Category: Uncategorized

The following column originally appeared in the mighty AdExchanger on Feb. 15, 2022

“I said ‘Hey, what’s going on?’” – 4 Non Blondes

Back in 1994, when a 23-year-old Netscape engineer single-handedly enabled third-party cookies by default, digital advertising was a $50 million business. Now it’s at $450 billion, and a lot more people are involved.

It seems they don’t agree – not just on technical issues, which can be solved, but on existential ones. Like: “Is ad targeting and measurement good for our society or not?” Or: “Is requiring a person to opt in to ‘tracking’ fair?” 

The Privacy Sandbox was launched by Google’s Chrome team in 2019 as a test bed for ideas. They chose to take their ideas to the World Wide Web Consortium (W3C). This step was not required; Apple and others regularly make changes to products on their own.

As the W3C’s hard-working counsel and strategy lead, Wendy Seltzer, admits: “We can’t force anyone to do anything. We look for places where we can help find consensus.”

And in the past month, there’s been a flurry of Sandbox-related announcements: a potential replacement for the FLoC proposal, in-market tests for measurement and attribution ideas, a new working group.

Amid all this excitement, we’d be forgiven for thinking we’re on the brink of adopting universal standards for ad targeting and measurement. Not quite. We’ve become so used to a splintered internet that the whole idea of a self-regulated World Wide Web with the same rules of engagement for everyone seems as quaint as “Do Not Track.”

Building castles in the sand

As a cooperative venture, the Web relies on the goodwill of participants to survive. The W3C and its nerdier cousin, the Internet Engineering Task Force (IETF), are certainly doing their jobs.

Despite what we think, advertising is only a small part of the W3C’s daily grind. (It almost never comes up at IETF meetings.) Sandbox ideas end up in the Improving Web Advertising Business Group (IWA-BG), the Privacy Community Group (PCG) or the Web Incubator Community Group (WICG). Only the first one is focused on ads. The IWA-BG has 386 registered participants, 62 more than the Music Notation Community Group but 14 less than the more-popular Interledger Payments Community Group.

The main work of the W3C members consists of responding to issues on GitHub and holding conference calls, which are fun to audit. They’re definitely overworked. Two weeks before last Halloween, a new group called the Private Advertising Technology Community Group (PAT-CG) launched with a lot of momentum. At the group’s first gathering, one participant made the obvious point: “Many of us are struggling to take active part in all the groups active in this space.”

Like most committees, these ones can inspire angst. Frustration could be felt in the Twitter screed of one of the PAT-CG’s champions: “The folks in this group are *hungry to make progress*.”

What is clear is that the pro-advertising contingent is fighting uphill. During a presentation to the IETF last year, a Google engineer describing the FLoC proposal felt the need to justify the project by citing academic studies about the economic impact of cookie loss on publishers. In the same meeting, an Apple engineer talking about Private Relay, which masks IP addresses (and can break things like time zone and fraud detection), felt no need to justify promoting “privacy.”

The trouble is – and this is the crux of the issue – there’s still no consensus here on a very important, foundational question: What is privacy?

There’s a team called the Technical Architecture Group (TAG) within the W3C drafting a set of “privacy principles.” These are still a work in progress with many stakeholders, and the W3C’s Seltzer said in a meeting last fall that “it’s a tough challenge to bring all those perspectives together.”

But the ultimate success of this draft or a related privacy threat model that would herd the privacy cats isn’t clear.

So, what happens now?

Given its limited objectives, the Sandbox is succeeding. The Chrome team has received a lot of feedback and is reacting. According to the latest updates, four proposals have completed or are currently in trials (Trust Tokens, FLoC, Core Attribution and First-Party Sets). At least two more will enter trials this year.

Results are mixed, but that is just how engineering works: blunt feedback and iterations. FLoC itself has flown through an initial test, a redirection and recent relaunch, and it has hatched a whole aviary of suggested improvements. Missing in all this is a promise of cross-browser, Web-wide solutions.

The impact of FLoC is instructive in another way – one that’s reminiscent of the “Do Not Track” experience. In the latter case, a member of the W3C working group, Ashkan Soltani, grew frustrated and ended up helping to draft the CCPA and CPRA regulations. (Soltani is now in charge of the California Privacy Protection Agency.)

Similarly, a vocal member of the W3C Privacy Sandbox, James Rosewell, drafted a complaint that, in part, led to Google’s agreement to cooperate with the UK’s Competition and Marketing Authority. This agreement was accepted by the CMA just before Valentine’s Day, while a coalition of European publishers filed another complaint.

Seems like, in the end, the future of the cookie may just be worked out between the parties with the power here: Alphabet and the regulators.

Follow Martin Kihn (@martykihn) and AdExchanger (@adexchanger) on Twitter.

This article originally appeared in The Drum on 2/26/21

When I’m thirsty, I go with water. When I’m hungry, I drink beer.

It wasn’t me who made the decision. It was the people on Reddit!

Wow no cow, no beef. This was so good it even tasted like bacon.

Imagine for a moment you are in a creative brainstorm, and a junior copywriter swoops in bravely with the above. You might pause for a moment, inhale, and say, “it’s a start.”

Now what if I told you that copywriter was a machine who had been given a specific prompt (in italics) based on recent spots from Super Bowl LV? Well, it was a machine.

Those lines – and dozens less sensible – were generated on my MacBook Pro using a pretrained open-source natural language A.I. model called GPT-2, built by the Elon Musk co-founded OpenAI. It was “steered” by a list of words taken from Super Bowl ads using another open source code library called PPLM, built by Uber Engineering.

Loading and learning the models took about an hour. And given a few-word prompt, GPT-2 happily takes about five minutes to churn out 20 “ideas,” without breaking for lunch.

Text generation – or robo-writing – has made startling leaps in the past few years, moving from punchline to something that may deserve a seat in the creative lounge. Believe it or not, the best robo-writers are almost the equivalent of that most annoying/wonderful phenomenon: the eager beginner, completely inexhaustible but creatively uneven.

Most of GPT-2’s “ideas” were not quite ready to be presented; some were nonsensical. Oddly, it had no clue what to do with the prompt “Jason Alexander.” And one of its “Wow no cow” completions was “Wow no cowbell can be quite like the best in the universe.”

Which is probably true and not at all helpful.

In the near future, the smartest creative teams will be those that can use A.I. writers in productive ways, as a computer assist to a creative session and a source of ideas that might spark better ones.

One trillion parameters

At first, GPT-2’s creators were so afraid of its power falling into the wrong hands that they were reluctant to release it. They relented and rely now on academic partnerships to limit bad actors like cyber-propagandists. Although not open source, GPT-2’s successor — called GPT-3 — is available to try on application as an API. The full model was recently licensed to OpenAI’s major investor, Microsoft.

GPT-3’s largest setting has 175 billion parameters. Think of these as individual knobs that the model has to tune, based on human writing samples, in order to predict the next word in a sentence. Google just open-sourced an even larger text model called Switch Transformer that reportedly has more than 1 trillion parameters.

The human brain has about 3 trillion synapses. Let’s leave that right there.

GPT-3 takes GPT-2 out of the sandbox and sends it to middle school. Give the model a prompt (e.g., “Once upon a time…” or “It wasn’t me…”), and it can continue at some length, generating text that is often plausible and sometimes uncanny. Early testers ranged from awed to skeptical – and often both.

Hype grew so heated last summer that OpenAI’s chief executive Sam Altman took to Twitter to reassure humanity that GPT-3 “still has serious weaknesses and sometimes makes very silly mistakes.”

The most angst issued not from writers and poets — who are depressed enough already — but ironically enough from computer programmers. It turns out that computer code is also a language that GPT-3 likes to write.

In fact, that’s what makes this new generation of robo-writers different: they are flexible models, verging into the space called artificial general intelligent (AGI). This is the kind of intelligence we have: not pre-trained in any particular discipline but capable of learning. GPT-3 seems to perform well on a range of language tasks, from translation to chatting to impressing electronic gearheads.

Ad copywriting isn’t such a leap. As a tool to build creative prompts from catch phrases ready for human filtration, so-called Transformer models make a lot of sense.

From the refrigerator to the gallery

Even as AI agents got noticeably better at diagnosing fractures and targeting drones, their creative efforts were conspicuously weak. Robot “art” looked like it belonged on a refrigerator in a garage, and robot “poetry” not even on a prank e-card. This is changing.

Robo-writers are already employed in shrinking newsrooms. So far, they’re mostly stringers on the high-school sports reporting, weather and stock market desks — churning out endless Mad Lib-style pieces in routine formats about games and finance that no sane journalist would want to write, even for money.

Computers thrive at tedium. It’s their métier. The for-profit OpenAI takes a brute force approach to its innovation. Two years ago, it gained notoriety for developing a machine that could beat the world’s best players of a video game called Dota 2. It did this by having software agents play the equivalent of 45,000 hours of games, learning by trial-and-error.

The GPT family of tools were also developed by pointing software agents at a massive corpus of data: in GPT-3’s case, millions of documents on the open web, including Wikipedia, and libraries of self-published books.

GPT-3 is exposed to this mass of human-written text and builds a vocabulary of 50,000 words. Its model’s weights predict the next word in a sequence given the words that came before – and develops meta-learning beyond simply memorization. It requires a prompt and can be guided by sample text that provides a “context,” per the Super Bowl examples above.

It’s a trivial matter to drop a prompt like “Car insurance is …” into the GPT-3 Playground, tweak a few toggles, and generate snippets of sensible prose. It’s not much harder to guide the model with a sampling of, say, action movie and comic book plots and generate stories at least as coherent as those of some recent superhero movies.

To answer the obvious question, it can be shown that GPT-3’s creations are not just plagiarism. But are they pastiche? The model learns to predict words based on its experience of what others have written, so its prose is predictable by design. But then again, isn’t most writers?

Limitations include a “small context window” – after about 800 words or so, it forgets what came before – so it’s better at short pieces. It has a short attention span, matching our own.

For this reason, people who have spent more time with the model grow less impressed. As one said: “As one reads more and more GPT-3 examples, especially long passages of text, some initial enthusiasm is bound to fade. GPT-3 over long stretches tends to lose the plot, as they say.”

But ad copywriting isn’t about sustaining an argument or writing a book. It’s about trial and error and inspiration. Increasingly, that inspiration may be coming from a robot.

This post originally appeared on the blog of the NYC Data Science Academy, where I am a student. We were assigned the Ames Housing dataset to practice Machine Learning methods such as Lasso, Ridge and tree-based models. I added the Plotly Dash out of sheer exuberance.

The Ames Housing dataset, basis of an ongoing Kaggle competition and assigned to data science bootcamp students globally, is a modern classic. It presents 81 features of houses — mostly single family suburban dwellings — that were sold in Ames, Iowa in the period 2006-2010, which encompasses the housing crisis.

The goal is to build a machine learning model to predict the selling price for the home, and in the process learn something about what makes a home worth more or less to buyers.

An additional goal I gave myself was to design an app using the Plotly Dash library, which is like R/Shiny for Python. Having been a homeowner more than once in my life, I wanted a simple app that would give me a sense of how much value I could add to my home by making improvements like boosting the quality of the kitchen, basement or exterior, or by adding a bathroom.

Conceptually, I figured I could build a model on the data and then use the coefficients of the key features, which would probably include the interesting variables (e.g., kitchen quality, finished basement). These coefficients can predict what a unit change in a feature would do to the target variable, or the price of the house.

Data Preparation

One advantage of the Ames dataset is that it’s intuitive. The housing market isn’t particularly hard to understand, and we all have an idea of what the key features would be. Ask a stranger what impacts the price of a house, and they’ll probably say: overall size, number of rooms, quality of the kitchen or other remodeling, size of the lot. Neighborhood. Overall economy.

To set a baseline, I ran a simple OLS model on a single raw feature: overall living space. My R^2 was over 0.50 — meaning that more than half of the variance in house price could be explained by changes in size. So I didn’t want to overcomplicate the project.

The data had a lot of missing values. Many of these were not missing at random but rather likely indicated the feature was not present in the house. For example, Alley, PoolQC and Fence were mostly missing; I imputed a ‘None’ here. Other values needed special treatment.

LotFrontage was 17% missing. This feature turned out to be closely related to LotShape, with IR3 (irregular lots) having much higher LotArea values. So I imputed values based on LotShape.

There were a lot of categorical values in the data which I knew I’d have to dummify. Before going there, I looked into simplifying some of them. ‘Neighborhood’ seemed ripe for rationalization. It was clear (and intuitive) that neighborhoods varied in house price and other dimensions. There didn’t seem to be a pattern of sale price and sale volume by neighborhood (except for NorthAmes), but Neighborhood and Sale Price were obviously related.

Now I know it isn’t best practice to create a composite variable based only on the target, so I created a new feature called “QualityxSalePrice” and split it into quartiles, grouping Neighborhoods into four types.

There were many features related to house size: GrLivArea (‘Graded’ or above-ground living area), TotalBsmtSF (basement size), 1FlrSF, even TotRmsAbvGrnd (number of rooms). These were — of course — correlated, as well as correlated with the target variable. Likewise, there were also numerous features related to the basement and the garage which seemed to overlap.

To simplify features without losing information, I created some new (combined) features such as total bathrooms (full + half baths), and dropped others that were obviously duplicative (GarageCars & GarageArea).

Looking across both continuous and categorical features, there were a number that could be dropped since they contained little or no information; they were almost all in one category (e.g., PoolArea, ScreenPorch, 3SsnPorch and LowQualFinSF).

Finally, I dummified the remaining categorical features and normalized the continuous ones, including the target. The sale price (target) itself was transformed via log, to adjust a skew caused by a relatively small number of very expensive houses.

Data Exploration & Modeling

Everybody knows (or think we know) that it’s easier to sell a house at certain times of year, so I looked at sale price by month and year. These plots showed a peak in the fall (surprisingly, to me) as well as the impact of the 2008 housing crisis.

Correlation for the continuous variables showed a number aligned with the target variable, and so likely important to the final model. These included the size-related features, as well as features related to the age of the house and/or year remodeled.

Because my target variable was continuous, I focused on linear regression and tree-based models. I ran OLS, Ridge, Lasso and ElasticNet regressions, using grid search and cross-validation to determine the best parameters and to minimize overfitting.

The error functions for these models on the test data set were all similar. The R^2 for Ridge, Lasso and ElasticNet were all in the range 0.89=0.91. The significance test for the OLS coefficients indicated that a number of them were very significant, while others were not.

I also ran tree-based models Gradient Boosting (GBM) and XGB (also gradient boosting) and looked at Feature Importances. The tree-based models performed similarly to the linear regression models, and pointed to similar features and being significant. Of course, this is what we’d expect.

In the end, the most important features across all the models were: OverallQual (quality of the house), GrLivArea (above-ground size), YearRemodAdd (when it was remodeled), NeighType_4 (high-end heighborhood), and the quality of the kitchen, finished basement and exterior. If nothing else, the model fits our intuition.

Feeding the features through the model one by one, additively, it became obvious that the most oomph came from 20-25 key features, with the rest more like noise.

Plotly Dash App

Settling on a multiple linear regression as the most transparent model, I ended up with an array of 23 features with coefficients and an intercept. To use them in my app, I had to de-normalize them as well as the target.

The app was aimed at a homeowner who wanted to know the value of certain common improvements. The coefficients in my linear model were the link here: each coefficient represented the impact of a UNIT CHANGE in that feature on the TARGET (price), assuming all other variables stayed the same. So I just had to come up with sensible UNITS for the user to toggle and the impact on the TARGET (price) was as simple as multiplying the UNIT by the COEFFICIENT.

Since I was building an app and not trying to win a data science prize, I focused on the significant features that would be most interesting to a remodeler. Some things you can’t change: Neighborhood and Year Built, for example.

But some things you can: in particular, features related to Quality would be in the app. These were Basement, Kitchen and Exterior Quality. Other features could be changed with significant investment: Baths (can add a bath), Wood Deck (can add a deck or expand one), and even house size (since you can — if you’re crazy and rich — tack on an addition somewhere).

I also included a couple Y/N features since they could affect the price: Garage and Central Air.

Plotly Dash

I knew Plotly as a way to create interactive plots on Python, but the Dash framework was new to me. It was introduced to me by one of our instructors and an active online community and examples also came to my aid.

Dash is similar to R/Shiny in that it has two key components: a front-end layout definer, and a back-end interactivity piece. These can be combined into the same file or separated (as they can in Shiny).

LAYOUT

The general pseudo-code outline of a Dash App starts with importing the libraries and components. It’s written in Python, so you’ll need numpy and pandas as well as Plotly components for the graphs.

Dash converts HTML tags into HTML when they appear after “html.”, which is straightforward. I included Div tags, H2 (headers) and P (paragraph) tags. Different sections of the Div are called “children” (as they are in HTML). I used “children” here because I wanted to have two columns — one for the inputs and the second for the graph of adjusted house prices.

The rows and columns can be done in a couple different ways, but I used simple CSS style sheet layouts (indicated by “className”).

FUNCTIONALITY

Interactivity is enabled by the layout and the callbacks. In the layout, as in Shiny, I could specify different types of inputs such as sliders and radio buttons. There aren’t as many or as attractive a selection as you find in Shiny, but they get the job done. I used sliders and buttons, as well as input for the avg starting price.

The functional part of the app starts with “@app.callback,” followed by a list of Outputs (such as figures/charts to update) and Inputs (from the layout above). The way Dash works, these Inputs are continually monitored, as via a listener, so that any change to a slider or other input will immediately run through the function and change the Outputs.

Right after the @app.callback section, there’s one or more functions. The parameters of the function represent the Inputs from @app.callback, but they can have any name. The name doesn’t seem to affect their order, since they’re read in the same order as the callback Input list.

This function includes any calculations you need to do on the Inputs to arrive at the Outputs, and if the Output is a graph or plot — as it almost always is — there’s also a plot-rendering piece.

In my case, I wanted the output to be a normal distribution of potential new values for the sale price, given the changes in the Inputs. For example, if someone moved the “Overall Quality” slider up a notch, I needed to generate a new average saleprice, a difference (from the baseline) and a normal distribution to reflect the range.

I did this by turning each coefficient into a factor, adding up the factors and multiplying by current sale price. I then generated a random normal distribution using Numpy with the new target as the mean and a standard deviation based on the original model.

The final dashboard looked like this:

CONCLUSION

Granted, there are a caveats around the app. It’s based on the Ames housing dataset, so other areas of the country at different times would see different results. It requires estimating a ‘starting price’ that assumes the default values are true, and this estimate might be difficult to produce. But as a proof of concept, it has potential.

I think there’s definitely a data-driven way to estimate the value of some common improvements. This would help a lot of us who are thinking of selling into the housing boom and wondering how much we can reasonably spend on a kitchen remodel before it turns into red ink.

The following article first appeared on the Salesforce Marketing blog in January.

Last May, at the height of the COVID pandemic’s first wave, PepsiCo raised some bubbles by launching not one but two websites where consumers could browse and purchase a selection of the company’s more than 100 widely munched snack and beverage brands. At a time when many outlets were closed and people ordered more food online, the company quickly made it easy for consumers to buy directly.

One ecommerce site, PantryShop.com, took a lifestyle-themed approach, offering bundles of products in categories such as “Rise & Shine” (Tropicana juice, Quaker oatmeal, Life cereal) and “Workout & Recovery” (Gatorade, Muscle Milk, and Propel electrolyte-infused water). A second site, Snacks.com, offered a more straightforward lineup of Frito-Lay brands such as Lay’s, Tostitos, Cheetos, dips, nuts, and crackers.

These platforms complement PepsiCo’s retailer channels to ensure the company continues to deliver consumers their favorite products on the right platform, in the right place, at the right time. 

Whenever, wherever, however has been the mantra of ecommerce digital marketing for a while, but it’s become more important in the age of COVID-19.

Whenever, wherever, however has been the mantra of ecommerce digital marketing for a while, but it’s become more important in the age of COVID-19. Most of Salesforce’s customers – particularly those in high-velocity consumer categories such as consumer packaged goods (CPG), retail, and restaurants – have had to become extraordinarily flexible over the past 10 months to adapt to ever-changing local guidelines and consumer behaviors.  

What was most striking about PepsiCo’s foray into direct-to-consumer (D2C) commerce and marketing was its speed. “We went from concept to launch in 30 days,” said Mike Scafidi, PepsiCo’s Global Head of Marketing Technology and AdTech. “Within 30 days, we stood up our ecommerce capabilities and started delivering direct to our consumers.”

PepsiCo’s products are consumed more than a billion times daily in 200 countries. It boasts 23 brands with more than $1 billion in annual sales. How does such a large and complex global company pull off such impressive footwork?

The answer, Scafidi said, was preparation. “Digital is inherently disruptive,” he explained. “We’ve been training for this for a long time. We’ve been preparing to adapt to disruption for 20 years.”

Planning for change is a skill

Scafidi and I spoke from our remote locations about “Building Resilience and Adapting Your Marketing Tech in Uncertain Times” during Ad Age’s CMO Next conference, co-hosted by Ad Age’s Heidi Waldusky. Scafidi stressed that tumultuous times took PepsiCo back to basics — inspiring the company to lean on skills it had been developing for years — especially in consumer research and media measurement.

Part of the reason PepsiCo was able to launch Snacks.com and PantryShop.com so quickly, he said, was “we were leaning on what we were doing already.”

He reminded me of an analyst quote I read recently on embedding resilience into sales and marketing plans: “[T]he more an organization practices resilience, the more resilient it becomes.”

Over the past year, many organizations have had plenty of time to practice being resilient. As stores shut down and millions huddled at home, there was a surge in digital activity across all channels. Media consumption soared: people around the world watched 60% more video, for example. And they shopped. Salesforce’s latest Shopping Index shows that comparable online sales were up 55% in Q3 of last year after climbing 71% in Q2.

We’ve heard from many of our customers that they needed to launch new capabilities faster than ever before. Otherwise, they’d lose business. Curbside pickup, buy online, pickup in store, expanded digital storefronts, appointment scheduling, contact tracing – the list goes on.

Our desire to help customers adapt to rapid digitization inspired us to launch Digital 360, a suite of ecommerce digital marketing products combining marketing, commerce, and personalized experiences under a single umbrella. With it, Salesforce Trailblazers like Spalding and Sonos were able to scale their online commerce dramatically, making up some of the shortfall in brick-and-mortar sales.

When times are changing, it’s too late to build up basic skills. If you have a foundation in place, that allows you to adapt.

MIKE SCAFIDI, PEPSICO GLOBAL HEAD OF MARKETING TECHNOLOGY AND ADTECH

Unilever also faced dramatic market shifts in the recent past. Keith Weed, the company’s chief marketing and communications officer, pointed out back in 2018 that the pace of change “will never be this slow again” – not knowing just how fast that pace would get. And like PepsiCo, Unilever met the hyperfast present by relying even more on its customer research skills.

“We know that people search [online] for problems, not products,” Weed said. So the company created Cleanipedia.com, which offers detailed solutions to cleaning problems in 26 languages. Built before COVID-19, the site was ahead of its time and has attracted 28 million visitors to date.

Building a foundation that scales to meet customers wherever they are

When times are changing, it’s too late to build up basic skills. “If you have a foundation in place, that allows you to adapt,” Scafidi said.

For example, the PepsiCo team was able to rapidly restructure its internal media measurement analyses because it had already put in the work to develop an ROI Engine, which helped determine the real impact of its advertising, promotions, and email. The ROI Engine automates data inputs, processing, and algorithms to improve paid media optimization decisions. Combining the ROI Engine with a customer insight capability called Consumer DNA, “We were able to stabilize our understanding of the consumer and adapt to where they were,” Scafidi explained.

PepsiCo’s Consumer DNA project is an example of a custom-built tool that allows the company to gain a 360-degree view of the consumer to enable more effective and targeted media buying and marketing activation.

At Salesforce, we help our customers engage with their customers. To do this in 2020, we too relied on core skills, built up over years, to adapt to an environment that seemed to change by the second. The result was launches like Digital 360 and Work.com. The latter helps companies safely reopen their workplaces. We also introduced a customer data platform (CDP) called Customer 360 Audiences, which serves as a single source of truth to build a unified profile for marketers and others.

The Ancient Greek philosopher Heraclitus said, “the only constant in life is change.” As customers like PepsiCo show us, the best way to adapt is to build core skills that can help you pivot quickly in the future.

Happy to announce that the book I co-wrote with Chris O’Hara on Customer Data Platforms has just been published by Wiley and is now available at fine booksellers everywhere such as this one.

Helpfully, the title is “Customer Data Platforms: Use People Data to Transform the Future of Marketing Engagement“.

Note the key phrase: Customer Data Platforms (CDP). Only the hottest mar-tech category to appear in at least a decade, and we literally wrote the book on it. At least, the first substantial mainstream book on this key topic. We cover the category from data integration and identity management to exploration, activation and A.I.

It’s accessible and a quick read, full of charming illustrations and architecture drawings. If you’re interested in this fast-growing tech category that points the way toward the converged platform future, our book is a great place to start.

Happy launch-day, my friends! Here’s to a more integrated, privacy-friendly and analytical future! The future is bright. peace mk

The following column originally appeared in the mighty AdExchanger on 11/10/20

Measurement is a footnote – outglamorized by targeting and the opera of browsers, pushed into the corner during debates about the future of media. But it’s arguably more important than aim and requires more discipline.

A few weeks ago, the World Federation of Advertisers (WFA) released a statement and a technical proposal for a cross-media measurement framework. It was the outcome of a yearlong global peer review and an even longer discussion among participants including Google, Facebook, Twitter, holding companies, industry groups such as the ANA and MRC and large advertisers including Unilever, P&G and PepsiCo.

Reactions ranged from enthusiastic to less so, but few people seem to have read more than the press release. After all, it’s not a product and could be yet another in a parade of grand ambitions in online-offline media measurement, dating back to Facebook’s Atlas.

But it describes a realistic scenario for the future of measurement. Sketchy in spots, the WFA’s proposal is ironically the clearest screed of its kind and is worth a closer look.

To be sure, this is a project focused on solving a particular problem: measuring the reach and frequency of campaigns that run on both linear TV and digital channels, including Facebook, YouTube and CTV. In other words, the kinds of campaigns that cost participating advertisers such as P&G a reported $900 billion a year.

And P&G’s own Marc Pritchard is on record calling the proposal “a positive step in the right direction.”

The need is clear. Advertisers today rely on a patchwork of self-reported results from large publishers, ad server log files, aggregate TV ratings data and their own homegrown models to try to triangulate how many different people saw their ads (reach), how often (frequency) and how well those ads fueled desired outcomes, such as sales lift.

The latter goal is acknowledged in the current proposal, which doesn’t try to solve it. But the WFA, building on previous work from the United Kingdom’s ISBA, Google, the MRC and others, lays out a multi-front assault on reach and frequency that covers a lot of ground.

How does it work?

The proposal combines a user panel with census data provided by participating publishers and broadcasters, as well as a neutral third-party data processor. The technical proposal spends some time talking about various “virtual IDs” and advanced modeling processes that are loosely defined – and the goal of which is to provide a way for platforms that don’t share a common ID to piece together a version of one.

Needless to say, a lot of the virtualizing and modeling and aggregating in the WFA’s workflow exists to secure user-level data. It’s a privacy-protection regime. It also engages with the much-discussed third-party cookieless future.

Panel of truth

The proposal leans heavily on a single-source panel of opted-in users. At one point, it calls this panel the “arbiter of truth,” and it’s clear most of the hard work is done here. Panelists agree to have an (unnamed) measurement provider track their media consumption online and offline. Panels are a workhorse of media measurement as provided by Nielsen and others, but they are expensive to recruit and maintain. It’s not clear who would build or fund this one.

In the past, other panels have struggled to collect certain kinds of cross-device data, particularly from mobile apps. Panels also get less reliable in regions or publishers where they have less coverage, a problem that could be addressed by joining multiple panels together.

In addition to the media consumption, demographic and attitudinal data it provides, the panel is used to “calibrate and adjust” much more detailed census data voluntarily provided by publishers (including broadcasters).

Publisher-provided data

No walls here – at least in theory. Given that Google and Facebook support the WFA’s proposal, it’s implied they’re open to some form of data sharing. It’s already been reported – although is not in the proposal itself – that some participants will only share aggregated data, but it’s better than nothing. The WFA’s idea of “census data” includes publisher log files, TV operator data and information collected from set-top boxes.

This census data is married at the person-level with the panel data using a series of undefined “double-blind joins of census log data with panelist sessions.” Joined together, the different data sets can correct one another: The panel fills gaps where there is no census data, and the more detailed census data can adjust the panel’s output.

Virtual ID’s, anyone?

The census data will have to be freely provided, and so wide-ranging participation across many publishers is required for success. Another requirement is a way to tie impressions that occur on different publishers (which don’t share a common ID, remember) to individuals to calculate unduplicated reach and frequency.

In a general way, the proposal describes a process of assigning a “Virtual ID” (VID) to every impression. This VID may – but may not – denote a unique individual. How is it assigned? Based on a publisher-specific model that is refreshed periodically and provided by the neutral measurement provider. It appears to use cookies (and other data) in its first version, graduating to a cookieless solution based on publisher first-party data in the future.

The output here is a pseudonymized log file with a VID attached to each impression, overlaid with demographic data – at least TV-style age and gender cohorts – extrapolated from the panel.

Doing the math

In the final step, each individual publisher will perform some kind of aggregation into “sketches.” These sketches are likely groups of VIDs that belong to the same demographic or interest segment, by campaign. And it is worth noting here that the “sketches” can’t be reidentified to individuals and are somewhat similar to proposals in Google’s “privacy sandbox.”

At the penultimate step, each individual publisher sends their “sketches” to an unnamed independent service that will “combine and deduplicate VIDs” to provide an estimate of reach and frequency across the campaign. The WFA has a proposal for this Private Reach & Frequency Estimator posted on GitHub.

A GitHub explainer mentioning data structures and count vector algorithms is ad tech’s new sign of sincerity.

Finally, outputs are provided via APIs and dashboards, which support both reporting and media planning. End to end, it’s an ambitious proposal that has many of the right players and pieces to work. Its next steps are validation and feasibility testing led by the ISBA in the United Kingdom and the ANA in the United States.

Whatever happens, we’ve learned something from the WFA’s proposal. Even in a best-case-scenario, accurate global campaign measurement will definitely require heroic levels of cooperation.