Is Attribution The Solution To Ad Fraud?

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

Today’s column is written by Jim Spaeth, partner at Sequent Partners.

I’ve heard anecdotally that somewhere between 50% and 70% of digital advertising spend is wasted, completely ineffectual. Fraudulent impressions could account for 30% to 50% of all digital impressions.

Significant questions exist about marketers’ ability to know whether their campaigns are actually driving sales. That is why attribution – the ability to measure advertising and assess how it impacts the bottom line and reaches intended consumers – is the single most emerging critical topic in the industry today.

But can it be the savior for marketers in getting rid of ad fraud?

Marketers use attribution models to avoid underperforming impressions. In concept, that means fraudulent impressions should be identified as poor performers and their source dropped from the buy.

So, does that mean attribution is the solution to ad fraud?

Unfortunately, it’s more complicated than that.

Attribution treats every impression as if it’s valid and seeks to determine its incremental effect. Obviously invalid impressions from bots or click fraud don’t have an effect because they’re not viewed by a person. They’d be scored zero in an attribution model, but so would a valid impression that is delivered to the wrong target audience, not viewed or whose message doesn’t connect with the consumer.

Some of these problems related to poor targeting, media selection or creative strategy can be remedied. But there is no remedy for fraudulent impressions. Should advertisers just throw out all the underperforming impressions? Should they walk away from valid but underperforming media and throw the baby out with the bathwater?

By simply rejecting underperforming impressions, the advertiser misses the opportunity to learn if the creative is the issue. By understanding which creative executions are performing better than others, the advertiser can optimize the creative mix and lift ROI materially. Simply rejecting all underperforming impressions also misses the opportunity to analyze and refine targeting strategies or media selection.

Is there some way that attribution could distinguish between fraudulent and other underperforming impressions?

Yes, but this seemingly simple solution requires a high degree of granularity in the attribution analysis.

Fraudulent impressions are likely to be bundled with legitimate impressions. This will reduce the efficacy of the bundle and maybe, as a result, they are dropped from the buy. And maybe a source of remarkably underperforming impressions could be identified for further investigation to reveal the fraudulent source.

That’s definitely a “maybe,” though. Like separating wheat from chaff, this would require a continuous sifting that may not be possible in the time available or not even possible at all.

If anyone is still using last-click attribution, the inability to distinguish between fraudulent and valid conversions is downright dangerous. If the last click viewed is fraudulent, the programmatic platform will optimize spending toward the sources of that fraud.

So, is attribution the solution to ad fraud?

If we can assume the attribution work is being done with appropriate statistical models, at a fine level of granularity, by a thoughtful analyst, the answer is “definitely maybe.”

Follow AdExchanger (@adexchanger) on Twitter.

1 Comment

  1. Bob Sacamano

    Interesting viewpoints, thank you for sharing.

    I don't believe attribution alone will address ad fraud, because attribution models are generally applied for conversion-optimized campaigns, which are a subset (albeit a large subset) of the whole. In other words, a lot of ad dollars are spent driving video completions, which I don't believe are a common target for attribution vendors.

    Additionally, last-view IS an attribution model, and has done nothing to reduce fraud, even though fraudulent impressions are worthless in that model (and any other, as you suggested).

    And finally, regarding last-click. If the conversion is a sale, or otherwise not something a bot can do, I don't believe there is risk of an algorithm heavying up on bots any more than any other attribution model. I agree that last-click is an outdated metric but plenty of serious advertisers, especially banks and credit cards, are still using it for some reason.

    Without dismissing the severity of fraud, which needs to be eliminated, I feel there are larger problems surrounding attribution and outcome measurement in our industry. Advertisers expect sales results from a campaign in the extreme short term, often in days. "You've been live for two days, you’re driving lots of conversions, why haven't sales increased?" My honest belief is that most campaigns don't alter human behavior in the short term, and if they did, it's impossible to measure the impact of digital advertising relative to other variables that can't be measured, like weather, radio advertising, or "my grandfather always drove Fords, and so do I". This is true whether you’re using last-touch, regression-based multi-touch, or something else. Digital advertising creates more data than any other type, which (reasonably) leads people to want to use it to understand outcomes. However digital data alone will never be able to tell a consistent sales story.

    Digital advertising is a critical channel, probably the most important channel available to advertisers, but I believe this desire for ultra-short-term results is responsible for a lot of the problem areas in our industry. For example, if I need to deliver 100M impressions, perhaps only 1M will drive a conversion, leaving 99M impressions targeted to less-valuable users that probably weren’t necessary in the first place. THIS is where campaigns are vulnerable to bots—the “show an impression because I have to” ads that aren’t expected to have a big impact anyway. But if I tried to sell 1M impressions to an advertiser, they’d laugh at me.


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>