Today I am writing about machine learning in search – a hot topic and debate these days. I expect this to be one of several posts on the topic.  This particular writing intends to summarize both sides.

On the “Pro” side, Google is preaching that their new machine learning algorithms replace the need for keyword creation and manual bidding.  Additionally, Responsive Search Ads (RSA) negate the need to write ad copy.  With the elimination of this busy work, SEMs should now have more time back in their day which should be replaced with strategic input. If not, they lose their value to their clients.

On the “Con” side, having just attended SMX Advanced in Seattle, there is reluctance on the part of SEMs to adopt machine learning as aggressively as Google wants them to. Essentially, this debate comes down to those who are in the weeds vs. those who are not as well as Google who is pushing the agenda.

The first session I attended was more or less an open-mic “what are the hot SEM issues out there” type of thing. The moderator quickly steered the conversation to machine learning. The responses fell into one of three camps. Note that on subsequent days of the conference, there was some excellent content on how to use machine learning strategically, but this initial session provided an early view on the different sides of the debate.  

1. The writers in the industry who understand the debate but not the weeds.
2. Advertisers who fault their agencies for steering them away from machine learning rather than adopting it. Again, this is largely a not an in-the-weeds group.
3. The SEMS responsible for the performance of search campaigns who seemed to be largely in agreement with each other on some of the pitfalls of machine learning. Here are just two examples:

• About a year or so ago, App Campaigns were Google’s first foray into a 100% machine learning campaign type, and their performance is/has been mediocre at best. Like Google’s new “Smart” campaign types, SEMs have no levers to improve UAC performance.

• Responsive Search Ads are not only clunky and somewhat mind-numbing to set up, they serve weird and are seldom top performers vs. static copy.  

For the record, I have seen some of the new automated machine-learning strategies work well and consider them a good asset for the toolbox in order to improve an account’s performance.  To universally adopt them, though, would be nuts. As one Google exec said at SMX, in a year, these strategies should be exponentially improved vs. where they are now.  

Because even Google is saying machine learning has not been perfected yet, everyone should relax and move into testing mode rather than firmly jumping in one way or the other.

And why nuts to universally adopt them now? Here is just one example. One of our agency’s initial forays into Smart Shopping went well. The campaign was significantly outperforming regular PLA campaigns that were highly segmented and loaded up with years of optimizations. We overlooked the fact that we can’t read Remarketing from Gmail from Shopping performance as well as the fact that our display creative rendered oddly in some ad units such as the high-impression-volume 300×250.  

But for some reason, the wheels fell off and impressions pretty much disappeared. Could it be that another Smart Shopping advertiser joined the auction, and our client got booted?  Being that it’s a Smart campaign, we are running blind here, and Google can’t provide any insight, either. What if this happened at the height of Holiday?

There are other hot topics around this issue such as what happens if all of your competitors launch Smart campaigns, too.  Who gets the visibility? The one with the biggest budget which other than the feed is the only other lever at an SEM’s disposal?

Another question is what happens to the Marins, Kenshoos, and SA360s? We are reasonably sure that Google (at least for now) gives preference to machine learning campaign types in terms of top visibility. The bid management platforms facilitate automated bidding but they certainly don’t provide any clout aka visibility. It is unclear what the purpose of these platforms becomes other than amped-up reporting and, in the case of Marin and Kenshoo, bid management for other media platforms.

While Google says the SEMs newfound time should shift to strategy, it is more likely that, at least initially, it will go toward analytics to determine if these strategies are helping – or hurting – our businesses. We are already seeing studies on the performance of match type variants and should expect to see more like that.  

Lastly, lest this post imply that all of our results with machine learning have been negative, that is definitely not the case.  We will report on some interesting successes in future posts.