Solving the Inventory Problem, Part 2

Two weeks ago, I wrote about the strategic and policy aspects of the inventory problem in online publishing.

Today, I will take it one step further. I’ll identify and tackle the core of the operational issues involved in designing and implementing inventory yield management processes and systems. Let me warn you now that the next few paragraphs will get somewhat technical, so if you’re not interested in the operational goings-on of ad inventory management, you should skip ahead to the “Steps to a Solution” section.

The Starting Point: What’s Available and What’s Booked?

To determine accurate, reliable, and complete inventory figures, publishers must have a handle on both available inventory and booked inventory (already under contract). Most ad operations managers forecast inventory using a combination of manually assembled capacity info and contracted data from various systems (ad server, Web server, order management, registration, etc.). They then manually calculate trending, overlap, and availability. The resulting spreadsheets, although often unreliable, are the “bibles” of ad ops today.

The Challenges: Problems With Existing Systems

Here’s where we start getting a little esoteric.

Inventory overlap. The concept of inventory overlap is a bit complicated, but it is crucial. It is the bane of ad operations personnel, and is at the core of inventory management problems.

Virtually every publisher today offers advertisers multiple targeting criteria in delivering their campaigns (content section, time of day, IP-targeting). While this gives publishers more robust products for their advertisers, it also means that the same user can be counted in a number of different categories. This means the publisher’s inventory in terms of unique users can be said to overlap. For example, you might have 50,000 “business” users, and 20,000 “sports” users. With most sites, some of the business users are also sports users and vice versa. If you sell all 20,000 sports users, then you have also sold some, but not all, business users.

The same holds true if you manage your inventory in terms of impressions; each ad slot can be categorized in multiple ways based upon where it is, e.g. site, section, page, position, etc. A campaign targeted to “run of site X” also consumes inventory from particular sections, pages, positions, etc. Therefore, they overlap as well.

Overlap can be calculated as long as all targeting criteria (dimensions and values) are recorded in the ad campaign request. The amount of overlap is calculated as the percentage of users that fall into one criterion (business or sports or ROS) that also fall into the other. The problem lies in that all dimensions can potentially overlap, and so calculating overlap becomes an n-dimensional problem (meaning it can be enormous and incredibly complicated).

Different ad server delivery methodologies. The “secret sauce” of ad servers is how they target and prioritize. Taking into account how they might perform their various “black box” functions — which occur somewhat mysteriously inside the software with little user control — makes inventory forecasting particularly difficult. Since delivery methodologies are proprietary and limited to the ad servers themselves (DoubleClick, 24/7 Real Media), it is difficult to externally forecast deliveries in coordination with data from the other systems.

Narrow inventory views. Even if you have a reasonably accurate forecast for an ad impression or an audience segment, ad server inventory forecasts provide an answer that is useful at only one point in time, for a single set of targeting criteria. They can’t deliver broad views of inventory availability (or even capacity) across products and across times and across multiple audience segments. They can’t help you efficiently analyze and compare the inventory implications of alternative campaign delivery opportunities. This is essential if you want to maximize the yield of your audience or pages. As a result, almost all ad operations managers are consigned to a horrific amount of manual labor in developing and maintaining such broad views in spreadsheets.

Lack of a unique product key. One of the biggest hurdles to effective inventory management is the lack of a unique identifier for publishers’ ad products. Basically, publishers do not have an index of the products that they sell. All commercial ad serving products, which today represent inventory as a multiple variable, lack a single unique key for identifying an atomic element of inventory. An individual ad product (“banner on the business home page”) is typically represented as a set of name=value pairs, for example “site=X&page=Y&size=Z&position=C”.

Anyone who is familiar with relational database architecture will understand that this presents an immediate and far-reaching problem. If an “atom” is represented only in terms of various overlapping dimensions, the difficulty of viewing the universe of unique “atoms” becomes very difficult.

Fascination with inventory forecasting. Figuring out how to forecast future inventory capacity (or demand) from present behavior is somehow fascinating to many in ad operations. While it is not technically overwhelming, it is just plain unproductive. Many hundreds and thousands of hours have been wasted looking for a “scientific formula” that results in a trending forecast. Trending is, in relative terms, not a business-critical issue. If the other problems can be solved, future capacity (or demand) can effectively be assumed to equal present capacity.

Steps to a Solution

Make audience the center of your world. Most publishers do not have a consistent way to measure the performance of their business that truly crosses departments. Sure, everybody measures some basic financial metrics that have some cross-department applicability — revenue, profits, revenue per head-count, cost per page delivered, revenue per advertiser, advertiser renewal rate — but those metrics are not truly indicative of the performance of the organization against its core asset, its audience.

Stated simply, the business of ad-supported online media is to attract and retain an audience of value to advertisers. All operational systems and financial systems that have historically focused only on product (page tracking, ad serving, order management) need to include audience-centric measurements. Web analytics systems have to track by person, not just by page. Ad servers must be able to target and measure by person, not just by page. Ad order management systems need to crate and track orders by audience delivered, not just by pages and sections.

Create a single source of truth for online advertising data. Publishers can make tremendous strides in the inventory management process if they focus first on integrating data from various “silos” and presenting broad analytical views. Then they can tackle the more difficult problems of overlap and trending.

Build an ad product index on the SKU model. The lack of a unique identifier, or SKU (stock keeping unit), for advertising inventory is a key omission in the design of existing inventory management systems. In fact, it is a key omission in most commercial ad serving systems. Ad operations managers have argued for years that publishers need a single unique identifier for “atoms” of ad inventory. This is in keeping with traditional manufacturing and supply chain processes that make use of an SKU to uniquely identify every meaningful unique variation of inventory. Many proprietary ad serving systems built by larger publishers, including some of the portals, in fact already use just such a model.

Get started now. The inventory problem will not go away without planning and commitment. For most publishers, reorganizing data and systems around audience, creating a single source of truth, and creating an SKU-based ad product index won’t solve all inventory problems, but it will solve a lot of them. But ignoring the issues, or just one-offing solutions, will only make the problems far worse.