) that hurts the customer experience, they would’ve been able to capture more customers from Amazon than they eventually did.4 Amazon Marketplace has developed proprietary marketplace technology including multi-listing matching, single product page buybox logic, restricted products technology, transaction risk/fraud detection tech and so on. More importantly, Amazon Retail’s instock Prime selection, sharp pricing and delivery experience attracts customer traffic from which sellers of any product on the platform benefit. Sellers also benefit, when using FBA, from Amazon’s fulfillment and delivery infrastructure. These are experiences a new entrant cannot easily recreate, and as a result, Amazon exhibits Supernormal Profit tendency (demonstrated in the section below) through its exorbitant referral fees. Note that supernormal profit is profit greater than the opportunity cost rate of return. In other words, normal profit is one that would be deemed by Amazon itself as sufficient to make a marginal investment worthwhile; supernormal profit is margin significantly greater than this normal rate.

Amazon Marketplace Profitability

Since Q1 2017, Amazon has begun to provide more details on financials for Amazon Marketplace, but does not break out revenue from referral fees or profitability of the marketplace business; however, we can estimate revenue and profitability using proxies, some general assumptions and publicly available information.

In a press release in January 2017, Amazon announced that FBA delivered more than 2B units and constituted 55% of all third-party (aka Marketplace) units. Jeff Bezos, the CEO, has also announced in the past that about 50% of all units sold on the platform come from Marketplace sellers. In addition to these, I will use data from Amazon’s Q1 2017 results (page 13) that provide unaudited revenue numbers from third-party services for the preceding four quarters. This third party revenue is a combination of FBA fulfillment fees revenue, FBA referral fees revenue, and MFN referral fees revenue. For ease of analysis, I will try to breakdown the P&L for the MFN portion of Amazon’s Marketplace Business.5

I estimate that, in 2016, Amazon Marketplace’s Merchant Fulfilled business (Marketplace minus FBA, aka MFN) generated $44.6B in GMV and $7.8B in Revenue with an operating profit of $5.3B (operating margin of 69%) for an effective rake of 17% (see Table B). Very few companies generate such operating margins – examples are Mastercard (54% op margin) and Visa Inc. (53% op margin) in the payment processing industry.6 Apple, a company that sells premium and differentiated products, ended the year 2016 with a 28% op margin.7 Alibaba and eBay, who both operate marketplaces not unlike the Amazon Marketplace, ended the year 2016 with 30% and 26% operating margins respectively (Table C). Alibaba and eBay generated $22.9B and $8.9B in revenue over $547B and $84B in GMS, for an effective rake percentage of 4% and 11% respectively (Alibaba primarily in the form of search ads on the platform and not a direct rake). Based on Alibaba and eBay’s financial statements, I estimate Amazon’s breakeven referral fee to be between 5-8%, and at their current 17% effective rake, ~10 percentage points of referral fee is Amazon’s supernormal profit.8

 Threat of New Entrants

Another way to look at whether Amazon Marketplace is vulnerable to the “Wheel of Retailing” is to evaluate the opportunity for new entrants or competitors to reduce their referral fee for sellers, who can then pass off the lower fees to customers in the form of lower prices. In my view, an online marketplace has medium to low barriers to entry (while network effects is a strong barrier, there is no other significant barrier). A new entrant (or a competitor such as Walmart) could offer sellers a ~7% referral fee, customers an 11% discount and still make a profit. Table D demonstrates an example of how a competitor can offer higher profits back to consumers in the form of lower prices. Assuming the product is price elastic, this new entrant can attract sellers by offering higher profits to the seller even with a lower price to the customer.

While Amazon’s Marketplace’s parity clause (S-4 Parity with Your Sales Channels) in existing seller agreements mandates that sellers must “ensure that the price of an item you list on Amazon.com are at or below the price at which you offer the item via any other online channel,” it is difficult to enforce at scale – particularly when a new entrant provides an alternative marketplace that offers lower fee and one-click export and upload of Amazon listings into their website listings. Moreover, when used to deprive customers of lower prices, this clause may be scrutinized by the US FTC and DOJ under antitrust laws.9

In theory, it is possible for Amazon Marketplace to maintain these supernormal profits in the short run, wait for other players to signal willingness to enter, and then compete by matching their prices (in this case, referral fee). Moreover, a new entrant will not only find it difficult to build all the tools and systems that Amazon has painstakingly built over the last two decades, they will also need significant capital for investment in growth (because the lower rake will not provide sufficient cashflow for big initial capital expenditures). However, in the long run, these supernormal profits should eventually attract new entrants and/or competitors (who will capture market share by charging lower referral fee and passing off savings to the customer) that will erode profitability until only normal profit is available, thus reaching a long run equilibrium stage.

 Stratelogical View

In essence, Amazon Marketplace is a high fixed cost (primarily in software development) business where the marginal cost of one additional  sale of an item is minimal. Yet, Amazon charges its sellers several percentage points for access to its marketplace platform. Amazon Marketplace could reduce their marketplace referral fee to 7.5% and still maintain an operating margin of 28%10 that is more in line with other Marketplace businesses such as eBay and Alibaba. If Amazon doesn’t do so, perhaps Walmart (and/or other marketplaces) could lend them a helping hand by reducing their own fees and capturing first the sellers, and then customers from Amazon.


Note: This blog post does not contain confidential Amazon information; these are my personal views and does not represent the views of Amazon or its management.


  1. At first glance, it may seem that customers don’t care about Amazon’s physical fulfillment footprint – if the products are available elsewhere, why would they? In reality though, customers would pick based on the tradeoff between cost and faster delivery. FBA enables faster delivery for sellers at a lower cost than they could do themselves because Amazon leverages their scale to pass off better costs to sellers. Whether a competing platform with lower rake would be able to capture FBA sellers is debatable because it depends on the tradeoff between cost and delivery speed of products.

  2. In Q3 2017, Amazon announced revenue from Ecommerce to be $26.4B (page 13, Q1 Results) and US Dept. of Commerce announced the total US Ecommerce sales to be $115.3B.

  3. Yes, surprising, but Apple and Amazon has always had a frosty relationship; Apple has never allowed Amazon Retail to sell the iPhone even though they have allowed Walmart, Target and several others to do so. All iPhones available on Amazon are almost always sold by third-parties.

  4. Jeff Bezos, the CEO, has spoken about single detail page several times including both 2015 and 2016 shareholder letters.

  5. FBA has additional costs involved in storing, fulfilling, shipping and C-returns that are harder to tease out from Amazon’s financial statements. To estimate MFN revenue, I made two major assumptions – (1) that Average Selling Price (ASP) is the same across Amazon Retail, FBA, MFN, and (2) FBA fulfillment fees constituted 25% of marketplace revenue; I sampled a handful of Amazon products using FBA calculator to arrive at this estimate.

  6. These are also Paid Cross-side Platforms. and would’ve been vulnerable to lower priced competitors if not for their (what I think is an anti-competitive) clause that forced merchants to set the same retail price irrespective of whether customers paid by credit card or other forms of payment. More on that in another post.

  7. Apple Appstore is an exception; Apple charges a 30% rake for their app store because of their hardware differentiation – iPhone buyers are unlikely to go buy an Android phone even if the apps on Google Appstore are 50% cheaper; they would if they found Android hardware better or cheaper. In Amazon MFN’s case, however, customers will go to a platform that offers them cheaper effective prices because there is no hardware differentiation.

  8. I agree that absolute dollar profit and free cash flow trump percentage margins. For example, a profit margin of 1% on a revenue of $100B is almost always better than 70% profit on a revenue of $1B if you’re getting paid at the same time, and getting paid earlier is always better than later if you’re making the same money (especially positive cash flow cycles, where you get paid before your costs are paid out). In fact, Jeff Bezos puts it perfectly in this HBR interview when he says “Percentage margins are not one of the things we are seeking to optimize. It’s the absolute dollar-free cash flow per share that you want to maximize, and if you can do that by lowering margins, we would do that.” The reason I also like to look at percentage margins in addition to cash flow is that higher percentage margins indicate opportunity to undercut and pass savings to customers to capture market share – yes, exactly what Prof. McNair theorized all those years ago.

  9. This controversial clause has already been removed in EU after EU investigations: official German FTC Press Release, and coverage from Haerting DE news. US FTC and DOJ prohibit business practices that deprive consumers of the benefits of competition, resulting in higher price.

  10. This assumes sellers pass all savings from lower referral fee to customers.

Platforms – The Pricing View

In construction, a platform is something that lifts you up and on which others can stand. The same is true in business. By building a digital platform, other businesses can easily connect their business with yours, build products and services on top of it, and co-create value.

-Harvard Business Review, Three Elements of a Successful Platform

Two-sided Markets, Platforms, Marketplaces, or Aggregators – whatever you call them, they are all the rage. Facebook, Uber, AirBnB, Yelp and other breakout platform companies of the last decade have popularized this business model for entrepreneurs and VCs alike. First identified and analyzed in pioneering work by Nobel Laureate Prof. Tirole and research partner Prof. Rochet, “two-sided markets” or “two-sided platforms” refer to businesses that cater to two interdependent groups of customers.1 Since then, several theoretical and empirical papers have emerged, and platform markets has become a very active area of research in economics – primarily in network effects and reputational systems. Work by Prof. Jonathan Levin, who thinks of platforms as intermediaries that bring users together to enable economic or social exchange, and by Prof. Hagiu, who provides a mathematical function for multi-sided platforms to choose between reseller and marketplace business models, are more recent examples of interesting research in this area.

Outside academia, VC Bill Gurley is perhaps the most famous proponent of marketplaces and its potential.2 His 10 factors to evaluate a new marketplace and optimal marketplace pricing strategy blogposts are required reading for anyone interested in platforms/marketplaces. Recently, Stratechery author and blogger Ben Thompson has garnered significant praise for rightly explaining this business model through his  “Aggregation Theory”. In this post, I am going to expand on Bill’s pricing blogpost to classify platform businesses from a pricing perspective and then elucidate why it is important to do so.

When viewed through a pricing lens, there are four kinds of platforms:

Paid Cross-side Platforms – These platforms create value primarily by enabling direct exchanges between its consumers and producers, and benefit from cross-side network effects, i.e. the volume and nature of merchants attract users, and more users attract more merchants (more sellers -> more customers, more customers -> more sellers). Interaction is cross-directional, from sellers to customers and customers to sellers. Customers don’t necessarily interact with other customers (and merchants do not necessarily interact with other merchants), and even if they do (Uber Pool etc.), that communication is not the primary goal for the platform. Examples are eBay, Amazon Marketplace, Alibaba’s Tmall, Uber, Lyft etc. They are called “paid” platforms because merchants are charged a rake (or revenue share) when they sell to customers. These platforms are price elastic with respect to the rake that they can charge merchants because higher rakes are passed on to the end-customer as higher prices.3

Hardware Cross-side Platforms Cross-side platforms or marketplaces built on top of a hardware or physical layer is unique because these platforms do not attract users primarily by the volume and nature of the merchants on their platforms; rather the hardware attracts users, and the users then attract merchants. As a result, these platforms are more price inelastic (compared to other platforms) with respect to the revenue share they charge merchants. Examples include Apple Appstore, Xbox Gamestore, Google Playstore. Even a physical marketplace in a unique location that allows merchants to sell their wares and charges them on a revenue share model falls in this category.

Free Cross-side Platforms – These platforms enable exchanges between its consumers and producers almost always for free, and use an ad-based model (free for users, producers pay for ads) for monetization. They benefit from cross-side network effects. Google, Yelp are examples.

Same-side Platforms – These platforms create value by enabling direct exchange amongst its users, and are monetized either by charging the users directly (rarely), or by charging advertisers to sell services to users (most often). They benefit from both same-side effects (more users -> more users) and one-way cross-side effects (more users -> more advertisers). Note that more advertisers do not directly mean more customers. In fact, more ads -> bad customer experience -> less customers. WhatsApp, Facebook, WeChat, Skype are examples.

So why is a pricing view of these platforms4 important? Because the rake that a company (rather a product) can charge depends on where it falls in this spectrum. Paid Cross-side Platforms are more price elastic with respect to the rake that they can charge merchants because a competing platform with lower rake can attract users with lower eventual consumer prices. Hardware Platforms get away with higher rake because of their hardware differentiation. In other words, Apple’s Appstore can get away (and they have for the past 10 years) with charging developers a high rake (30% revenue sharing) because a competing platform with lower rake for merchants is unlikely to cause users to shift (as long as Apple’s hardware continues to attract users). In Free Platforms, prices charged to advertisers don’t (directly) affect the prices on customers, so theoretically new platforms cannot undercut the price (charged to advertisers) of the dominant platform and capture more users. Similar to Free Platforms, prices charged to advertisers in Same-side Platforms don’t (directly) affect the prices on customers, so theoretically new platforms cannot undercut the price of the dominant platform and capture more users. The way to capture marketshare is to offer differentiated service (Facebook vs. Myspace, Snapchat vs. Facebook).

From a cost point of view, there are three costs at play in any business. An upfront capital expense (in case of platforms, this is incurred in developing the platform and is primarily in software development and in some cases physical assets and hardware costs), and variable costs that vary with each unit of output (i.e. costs that vary with the number of sellers and customers serviced such as platform maintenance costs, customer service, payment processing etc.).

So in cross-side platforms (other than hardware), your pricing/rake strategy can take two approaches:

  1. In the absence of sufficient funding, you charge sufficiently high enough to account for your variable costs and the depreciation of your capital expense. While this rake is likely to provide sufficient cash flow to keep you in business, it will also mean slower seller (and thus customer) acquisition for your platform. The less the rake, the faster your seller acquisition. As you grow, though, you should continue to reduce your rake to build a moat around your business and make it harder for new entrants. That is, as you reduce your rake, other startups like yourself will not be able to enter the market. Think about it this way: if you had to charge the high rake in the beginning, arguably new startups would have to do the same, unless they are deeply funded.
  2. The second (and, in my view, better strategy) is to charge a rake that is only slightly higher than your variable costs. This way, you will acquire sellers (and thus customers) faster, and with each seller and customer acquired you will discourage other new entrants from competing. In the long-term, your fixed costs will be (almost) depreciated completely anyway, and so a rake that is only slightly higher than the platform’s marginal cost of serving one additional customer will make it cost prohibitive for any new entrant platform trying to enter the market.

In both strategies though, in the long-term, if your cost of serving a $100 order is $5, then unless you are a hardware cross-side platform (like Apple’s Appstore) you should not be charging sellers a rake of 20% giving you an operating margin of 75%, because someone else could charge lower rake and capture both sellers and customers.

So the next time you are deciding on the pricing strategy for your Uber for X startup, remember – some rakes are more hazardous than others.

PS: In traditional software parlance, platforms are those where other developers can build on top of your application (platform). I take a more broader view of platforms, one that includes marketplaces for sellers and buyers to connect, or advertisers and customers to interact.


  1. Think sellers and buyers at Amazon, riders and drivers of uber, and so on.

  2. With winning bets on several marketplace successes such as eBay and Uber.

  3. That said, Uber, Lyft and other “price setting” Cross-side Platforms are less price elastic because they set the final price to the customer irrespective of the rake. Lower rake directly affects bottomline instead of topline. Also, a managed platform can charge higher rake if customers or merchants value the platform’s “management”.

  4. Some companies sit across multiple models but products typically pertain to one model. For example, Google is a Free Cross-side Platform for search, but Gmail is a Same-side Platform.

Hello World

Welcome to Stratelogical  – a blog that offers insight, analysis and opinion on business and technology strategy.

The act of writing is the act of discovering what you believe.

– David Hare

Through this blog, I seek to explore and discover what I believe in, and express my discovery and exploration to the world.

I am an engineer with a graduate degree in business and several years of experience in technology and engineering companies. During the day, I am a Product Manager at Amazon, and when I’m not Product Managing, I spend way too much time reading tech and business news, and seeking interesting coffee.

I write a lot about my views on Amazon; but the contents of this blog are my personal views and does not represent the views of Amazon or its management. Data and information on this blog will never represent confidential Amazon information.

With my full-time day job, I intend to post only bi-weekly or monthly. To stay updated, please follow Stratelogical on Facebook, LinkedInTwitter or RSS.

I look forward to making this blog worth your time.

© 2018 Stratelogical. All rights reserved.

Theme Hoffman.