How an Ad Exchange Might be Constructed

In this post (and several others), I’d like to explore what the institution of an exchange might look like. Believe it or not, I’ve done a lot of thinking about this, both directly and as a background process over the course of the past 2 years. I said in my first post that I am lucky in that I work for one of the larger online companies. I am fortunate enough that, just by virtue of who I work for, I get to discuss this concept with a variety of experts out there. My particular role is interesting in that I get to think about a variety of online advertising types (Display, Search, Rich Media, Video, and others). I’m talking a lot about how lucky I am here, so I’ll add this: I’ve worked in several high tech industries, and I’ve been impressed by how smart people are. I find this to be particularly true in the online advertising world in which I currently exist.

Some Definitions

Yeah, I should start with a few definitions to help things along. Probably should have done that before that last post, so apologies there.

Term	Definition
Advertiser	For the purposes of our discussion, anyone who wants to place advertisements online. They may operate in a variety of formats (search, display, etc.)
Publisher	For the purposes of our discussion, anyone who has online content for which they need monetization. Presumably they subscribe to the web 2.0 theory that advertising is a great way to monetize.
Online Ad Network (usually just “ad network”)	The wikipedia definition of ad network is good. I will say here simply that ad networks are reps of aggregated advertisers. That is, the ad network strikes a relationship with many advertisers (demand) and represents that demand to publishers. An important piece of information for the purposes of our discussion is that the owner of the ad network makes her money on the spread between what their advertisers are willing to pay and what the publishers are willing to take for placements. Some examples of ad networks: ValueClick, AdSense, and others listed in the wikipedia link above.
Ad Platform	An ad platform is a technology by which advertisers and publishers are matched. More specifically, they are technologies that enable the matching of online advertisements to online placements and the transactions which support such advertising. The owners of such platforms can be large publishers (Google, Yahoo!, and Microsoft), as well has ad networks.
Guaranteed or Reserved Inventory	This is online inventory sold on a guaranteed basis. It is sold on a forward basis, the “guarantee” being that your placement will run at a specific time in a specific place. DART, DoubleClick’s ad platform, currently exists in this space as does aQuantive’s new acquisition Accipiter.
Remnant Inventory	Remnant Inventory is more or less what’s leftover after guaranteed inventor has been sold. The wikipedia entry for remnant advertising is useful for context.
Online Advertising Broker (or simply “broker”)	A broker of online advertising is a construct that will exist in the exchange world that will aggregate both supply (advertising inventory) and demand. Brokers are a subset of ad networks (defined above), since not all ad networks have supply.
Home Inventory and Home Advertisers	I use these terms to describe the inventory and advertisers that a broker has within it’s own advertising ecosystem. For example, imagine you are Yahoo and that you have homepage inventory. You also have advertisers. Those advertisers bidding on your homepage inventory is home advertisers bidding on your home inventory
Foreign Inventory and Foreign Advertisers	I use these terms to describe inventory and advertisers that a broker accesses through the exchange. For example, imagine you are Yahoo. When the exchange exposes Microsoft inventory to you through, you can then bid on that Foreign Inventory. When you as a broker submit your inventory to the exchange and the exchange supplies a winning bid, then that bid has been provided by a Foreign Advertiser

Trade in CPM and Impressions

Right now, ad platforms offer pricing in a variety of formats: CPM, CPC, and CPA are the biggies (if you need definitions here turn to this page on online advertising in wikipedia and scroll to the section on payment conventions). Note that ad platforms generally offer CPM for guaranteed placements and mostly CPC and CPA for the remnant (although I’m sure they would be perfectly happy to take CPM). Let’s envision, for now, that our ad exchange operates solely in the remnant world. It’s cleaner that way.

I’ve spoken with some pretty high falutin’ experts in this industry who get extremely excited by the idea of trading for clicks on an exchange. I won’t say who they are, because I believe them to be intelligent people who are in this case exceedingly wrong. The exchange must trade the lowest common denominator item, which is the impression.

What the publisher has available to offer to the exchange is an impression. She cannot offer a click because at the time she needs to conduct the transaction, the event of a click has not occurred yet. She needs to match an advertisement to that impression in order to generate that click, and for that she must turn to a broker or (through that broker) the exchange. To construct a marketplace in which the publisher offered clicks instead of impressions to the auction, we would force the publisher (and her broker) to absorb the risk of click-through on the impression. In my opinion, a foreign broker is much better positioned to balance reward against this risk and I will explain how in “How the Broker Makes Money on Foreign Inventory” below. For now, suffice it to say that a broker will not be able to bid conditionally on impressions – they must take ownership of them outright.

Here’s roughly how I think it might work.

Let’s say you have a trading environment and you have the seller/rep of inventory and the buyer/rep of demand. At an extremely high level, if you have a transaction here your seller will say “I have an ad impression available for trade, my ask is $X.” The entity representing this inventory is a broker as defined above. The ask of $X will effectively be the maximum monetization that broker can generate without leaving its own network for the exchange.

The exchange will represent that inventory to the rest of the marketplace as “I have an impression from Broker A for sale, the minimum bid is $X.” It will also expose to the auction any other relevant and available information about the ad impression (publisher, placement on site, characteristics about the viewer, and anything else it has). The other brokers will gather bids from their advertisers for that inventory and submit them to the auction. The exchange will resolve to one of the following outcomes:

1. The exchange has no bid for that inventory. Broker A will use it’s own system to match the inventory to demand.
2. The exchange cannot meet the SLA associated with the inventory. That is, it cannot supply the inventory with demand fast enough for a suitable consumer experience. Broker A will use it’s own system to match the inventory to demand.
3. The ask (or reserve price) has not been met. Broker A will use it’s own system to match the inventory to demand.
4. The ask has been met. The new bid is $Y. It then passes all relevant information to Broker A about the associated advertiser and broker with the transaction, and the new ad is displayed.

I’ve recommending that the exchange be based on pure CPM. Does that make the name of this blog a misnomer? On to how people make money in this new environment

Economic Models in the Auction

By “economic models” I mean “business models.” For the purposes of our discussions below, I’d like to simplify the discussion by imagining that the exchange would charge no transactional fees on trades. That’s not going to happen, I don’t think, but I don’t believe the business motivations change below once you add on some flat CPM given to the exchange on each transaction (like a penny).

First I’d like to start with a digression on yield management in search.

How Yield Management Works in Search Today

Let’s look at the prevailing yield management solution out there right now for search: the hybrid eCPM auction. Google is the best example. Microsoft built their search auction in Ad Center to utilize the same kind of solution and Yahoo has recently converted as well because it is more efficient than their old solution.

Until a year or two ago, Google’s ranking algorithm for search was described to advertisers as “bid times CTR.” That meant that Google took your max bid, multiplied it by your expected click through rate, and created an index value. That value was ranked from high to low with all the other advertisers’ index values and the ad with the highest index value was shown in the best slot, the ad with the second highest was shown in the second slot, etc. The mathematically inclined out there amongst you will note that bid x CTR x 1000 is equivalent to CPM. That means that Google’s algorithm is a yield management solution – it seeks to optimize revenues. In fact, this algorithm had several advantages over its competitors at the time (in which the rankings were based solely upon bid):

• It is an optimal yield management algorithm, as I just expressed above
• It takes into account the propensity of the consumer to click on the ad (that’s the CTR part). That essentially gave the consumer a vote in the process and created an ecosystem of relevance wherein advertisers were incented to continuously improve the relevance of their ads in order to increase clicks and decrease cost per lead or CPC.

Google has since changed their algorithm somewhat calling it “bid times ‘quality score.’” Quality score as a nomenclature is simply an attempt to obfuscate to the advertiser what’s going on here. Basically, Quality Score was an enhanced estimate of CTR that took into account more factors that Google had available to them in forecasting clicks. In addition, this enhanced estimate is manipulated to take into account longer term user satisfaction in the forecast, essentially to weight long term returns over short term returns a bit (the primary example here is about spam: a ringtone ad may monetize more effectively than a more relevant ad and so would be ordered ahead it. But you can’t let that happen because if all your ads are ringtones then over time you aggravate consumers and lose share).

How the Broker Will Make Money on Foreign Inventory in the Exchange

I said earlier that the “the foreign broker is much better positioned to balance reward against [the risk of click-through].” In fact, one of the pleasant outcomes associated with the exchange model is that the economic model for the foreign broker will incent intelligence and rational behavior. Here’s how:

Let’s imagine the broker bidding on foreign inventory accumulating data in a variety of ways:

• Observing their propensity to win bids in the exchange
• Collecting any tag or cookie data they have associated with transactions originating from the exchange
• And of course, the bidding characteristics and performance history of their advertisers

If the exchange offers foreign inventory to the broker, the broker will have to evaluate the following:

1. What intelligence and data do I have regarding this transaction, both from the exchange and from my own data sources (profile, behavioral segment, IP characteristics, etc.)?
2. Taking into account 1 above, what is the highest bid I can elicit from my advertisers? If they are bidding on CPC basis, I must convert that CPC bid using my data and my algorithms to eCPM. If I must bid true CPM I must evaluate my confidence level around my eCPM and take that into account when I render a final bid.
3. Given the performance of the exchange and the bidding behavior of the other brokers in the system, what is the probability that I will win the auction for a range of bids? What is this probability expressed as a curve? Given all of the above, what bid maximizes my expected value in the exchange auction?

What the broker must do here is construct algorithms to support the following value equation (forgive the bogus math symbols etc, hopefully this will convey the point):

If

EV = the expected value of the bid submitted to the auction; and

B_MAX = the maximum bid the broker has from an advertiser for the specific inventory (given all intelligence on the impression); and

B_SUB = the bid the broker actually submits to the auction in the exchange; and

P₀ = the probability of winning the exchange auction given BSUB

Then

EV = P₀(B_MAX – B_SUB)

The job of the broker in performing against foreign inventory is to optimize against that equation over time. This will maximize their value. The tricky part is that as the delta between BMAX and BSUB increases, P0 decreases. That’s why in 3 above I recommend computing a probability curve. Good luck with that, algorithm.

This is geeky stuff. I wonder if anyone will read it.

Again it’s late and I’m exhausted so I’ll close here, though there is more to discuss in constructing the exchange.

Thanks for reading

DM