Gaining Clarity in the Spot Buy

Data-Driven Thinking"Data Driven Thinking" is written by members of the media community and containing fresh ideas on the digital revolution in media.

Today's column is written by Roland Cozzolino, CTO of MediaMath, an online advertising technology company.

Whenever someone wants to understand the basics of RTB, I like to tell them that they need to understand its data side, if you will, along with the importance of transparency, since that makes for effective ad trading.

What Transparency is NOT

Web servers and sites sort of know your machine.  In fact, that server only knows your machine specific to the session you are on, and the browser you are using, if and only if you have cookies enabled.  Switch browsers and you look like a new "thing" (let's avoid the term person, since we never, I repeat, NEVER know people).  That combination of items, machine + session + browser + cookie, is our unique "thing".  Furthermore, cookie data is specific to the server, meaning this data is only visible if our server under a specific domain placed the cookie.  Outside of that domain, we know nothing.  In fact, the machines visibility into the end-user is but a minute fraction of reality.  Little known fact, credit card companies know more about you than your family does.  So please don't confuse transparency with some sort of personal identification.

I Digress, a Little History

Publishers have historically driven our industry.  They sell their home page to Bob's Cat Emporium at $50 CPMs.  Bob is happy since it appears to be working.  Bob then finds out he could buy another page on Publisher X for 50 cents on a remnant exchange.  Bob is no longer happy, nor is Publisher X since they make all their money via the 5% of direct ads sold.  What does this mean?  Publisher X does NOT want you to know who they are.  This poses one big problem: at the end of a campaign, after the ads are served and the beer is flat, advertisers like to know what worked and why.  When the spot buy is totally opaque, nobody can make an intelligent guess, thereby diluting the value of all impressions going forward.  Fun little problem we got ourselves into.  So how do we make sense of the spot buy?  Let's look at what data is available and quantify that to something viable.

Houston We Have a Clue

From now on consider transparency as seeing the all the various data points on an impression.  That being said, every impression has certain set of inherent characteristics, which include:

  • Geographic Information (city, state, country, …)
  • Date and Time (month, day, year, hour, minute, second, time zone, …)
  • Size of an Ad (728x90, …)
  • Publisher (when available, see above)
  • Domain Level Cookie Data
  • Browser (IE, Firefox, Safari, …)
  • Operating system (OSX, Windows, Linux, …)
  • Price we paid to buy this impression

Then there are attributes you can buy into or may be lucky enough to get if you ask nicely:

  • Third Party Data Segments (Blue Kai, Monitored, IXI, etc.)
  • Site Classification (auto, finance, etc.)
  • Social Demographics (friends of friends)

And finally, there is the random data we all want, but exchanges and ads vary in what they offer:

  • Above/Below the Fold
  • Coordinates on the Page
  • Size of the Browser/Resolution
  • Search Terms (if any)
  • Ad Interaction (hover, click, etc)
  • What someone historically pays for this class of impression

That is actually a lot of data (I definitely forgot some, but that’s enough for now).  Take all of those points and multiply each one by their respective degrees of freedom +1, them multiply those together to get the total combination .  That is one giant, multi-dimensional thing.  So breathe for a second and realize, you do NOT want to analyze this much data.  Why, you ask?  Because you can't do anything with it, it's too big.  Go ahead, plug it into your favorite stat package and watch the machine cry. Let's entertain that you actually could process all this stuff.  If so, you will narrow your audience so much that you will very quickly stop spending money.  Furthermore, the web and advertising is somewhat chaotic, thus choking the inventory prevents you from ever finding new pockets of value.

Effective Trading

So there has to be a better way, right?  Thankfully there is and its not brain surgery.  First and foremost gain insight into the various data points available.  Spend time determining what your client wants and what you believe will work, then bust out the shotgun and start firing blindly.  Go ahead, spend some money.  Now get some reports that aggregate this information in a clear and concise manner and start finding what works.  Remember what I said before, don't ask for all the data, and just ask for subsets so it's manageable.  Now break out the rifle and start aiming.  Keep in mind when you see something that you like, you should and will pay more.  Furthermore, don't stop firing blindly because you never know what the next day brings (feel free to dust off a stochastic book if you like that kind of torture).  Repeat this over and over again till it paints a clear and concise picture.  In the end there are 3 things that should come out of your mouth when talking to the client:

  1. We verified the audience YOU believe loves your offering is accurate (people like being told they are correct).
  2. We found a completely new audience you never knew existed (this is the performance side of trading and bless you if you can pull this one off).
  3. We figured out who does NOT want your product (give us more money, we are narrowing this down to something tangible).

Any of those 3 statements will lead to better campaign management and a happy customer.  What isn't acceptable is, "We got lift, increased ROI, saved money … but … have no idea how that happened".

Summing It Up

RTB offers a ton of information per impression and understanding the data points is paramount to successful trading.  Don’t expect to take all of this data and make any sense of it, its just too big.  If you do, you will find yourself caught up in the details while missing the big picture.  And finally, make no assumption you know the audience ahead of time, let the audience show itself via the data.  Now that’s real transparency.

7 Comments

  1. "Don’t expect to take all of this data and make any sense of it, its just too big"

    While most stats packages will make your machine cry when faced with too much data, we can distribute data and computation over many machines to make sense of it. Standard stats won't work very well; only machine learning (AI) approaches are likely to yield useful and actionable insight.

    The harder challenge for us is dealing with data vendors and ad buyers that still want to buy inventory with old fashioned demographic categories or arbitrarily delineated intender segments.

    Reply
  2. Nice article Roland. Agree that we can learn a lot about audiences in the process.

    But I respectfully disagree about the size and scope of data. Big? yes. Hard to do? yes. Even harder to well? oh yeah. But we eat hard for breakfast.

    All data has value and we spent a lot of effort on learning how to use it for targeting and bidding. Data is the new currency and we don't leave a single penny on the floor.

    Reply
  3. Ken, I agree that all data has value. However, most (not all) companies who actually manage campaigns cannot do this. If you are an agency that does not have the luxury of big data machines, quant guys and your own platform to trade on, you are most likely looking at excel spread sheets when making price changes. If this is the case (and I would bet it is for more than 90% of our industry), limiting the data to manageable pieces is a feasible solution.

    Reply
  4. Roland, that is exactly the antiquated inefficiency that we are trying to address. What would the point be of our companies if we bought media like the agencies with a spreadsheet and a prayer?

    Reply
    • Zach,

      That's Roland's point. Part of the value DSPs such as MediaMath and Invite bring to the table are solving for media buying logistics which have been historically nightmarish. I thought that was obvious.

      Good article. I see the analogy of shotgun -> rifle. It's about tightening the aperture as you learn more. Right?

      Reply
  5. Zach, your sentiment is understood but this article was purposely not intended to showcase MediaMath or any other DSP. It is intended for anyone trying to buy ads on remnant and do so without getting inundated with all the variables. Like financial trading, there are professionals and the everyday home trader. The everyday home trader uses e-trade to execute (with a ton of information to keep them busy for decades). They would be crushed by a pro due to execution times, better information, faster feeds, etc. We, as DSP's, have all this. We (at MediaMath) process terabytes of information across double digit exchanges to buy as effectively as possible. We've proven this for over 3 years, but I have an arsenal of machines and an amazing team of industry vets who know how to pull this off. Reality is, if you really want to buy effectively, use a DSP across every exchange possible a let the machines do the work, then get some aggregrate reports that show a nice picture. That would not have been a very interesting article though. Furthermore (although I did find your tweet funny), a shotgun approach to buying should really be thought of as a "control group". That means understanding the market, understanding the basics of the campaign, what the client wants, and what you know historically has been successful. Just don't be too stringent because tomorrow's another day. RMX calls it "learn", we call it "watermarking". I, of all people, highly recommend using DPS's as your primary execution arm and data collector when running campaigns, but not everyone is in that boat ... thus a simple article on what data is out there and a request to think a little before you start bidding.

    Reply

Add a comment

XHTML: You can use these tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>