0 VisualIQ Connects Attribution to Media Buys

As I threatened – I mean, promised – a couple of weeks ago, I’ll be writing about marketing measurement systems over the next month or two. The hottest topic within this segment is probably algorithmic attribution, which uses advanced statistics to calculate the incremental impact of each marketing contact on final results. Unlike marketing mix models, which work with aggregated data such as total TV spend per month, the algorithmic systems start with individual ad impressions. The systems build a history of impressions for each person, identify groups of people with similar histories, and then attribute different results to differences in those histories. For example, there might be two groups of people, one having received three impressions and the other having received the same three plus a fourth. Any difference in outcomes would be attributed to that fourth impression. The actual details are more complicated, but you get the idea.

The advantage of algorithmic attribution is it is based on actual data, rather than arbitrary fractional weights based on someone’s opinion. (I wrote a more detailed critique of fractional attribution back in 2010;  I’m pleased to see that algorithmic attribution has become more common since then.) It goes without saying that algorithmic attribution makes more sense than assigning all value to either the first or last marketing touch.

VisualIQ was one of the pioneers of algorithmic attribution, dating back to its founding as Connexion.a in 2005. The company has grown rapidly in recent years. This reflects both greater market interest and, perhaps even more important, expansion of the product to connect directly with ad buying platforms. This makes it easier to convert the system’s findings into profitable marketing results.

Like other attribution systems, VisualIQ must start by assembling data about the client's advertising programs. The system gathers individual level data either by embedding its own pixel in Web ads, emails, and landing pages, or by importing impression history from other sources. It can stitch together impressions from several different sources to create a unified individual history. It can also import non-individual-level data, such as TV spend, weather, and competitive conditions.

The system can use this data to build two different types of models.

The IQ Envoy module builds “top down” models based on aggregate data; these are similar to traditional marketing mix models but also show the impact of one channel on another (for example, the effect of TV ads on direct mail response). This lets the system estimate the true net contribution to final results of spending in each channel.

Top down models are fairly customized, but Visual IQ has streamlined its process so it can build a new one in six to eight weeks, which is quick for a mix model. and make regular updates. Once the model is built, marketers can easily see the impact of different budget allocations by moving on-screen sliders. Constraints built into the model prevent users from creating scenarios that are unrealistic.

The IQ Sage module builds “bottoms up” models, which use individual-level data. Users can include real-world constraints such as minimum or maximum spend by channel and diminishing returns from higher spending levels. Once the parameters are set, the system generates a graph showing the best possible results for each spending level. Users can pick the level they prefer and have the system generate an optimal media plan. This can be exported to Excel for manual refinement. Or, users can send the plan directly to real time bidding platforms and online ad exchanges for automated execution.

VisualIQ has two additional modules. Insight IQ allows users to explore the data loaded into the system.  This is important because marketers may have never seen it all loaded and correlated. Audience IQ shows the demographics of best customers and results for different segments.

As you might imagine, algorithmic attribution needs a large volume of data to build an effective model. VisualIQ has found that clients need an ad budget of at least $5 million per year. Those companies typically gain a 15 to 20% increase in media efficiency.

Cost for VisualIQ depends on the modules used, data volume, number of channels and other variables. Pricing for a system including IQ Sage and IQ Envoy starts around $150,000 per year.

0 MMA Modernizes Marketing Mix Models

I’ve been spending a lot of time recently looking at marketing measurement systems. This means that you, Dear Reader, will be spending a lot of time reading about them. A good place to start is Marketing Management Analytics, known to its friends as MMA.

MMA was founded in 1989 and is one of the pioneers in marketing mix modeling. Mix models remain the heart of the company’s business. But while traditional mix models look at direct correlations between advertising and sales, MMA’s current approach takes a more layered view. This includes what the company calls “multistage” attribution, which looks at intermediate touchpoints between an advertisement and the final purchase, and “customer cascade analysis”, which measures the long-term impact of advertisements on brand equity. The company has also beefed up its consulting services to help make its findings more actionable.

MMA’s foray into attribution is intriguing, since it puts the company into some degree of competition with attribution specialists like VisualIQ, Adometry, and ClearSaleing. But MMA works with aggregate data such as total spend and impressions, with a major emphasis on mass media like television. Those other vendors work primarily with data about individual buyers, which comes largely from digital and direct media. MMA's clients are traditional mass media advertisers, in consumer packaged goods, automotive, financial services, retail, pharmaceuticals, and communications, and it is working for CMOs who are allocating budgets across channels. The other vendors' clients are concentrated in ecommerce and they are answering more tactical questions about spending within the digital channels. What they all share is the goal of measuring the incremental impact of expenditures in specific media.

MMA recently released the latest version of its flagship software, Avista.  The system is still focused on traditional marketing mix models, although it can incorporate the "multistage" approach of measuring the impact of one channel on another.  The new release, Avista 8, was designed to make it easier for marketers and media planners work directly with the system, rather than relying on technical experts. The main interface displays curves that represent the relationship between spending on each tactic and final sales. Marketers use sliders to adjust the spending levels and the system then estimates the sales that would result.

Avista can also run optimization routines to automatically find the most effective spending mix. Users can limit how much spending on any one tactic can increase or decrease, can create groups of tactics that draw from a shared budget, and can choose the target of the optimization (maximum profit with a given budget, minimum spend to reach a target revenue level, etc.). Outputs can show details by brand, product, region, sales channel and time period. Users can save scenarios and compare them to each other. Once they’ve chosen a scenario, Avista can convert it to a high-level media plan for buyers to execute.

The system also has a forecasting feature that runs the same models but also lets users change assumptions about factors other than marketing spend, such as weather, competitive behavior, and distribution channels. Results can be displayed on reports, which in turn can be assembled into custom dashboards.

MMA also offers its clients a data access tool called MarketView, which lets them view and lightly analyze the data assembled as model inputs. This is a popular service by itself, because model inputs often include data the marketers have never seen before. Giving them early access helps to speed the modeling process by letting them verify the quality of the data.

0 OMMA Metrics Conference: Online Ads Must Prove Real Value To Succeed

I took a break from my usual obsessions yesterday to attend the New York edition of MediaPost’s OMMA Metrics and Measurement conference. It was a good chance to dive into this particular sector of the marketing analytics universe. (Another version of the program will be presented in San Francisco in July; the company also live streams a free Webcast. You can also download selected presentations from yesterday.)

If there was an overriding theme to the event, it was frustration that online advertising isn’t attracting as much money as it should. There was more than a little ”TV-envy”: the feeling that TV gets more advertising because buying is based on simple, widely-accepted audience measures. Some speakers argued for duplicating this situation, by removing some middlemen and creating standard online audience measures.

Others pointed to a deeper issue: that online marketers can’t measure the value of their efforts in terms of revenue or brand metrics like awareness and preference. In this view, TV buyers accept simple measures like Gross Rating Points because these measures have proven over time to correlate with real business results. Media mix modeling has more recently confirmed this. But except for direct response, online media can’t show the same relationship. This forces online marketers to report endless (but never complete) data about who saw what and how they acted, in the hopes that piling on enough details will somehow make advertisers happy. It never does.

This is the online version of the old joke about the drunk who loses his keys in the alley but looks for them under the streetlamp “because the light is better”. Moral of story: no volume of irrelevant data can substitute for the information you really need.

In the case of online advertising, the dark alley is the connections between ad placements and final business results. Several speakers touched on parts of this. IBM’s Yuchun Lee gave an opening keynote that highlighted the pervasive influence of online information over all customer activities, not just online purchases. Adometry’s Steve O’Brien explicitly stated that attribution must measure the incremental impact of each marketing effort on final results (although I think he limited this to online results). ForeSee Results’ Larry Freed stressed the need to trace all online and offline behaviors to understand their true role in final outcomes. Others cited studies where careful measurement found that indirect results showed online to be much more powerful than direct attribution alone.

Yesterday’s speakers also raised the problem of scalability: that is, being able to duplicate and expand on success. This is one area where TV envy makes sense, because it’s easy to add more Gross Rating Points and be reasonably sure of getting the expected results. Online ad buying is more like buying print ads or mailing lists: you may have some sense of the audience demographics, but the only way to really know how it will perform is to run a test. But this isn't a measurement problem: simple, standard measures that hide true audience differences are only going to be unreliable predictors of actual results. What’s really needed are better testing methods to predict as quickly and cheaply as possible how each new audience will perform. The trick is you’re not just looking at immediate response, but all of those indirect effects that are so tricky to capture in the first place. Now you have to predict them in advance as well as measure them after the fact.

Nobody said it would be easy.

0 Marketing Performance: Plan, Simulate, Measure

Let’s dig a bit deeper into the relationships I mentioned yesterday among systems for marketing performance measurement, marketing planning, and marketing simulation (e.g., marketing mix models, lifetime value models). You can think of marketing performance measures as falling into three broad categories:

- measures that show how marketing investments impact business value, such as profits or stock price

- measures that show how marketing investments align with business strategy

- measures that show how efficiently marketing is doing its job (both in terms of internal operations and of cost per unit – impression, response, revenue, etc.)

We can put aside the middle category, which is really a special case related to Balanced Scorecard concepts. Measures in this are traditional Balanced Scorecard measures of business results and performance drivers. By design, the Balanced Scorecard focuses on just a few of these measures, so it is not concerned with the details captured in the marketing planning system. (Balanced Scorecard proponents recognize the importance of such plans; they just want to manage them elsewhere). Also, as I’ve previously commented, Balanced Scorecard systems don’t attempt to precisely correlate performance drivers to results, even though they do use strategy maps to identify general causal relationships between them. So Balanced Scorecard systems also don’t need marketing simulation systems, which do attempt to define those correlations.

This leaves the high-level measures of business value and the low-level measures of efficiency. Clearly the low-level measures rely on detailed plans, since you can only measure efficiency by looking at performance of individual projects and then the project mix. (For example: measuring cost per order makes no sense unless you specify the product, channel, offer and other specifics. Only then can you determine whether results for a particular campaign were too high or too low, by comparing them with similar campaigns.)

But it turns out that even the high-level measures need to work from detailed plans. The problem here is that aggregate measures of marketing activity are too broad to correlate meaningfully with aggregate business results. Different marketing activities affect different customer segments, different business measures (revenue, margins, service costs, satisfaction, attrition), and different time periods (some have immediate effects, others are long-term investments). Past marketing investments also affect current period results. So a simple correlation of this period marketing costs vs. this period business results makes no sense. Instead, you need to look at the details of specific marketing efforts, past and present, to estimate how they each contribute to current business results. (And you need to be reasonably humble in recognizing that you’ll never really account for results precisely—which is why marketing mix models start with a base level of revenue that would occur even if you did nothing.) The logical place to capture those detailed marketing effort is the marketing planning system.

The role of simulation systems in high-level performance reporting is to convert these detailed marketing plans into estimates of business impact from each program. The program results can then be aggregated to show the impact of marketing as a whole.

Of course, if the simulation system is really evaluating individual projects, it can also provide measures for the low-level marketing efficiency reports. In fact, having those sorts of measures is the only way the low-level system can get beyond comparing programs only against other similar programs, to allow comparisons across different program types. This is absolutely essential if marketers are going to shift resources from low- to high-yield activities and therefore make sure they are optimizing return on the marketing budget as a whole. (Concretely: if I want to compare direct mail to email, then looking at response rate won’t do. But if I add a simulation system that calculates the lifetime value acquired from investments in both, I can decide which one to choose.)

So it turns out that planning and simulation systems are both necessary for both high-level and low-level marketing performance measurement. The obvious corollary is that the planning system must capture the data needed for the simulation system to work. This would include tags to identify the segments, time periods and outcomes the each program is intended to affect. Some of these will be part of the planning system already, but other items will be introduced only to make simulation work.