Presenter: Jay Friedman, Goodway Group, Partner & COO
We were told that programmatic was so advanced, the algorithms would optimize the media for us. What happened? In this session, Jay Friedman talks through how in the past, the promise of programmatic has fallen short, but more importantly, he talks through why programmatic’s future is so bright. Advancements in data science, people-based marketing and all channels moving towards being programmatically traded give us the programmatic vision of the future.
9. 5 ways to make your programmatic
campaign better…Right now.
Create bid
expressiveness
Optimize your
supply path
Non-gameable,
meaningful metrics
Services must
pay for
themselves
Media
consolidation
Promise of programmatic (fields of flowers)
The future is programmatic – all of it – OOH, broadcast, everything
How do we get there? How do we reconcile? Most importantly, how do we get our media to work and deliver real ROI. Whether programmatic has failed you or worked well, today we’ll discuss specifics you can take back to improve how your campaigns perform, and why you can look forward to all media being bought programmatically.
But before we get into this sunny future, we need to talk about the elephant in the room and overcoming some negative feelings you may have around programmatic.
[depending on the vibe of the room, you can ask them to shout-out programmatic word association, or do show of hands who has positive feelings about programmatic, etc.]
You probably have these associations because there has been a long line of vendors knocking on your door, promising amazing results but delivering on the good end, sub-par performance, and on the bad end, straight-up fraudulent results.
And if you’ve had this kind of experience, the snake oil that everyone was selling was an algorithm. Who here really knows what an algorithm is? [invite audience to engage] An algorithm is a set of rules a computer uses for solving problems, and it usually involves math and calculations. Remember when you were a kid and you tried to send notes in secret code to your friend where instead of using the letter A, you used the number 1; and instead of using the letter B, you used the number 2…and so on. That is a really basic example of an algorithm – when you want to use this letter, use a corresponding number instead. And since computers and ad servers are so good at data processing, we can implement extremely complex algorithms and apply them to your media buy.
To illustrate a more relevant example of an algorithm in adtech and what it takes to build a good one, I’d like to talk through how we approached building an algorithm using the Time in View metric and it’s impact on conversions. Time in view is simply the measurement of how many seconds an ad was measured as viewable to the user. This is an example of what’s required to do original data analysis and develop algorithms to help optimize your campaigns. This is also the depth needed to interrogate a vendor about their methods to know what you’re really getting.
How would you go about determining the impact TiV has on conversion? Like anyone, we started simply. We took all of our converting users and matched them to user impressions with TiV data to see if longer TiV created more conversions.
Problem #1 – That Initial analysis showed absolutely no benefit to time in view. But that’s impossible, right? Well, users have varying amounts of impressions leading up to a conversion, so 2 users might each have 20 seconds in view, but one has 2 of 10 seconds and one has 10 of 2 seconds. So we had to control for frequency, analyzing every converting user based on the frequency bucket they fell in.
Problem #2 – Still no correlation. But wait, re-targeted users are inherently different from prospecting users. And contextually targeted users are inherently different from behaviorally targeted users. So we controlled for user type.
Problem #3 – Still no correlation. Different devices measure differently. Let’s control for device.
Problem #4 – every advertiser we run has different types of conversions and different creative. Ok, let’s not look at the data in bulk and control for advertiser and creative treatment. Nothing obvious.
Problem #5 – We’re using basic analysis methods. Let’s try something more advanced like bootstrap aggregation or multiple regression types. Wait, yes, there is something.
Problem #6 – What we found requires multiple models and is essentially unexplainable to a client. Like you need a master’s in math and statistics. How are we going to communicate the value?
My point in telling you all of this is that building a good adtech algorithm is a long process of testing and learning different sets of problem-solving rules across infinite combinations of campaign variables, and often explaining the results of the testing to a general audience is not easy, especially when trying to build in nuanced quality metrics.
This example also explains why a programmatic vendor’s magic algorithm may not have worked the magic you hoped for. The simplest algorithms to build and explain to a client simply optimize to the common set of metrics that you all are most familiar with – CTR or eCPA, but this is like prescribing pepto bismol for every illness. Most times it’s wrong, and it’s only right when it’s not that serious. In reality though, each campaign goal requires a different algo – CTR, eCPA, brand awareness/lift, leads, purchase, foot traffic. The algos don’t and can’t work the same. Most platforms don’t build for all scenarios and instead build for purchase and CTR.
When platforms build algorithms for purchase and CTR, there are numerous problems that manifest. First, most of these algos are built using last-click purchase and CTR, which games the entire system giving you misleading results on the surface with little real-life business impact. Last-click attribution is a game where you organize your media around users who were going to click or purchase anyway before ever seeing your ad.
The other major problem that pops-up with algos has to do with inventory quality. There is a high correlation among low prices, high CTRs and fraud or suspicious inventory. If you’re working with a platform that builds its algos without accounting for inventory quality, then you’re going to get results that look great on the surface (low prices! Low cost pers! High CTR) but deliver minimal business impact.
No algo works for every campaign in its intended set. Even if every campaign is of similar nature, at least some campaigns will experience worse performance. Explain positive/neutral/negative.
But we can’t look at it as, “some campaigns will be worse.” We need to look at campaigns as having no performance, and then choose the non-algo performance or the algo performance. No one would choose the non-algo. Plus, if it makes it worse, turn it off!
Let’s pause for a minute to address something you may be thinking, and something I hear often. “Why does this have to be so difficult?” This is what happens to every industry. Think about the people who made the first tires. “Make it out of rubber and make it round!” and that was genius. Now you need a PhD in chemistry to innovate a tire. Advertising just took a LONG time to get more sophisticated. Of course, just like we have the option of buying bad tires, any of us can still buy each site directly via an IO and have most of our money run off-target.
What we are talking about isn’t easy, but At the same time, we should be ecstatic that we’re so early on in the complication because this means there is still significant greenfield for innovation. Think about how much media is NOT programmatic yet. Think about how new cross-device targeting is, and people-based marketing. Those who embrace and tackle the difficulty have huge upside. This is the 1984 of the personal computer industry and there is still so much unsolved.
So this is the question, are you someone who is going to solve and innovate with data science and modeling, or someone who will wait? There isn’t a wrong answer to this question. But this is a choice agencies and marketers need to make right now, and I think this reality is often skipped over in favor of a vendor who says, “Don’t worry, we make it easy!”
The reality is, it’s hard. It’s not impossible but takes a team and resources – you need data scientists visibility into thousands of campaigns across multiple verticals, with different goals, etc. Regardless of whether or not you want to go to this level, let’s cover 5 ways we can make campaigns optimize better right now. And let’s make it a goal to make at least 3 of these be thing we hadn’t considered before so we’re delivering real value to you today.
1---- Experiment with and create bid expressiveness. Most traders have less than 10 unique bids per ad group. So, within an ad group or targeting type we often see less than 10 unique bid values being submitted to exchanges. The world is no longer a 2nd price auction and there is a very good chance publishers are taking advantage of those leaving bids fixed and less expressive. Great traders may have 100 per targeting type. Great algos achieve 1,000+
2---- Optimize your supply path. The same site can be bought from 10 different exchanges. First, how often are those legit impressions that site, not domain spoofed or arbitraged? [tell Business Insider story] But beyond that, you’re spending different CPMs with each exchange and you can optimize this to save 5-10% and improve performance at the same time. This can be done manually or algorithmically.
3---- Pick a non-gameable, custom metric that is actually meaningful to your business. Lots of people pick VCR, CTR, and viewability. Those are measures of quality and engagement but not end metrics, and they’re incredibly gameable by bots. Metrics such as foot traffic or offline sales lift are good examples of meaningful business-oriented metrics. This will reduce your fraud because you’ll naturally optimize to what’s working, and what’s working won’t be something bots can figure out.
4--- Only pay for tech/non-working media dollars that more than pay for themselves. Things like pre-bid fraud prevention that cost 2% of media are only valuable if they eliminate more than 2% of fraud. Don’t pay for something that’s 2% of media but eliminates 1% of a problem. Same with reporting and analytics. Before you pay $.05 for something, know what you’re going to get out of it and make sure the results more than pay for themselves.
5--- Consolidate your media buying with a single partner – the status quo at agencies, even the obligation, is to test and run with multiple vendors to see who does the best. This inevitably sets up the vendors to play a game of winning the last-click, which as we discussed in the beginning, is a detrimental strategy to being successful in digital. You lose control of frequency, you give each vendor less data to optimize with, and you parse out to more people in the industry confidential intelligence about your brand, which may ultimately end up with your competitors. Any algorithm is going to have a better chance of actually helping your campaign if you give it more data to work with and a clear goal.
Close out
Programmatic’s future is bright because it solves the problems of today’s marketers who have to navigate in and succeed by using a complex set of data across multiple screens, channels and media properties. And the problems of programmatic, i.e. the snake oil, are indeed solvable. When you lay out your programmatic strategy – whether that is to do it yourself or work with a partner or platform, you should not settle for a solution that doesn’t solve these challenges.