The Programmatic Waterfall Mystery

aripaparosellsiderThe Sell Sider” is a column written for the sell side of the digital media community.

Today’s column is written by Ari Paparo, CEO at Beeswax.

It may sound like the latest in young adult fiction, but this mystery is not dramatic or amusing, nor will it make your teenager ask you for money for the movies.

The mystery is this: In the supposedly super-efficient world of RTB, why would publishers continue to waterfall their demand sources?

Spoiler alert: This column will not actually answer the question. I merely posit possible explanations, which you will have to wait to see if the writer can tie together in an as-of-yet unscheduled Episode 2.

First, let’s start with some economics. As should be familiar to anyone who has anxiously waited for an airline upgrade, there’s a big difference in price between coach and business class, and it’s not easy to switch between the two. The economic theory behind this is that, as a seller, you can maximize your profit by touching different consumers on different portions of the demand curve.

Differential pricing maximizes yield on a demand curve


Before the exchanges, most publishers executed demand segmentation through a waterfall of tags. Tags from various ad networks would load only after someone else had passed on the impression. The lower you were on the waterfall, the fewer premium impressions you had access to, which was a form of quality discrimination. If you don’t believe this matters, consider the difference in ad impression value between the first page of a slideshow and the last one.

Differential pricing using an old-fashioned waterfall/daisy chain


Fast forward to today’s world, where every impression can be simultaneously auctioned to every potential demand source in real time, with business rules customized to maximize revenue. This should allow publishers to maximize yield, not just on two points of the demand curve, but theoretically on every point.

Diagram: How RTB should be the perfect solution to maximizing yield


So what do smart publishers actually do? They waterfall, or “daisy chain,” supply-side platforms (SSP). They send an impression to their preferred SSP with a relatively high floor price, then if the impression doesn’t clear, they redirect it to a second SSP with a slightly lower floor price, and repeat the process until AdSense clears the impression at pennies. This is essentially differentiating demand based entirely on the buyers being unaware that they could buy the same inventory cheaper. It’s as if you were able to repeatedly reload the American Airlines site and get cheaper prices for the same seat. When you ask publishers why they do it, they say, “Because it works.”


This is the ugliest secret in programmatic. I’ve personally verified it with more than a few top ad ops folks.

Any economist could tell you that this is a bad idea. But the ad ops folks insist that it works. So let’s trust them and try to answer the waterfall mystery: Why does it work?

Hypothesis No. 1: Different SSPs Have Different Demand Density

This seems like the obvious answer, but upon closer inspection, it doesn’t make sense. The top 80% of demand comes from a handful of bidders that integrate with everyone. Further, if one SSP had so much more demand than the others, it would always be the first in the daisy chain and gain significant share. But publishers insist they must move the order around over time and the SSP market remains a fairly stable oligopoly.

Conclusion: Unlikely.

Hypothesis No. 2: SSP Tools Produce Different Yields On The Same Inventory

The SSPs do have varied tools to manage yield. But once again, if one is dramatically and consistently better than the others, the market would reach equilibrium with the leader on top for most clients. This is not the case.

Conclusion: True, but unlikely to be the cause of the waterfall.

Hypothesis No. 3: Buyer Tools Differentiate By Supply Source

Do bidders price the same inventory differently if it comes from different supply sources? If so, this would cause some unpredictability in determining optimal auction settings, and switching between SSPs would be a crude way of generating incremental yield. But this could also cause reduced yield, since the combination of auctioneer, inventory and buyer would be hard to predict.

Conclusion: Possible, but hard to test.

Hypothesis No. 4: The Buyers (And Their Tools) Are Dumb

What if most buyers just don’t realize they can get the same impressions at lower prices and are instead consistently buying impressions for higher prices than they would ultimately clear for?

This would explain both the need for multiple auctions – to fool buyers into thinking they are distinct impressions – and the need to switch the order of the waterfall – to throw off the demand-side platforms once they have reached an optimized state.

Conclusion: This is a clear contender!

Hypothesis No. 5: Ad Ops Teams Like Messing With Things

What if yield is just going up because someone is paying close attention and tweaking settings a lot? See the Hawthorne effect.

Conclusion: Possible, but don’t tell ad ops.

I’m interested in hearing your thoughts. In a future article I’ll work through the strategic implications of this situation.

Follow Ari Paparo (@aripap), Beeswax (@BeeswaxIO) and AdExchanger (@adexchanger) on Twitter.

Enjoying this content?

Sign up to be an AdExchanger Member today and get unlimited access to articles like this, plus proprietary data and research, conference discounts, on-demand access to event content, and more!

Join Today!


  1. Nick Jordan

    Glad to see a discussion that actually includes economic theory and not just hand waving. Well done.

  2. Ari, congrats on the great article. The comparison of Waterfall vs Daisy Chaining SSPs is great when you see on a graph. Trying to answer your original question ( Why Daisy chain SSPs ?) I believe it’s a lack of knowledge or trust from the decision makers. The myth of monetizing 100% of the impressions comes from other business models and experiences, and may not be the right principle to manage Yield. In theory There is the “fight” of two different algorithms ( DSP x SSP ) , but this is a over simplistic analysis. What everybody is trying is to reach a equilibrium between “value” and “price”. But as the internet has no scarcity, I believe it’s more of a Gambling game. Before we add people to the mix, who can bend any limit of science to justify the price ( that’s why salespeople are top paid employees )- it’s is like having you computer trying to beat the other on Math Class. But the true winning is keeping the equilibrium “healthy” and not seeing 1 impression for US$$1000, but selling an average part for the hypothetical maximum revenue. In other words, it’s really complicated for most of the decision makers in the value chain, mostly because there is no absolute truth, that can be contested to “fire” you or promote you. Therefore, like IT people used to say in the 80’s, you do not get fired when you buy IBM, they were not the best, or cheaper, but would keep your job safe … so My thoughts on that is that people do that in order to be safe. Of course companies on edge, are making more money, but there is no way to prove it. What do you think ?

  3. As an SSP, we’d agree with all 5 hypotheses. 1, 2, and intuitive and somewhat rational (i.e., difficult to correct), while 5 is less rational and more difficult to correct. To me, Hypothesis 4 is the interesting observation: not that advertisers are stupid, but their desire for pace (budget or impressions per time period) may create a bid price that doesn’t align with ‘market value’. Arbs don’t pace, so they’ll never force a bid.

  4. Ari, this is a great article – having seen this issue on the inside myself, I’d say Hypothesis 2 is more of a player in this equation than you give it credit for, though.

    My sense is that this has quite a lot to do with frequency valuation on the DSP side, lop-sided cookie syncing, and differences in dynamic floor algorithms on the SSP side.

    I also think the Enhanced Dynamic Allocation is doing a lot of weird things to the market right now, because it causes impressions to route in unpredictable ways to supply side exchanges. There are a lot of weird ways you can configure this that lead to different outcomes (what rate do you assign to your SSP order lines, how specific do you target them, and how often do you update the rate? etc.) AdX is the one marketplace with a large, additional source of demand not available on the other SSPs (AdWords) so it seems to lead well, but seems weaker in the mid-range bid values from what I’ve seen.

    And speaking of AdWords, I don’t know who you’re talking to, but from my perch the RTB marketplace still has a lot of ‘dumb’ network demand that clears through the RTB pipes but isn’t really bidding like a DSP. I think that’s still very much providing differences in the yield & bid density curves, and creating inefficiency that publishers are trying to optimize around.

  5. I think it’s misleading to call waterfalling and daisy-chaining the same thing in this case. Daisy-chaining in the old ad net days often meant the practice of actually buying and reselling the impression multiple times. Unless I read this wrong, this isn’t transacting the same impression multiple times, but instead cycling from one possible buyer or seller to the next in order to facilitate a single sale. Am I reading it right?

  6. Alex Gray

    Thanks Ari, a very interesting read indeed. However, I think you’re ignoring the game theory aspect here. You say that daisy-chaining is “based entirely on the buyers being unaware that they could buy the same inventory cheaper”. This is like saying that an auction is based on the bidders being unaware that they could still have won if they placed lower bids. That might be right, but irrelevant; and likewise here: a buyer who’s more interested would buy the impression with higher floor price, because for them, it’s worth it. If they waited for a lower floor price, they could lose the impression to competition. Thus, daisy-chaining helps making sure buyers actually pay as much as they’re willing to, and close to the actual economic value of each impression.

  7. I think #3 is a clear contender. Buyers value SSPs differently, due to their stance on quality, anti-fraud, target market, etc. Many SSPs plug each other in as supply and demand sources – so you get SSP1 buying traffic on SSP2. Why would SSP1 ever have buyers going through it that they could get directly on SSP2? Because SSP1 may have a better reputation or relationship with the ultimate buyer.

  8. Sanjiv Ghate

    Great discussion and comments. Have you considered possibility that different extents of cookie matching overlap between bidders and exchanges can be influencing this behavior? A bidder maybe skipping the auction on the first exchange because the cookie is not matched and there isn’t much info for it to bid on? It’s possible that on the 2nd exchange the cookie is matching and it bids for it?

  9. Hypothesis 4 most likely. Outside Programmatic and just pure supply and demand. I used to work for on major global digital publisher brands and portals (Espn, msn, Supersport, etc.). We used to sell major online sponsorships, think Olympics, at premium price for Share Voice exclusivity 2-6 months prior to the start of the tournament. Depended on Share of Voice buys left, the price would drop signifcantly (50-90%) in the last month…. when the tournament started, you could almost get it for free or the inventory would be given to third parties to sell on our behalf. If you really want a good deal from publishers…. find out when their month ends – cut off normally around 24th. Phone the head of sales and ask what is on offer. Unfortunately they tend to be short term focused and devalue their inventory month end.

  10. Jesse Clemmens

    Great discussion. One additional element worth mentioning: within the SSP set there are a wide variety of platform policies to handle buy-side reselling. The most well informed ad ops people understand that each buyer beneath the SSP hood should bring unique demand. Publishers playing the waterfall game rarely enforce this for a variety of reasons and so in reality the dotted lines between exchanges 1, 2 and 3 (in the waterfall example graphic) are often extremely dotted or non-existent.

    Separate note, for publishers it is notoriously difficult and time expensive to run a good “bake off” or test of competing SSPs. Someone’s ALWAYS going to get the short end of the stick, be it chance cookie match overlap, seasonality, block-list inconsistency, etc. I bet that many diehards in the waterfall camp started out with the intention of picking a single SSP partner but ended up indefinitely juggling.

  11. Great article! I would say it isn’t easy to narrow this down to a single theory as a cause but rather look at the dissonance of the market as an emergent property of all the various factors.

    Even though RTB was supposed to open up exchanges, there’s still lots of data that is siloed and available on different SSPs, which in turn are valued separately by buyers depending on reach, performance, fraud, viewability, formats, audience, and support for any other necessary integrations by the buy side. Add in tech that’s great standalone but has to work through lots of disparate systems and there’s a recipe for the results here which can be strangely fixed through something as simple as a round-robin waterfall approach, arguably more normalization than optimization but it says a lot that it still works.

  12. Bas Vijfwinkel

    It’s simply a matter of cost and convenience.
    SSPs don’t have the capacity to send every request to every DSP they are connected to.
    Although every request to a DSP is just a fraction of a dollarcent, it adds up when you’re doing gazillion impressions. And sometimes scaling up poses additional issues regarding your infrastructure.
    By sending the requests first to a well performing DSP, you can cut your equipment related costs (servers/required network bandwidth) significantly.
    If that DSP generates a satifactionary revenue to support your business, why bother that you might raise your profit by maybe 10~20% by having to invest and maintain a significantly more complex system that can send out all request to everybody?

    And why do some DSP’s perform better? That all depends on what type of clients they are doing business with. Brand advertisers do actually pay better for traffic, especially if it’s combined with a cookie for retargeting.
    Also some algorithms are designed to win each single bid, not optimize the cost for all bids. That as major implications for the bidprice.

  13. Tai Conley

    Great discussion, and a topic I think about all the time. I challenge the notion that this creates economic inneficiency. An efficient market allows supply to meet demand at all points along the curve. And if a buyer is willing to pay for supply at a higher price, there is no economic inneficiency lost in the transaction. The entire point is that only fewer buyers are willing to meet supply at a higher price, but the seller is losing if it doesn’t post requests at a higher floor.

    Interesting to note that DSP response rates remain relatively low, and match rates vary across SSPs and DSPs. So even if a publisher only posted requests at a low floor, or with no floor with only one SSP, it wouldnt get the fill rate it needs. As a result, publishers are forced to use multiple SSPs

  14. Great article Ari!

    Allthough I think you are missing one huge point here, which is actually the most important one for publishers to daisy chain. Publishers want to make the most of every single impressions, obvious right? But how is it exactly that RTB works, oh yeah it’s 2nd. price auction based! What does this actually mean? Well, first of all it means that if a buyer values the publishers impression at CPM 50 (bidding price) and their floor price is 10 (and no other bidders), then there is a value-gab of CPM 40 and the bidder wins at 10,01. Okay, so did the publisher just sell his impression for XX% lower than the bidder was willing to buy it at, well thats kind of stu*** isn’t it?

    Now what if I, without any major dissadvantages could manipulate this 2nd. price aution by asking the market several time, at each price point, i.e 49, 48, 47 (in theory) I would actually be earning the EXACT price that the buyer would be interested in buying the impression for. Now this is the major driver for daisy chaining, ensuring the publisher the price the bidder is willing to pay without negotiation in any way, simple market forces at work.

    This raises questions regarding the bidding type could be reinvented to work as highest bidder win, but for the bid they placed, making more precise and accurate values of every single impression, since the bidders bids will not risk paying too much for the “product” right?

    Have a great weekend and enjoyfull reading.

  15. As a publisher that utilizes multiple SSPs, I utilize the top exchange as pure RTB (Private Marketplace and Open Auction) with higher floor rates using dynamic allocation. This enables buyers in each auction to bid on inventory that competes with guaranteed campaigns much higher in priority within the ad server.

    Our second SSP is the mediation layer that manages and optimizes all network partners and also competes with their RTB exchange. This delivers lower in priority, after guaranteed campaigns have the opportunity to deliver. Therefore, we expose this inventory at a lower price floor.

    The buyers within the first exchange could pass and buy at a lower rate within the second SSP exchange, however they could lose the impression entirely because of our direct sold guaranteed campaigns.

    Additionally, each SSP platform has rule and targeting capabilities which vary from platform to platform. The ability to segment and price inventory packages can be somewhat limited, so using the multiple SSP waterfall allows for a wider variety of rule sets.

    A third SSP is also used to keep the second SSP honest in terms of rate. It again is limited to pure RTB exchange volume, but in the event of a “race to the bottom”, there is a counterbalance to prevent one from negatively affecting overall yield. Or, to simplify, competition breeds success.

    Additionally, different SSPs offer different features like programmatic guaranteed, native/video support, reporting and analytics, DMP integrations, customer service, etc. With the land grab of tech like YieldEx, Chango, iSocket, Shiny Ads, etc, is it no wonder that Publishers need to ensure they have business relationships with multiple partners?

    Now, why do I set it up this way? Because, yes, it works.

  16. Great article. I would add that this isn’t just about SSPs, but also the individual publishing properties that advertisers are looking to target. Certainly properties have a higher brand value than others (very hard to measure) leading to either a real or perceived increase in value of the userbase. The idea that one can simply buy audiences doesn’t quite work, given that advertisers value the inventory as much as the user themselves. This ultimately means that whoever controls the supply will be able to charge more money, particularly if they are good at packaging multiple properties.

  17. Giovanni

    I’d add few other reasons which can contribute to this phenomenon:
    6) If SSPs are set in a waterfall-like setup, each SSP can work as a floor for the previous one. This means that publishers are practically operating with “soft floors”, making the highest bidders pay more, and still selling through lower priority SSPs whatever doesn’t clear the top floor.
    7) Optimization algorithms. Many DSPs, especially in the Learning phase, might treat the same impression differently depending on which exchange it comes from. So even the same exact impression might have different values for the same buyer depending on which SSP is managing it. Also, some DSPs have systems to Learn across SSPs based on predefined budget allocations. If one publisher works with more SSPs, it has a higher probability to be considered.
    8) Pacing – Most DSPs do probabilistic pacing, calculating the optimal % of response rate to be able to deliver evenly. If one campaign has a pacing score of 50%, it will basically flip a coin for every valid ad-request it sees in order to bid. if the same impression comes three times through three SSPs, the probability of buying it grows up to 87.5% (100% – 0.5%*0.5%*0.5%).

  18. The problem with all this daisy-chaining and waterfalling is the appalling damage to user experience, especially on mobile devices. Recently I did a survey on some of our pages using AdSense, and discovered 1204 extra requests on one page – and that was pretty typical. Asynchronous or not, that makes for a pretty rubbish user experience. Unfortunately, apart from removing all Google advertising from our site, there’s really little we can do. No wonder ad-blockers and hosts files are getting ever more prevalent.

  19. Great Article. I think #1, #3 and #4 are most likely correct. I have noticed the same advertiser won’t bid on SSP1 (and will let go impression) but open to bid on SSP2. So as a publisher I may loose him completely if I don’t daisy chain. As a publisher, I would rather prefer to have one SSP.

  20. I think you left off something important: doing this right requires hiring very high quality engineering talent. Right now that’s very, very difficult; you’re competing for _exactly_ the same talent that Google/Facebook/Apple/Microsoft/Amazon/HotStartup are after. And they’re voracious; here in Seattle, you can’t walk down the street without getting mobbed by recruiters. The Bay Area is even worse.