Prediction to Perfection: Can’t Get There From Here

By Michael Atkin

Recently there has been intense focus applied to the issues of brand safety in digital advertising. Major advertisers and several media agencies are suspending activity on certain platforms until they can have absolute assurance that their ads will not run adjacent to violent, offensive, terroristic or pornographic content. Due to the gigantic scale of potential content and audience targeting options in digital media’s long tail, there is much discussion of the development of filtering mechanisms that are exclusively machine based. These filters would use the kinds of predictive algorithms that have been so important in realizing the substantial ROI gains made through audience targeting in digital media. However, I believe a totally algorithmic solution, without human intervention or curation, is unlikely. Here’s why.

This problem is different

Using demographic, psychographic, behavioral, intentional, semantic and other data signals, machine processed algorithms using predictive analytic techniques have shown enormous power in refining audience segments for commercial message targeting. And this has produced significant ROI improvements for advertisers. So the thinking goes, just as specific audience prospects can be identified in real time in the vast array of digital audience opportunities, couldn’t this also apply to specific media resources in the long tail, to filter out offensive content adjacencies? If we could just interpret the right set of data signals, we could programmatically exclude objectionable content on the fly, just as we programmatically bid for the right audiences at the lowest cpm. And avoid human intervention with its related costs. However, at its root the brand safety problem is not the same.

I think our industry has done itself a disservice in allowing sloppy slogans to popularize the idea that when we target using predictive analytics, we can actually target at the individual customer or prospect level. But that’s not the case. No matter how complex the statistical process, it’s still just probability. The use of predictive analytics has always been designed for the improvement of the average ROI – and while we see striking improvements in that measure, it’s not at all the same thing as pinpointing only the individuals we’re interested in and excluding all others. But that’s the real nature of the problem we face in the brand safety area – an appearance on an individual media vehicle – not an average. Predictive algorithms cannot guarantee absolute brand safety in real time. And after the fact audits are not going to be helpful for a client whose brand reputation is precious.

The social media amplifier

The reason for this is that we are seeing a kind of reversal of the concept of “publishing”. In the past, the appearance of an advertiser’s commercial next to objectionable content certainly took place, but in effect the problem was hidden. Individuals could complain to an advertiser, and the advertiser might or might not adjust its schedule, depending on their own assessment of the situation. In certain cases, individuals would contact pressure groups who could make more collective noise, and those cases would often result in the advertiser changing its media selection. But now, each individual can and will publish their experience and their outrage online. And that outrage is instantly taken up and amplified in social media – which not surprisingly obey the age-old maxim “if it bleeds, it leads”. So the advertiser is faced with a much more ferocious and damaging public relations problem if they allow even a single instance of their message appearing alongside egregiously bad content. The situation has outpaced a purely statistical solution.

Implications

I think the implications for advertisers, media agencies and publishers are the following:

  • Machine algorithms alone can never provide a satisfactory degree of brand safety
  • The solution will involve some combination of active content curation, whitelisting/green zone content pools, and human monitoring
  • The long tail of advertisable environments will get shorter
  • Publisher reputations will reacquire their value
  • The real costs of brand-accountable digital advertising will become visible, acknowledged, and included in agency scopes of work

For clients who value their brand reputations, it appears certain that a solution to brand safety in digital media must involve a combination of both advanced statistical procedures and hands on human action.