How to Balance Search Ad Best Practices & Machine Learning Innovation

We sure have many ad options these days.

And many of us still rely on the old standard ads because they continue to outperform the new kids on the paid search advertising block.

How do you get your arms around this, you wonder, with a scalable strategy that supports broader targeting and testing objectives?

To make things even more complex, support from third-party management platforms has lagged, making it difficult to adopt changes amidst excitement from everyone in the industry about new ad options.

Scalable Solutions Require Scalable Tool Adoption

If the tools you rely on to track performance do not support full visibility into results, down to conversion level for each ad, you won’t be able to accurately determine why performance ends up the way it does.

Even if aggregate performance may improve, it’s important for your team to ask the hard questions.

What is more important:

  • Is it testing something new, perhaps in low sensitivity campaigns?
  • Is it to have a full understanding of new tools to know how to scale it?

If it’s the latter, have the conversations to reset expectations of how latest ad developments are not fully ready to be rolled out.

No one will want to be in a meeting in a couple months with pointed questions “why did this succeed or fail” and unable to answer.

Plus, your team’s limited resources can be prioritized to more actionable short-term items.

That said, regardless of readiness to move forward with adoption, we should step back to recognize a wider shift underlying the new ad developments.

Treating this as just another ad rewriting exercise misses the larger shift in user targeting happening at Google and search at large.

The new format specifics are there support the industries relentless march towards automation.

Beyond the new character limits, embracing automation is the real stake here.

For well over a year, Google is making a concerted effort to roll out machine learning elements across more and more components of its ecosystem.

It is easy to perceive this as Google taking away control from advertisers. Yes, automation means less hands-on management and often less predictable outcomes.

On the flip side, these SEM innovations will make your campaign assets – keywords, ads, bids, negatives – work smarter, not just harder.

The Case for Testing RSAs

Initial numbers observed across my early adopter clients support the promise of how automation improves performance.

While RSAs are seeing lower impression share than other ads, their CVR and ROI is far exceeding ETAs.

Below is sample data from a campaign that ran for a month, with full impression share and no bid changes or impact from seasonality.

While RSAs accounted for only 3 percent of impressions, thanks to much greater CTR and CVR they drove disproportionately more conversions at much better ROI.

Row Labels



RSA Share of Activity

Impr.  154,423  5,028 3%
Clicks  1,919 766 29%
Cost $1,665 $400 19%
Conversions  26  14 35%
Revenue $37,410 $16,556 31%


CTR 1% 15% 1,126%
CVR 1.4% 1.8% 35%
ROI $22 $41 84%

This is representative of observations seen across all areas of where RSAs were adopted.

With 80 percent significance in the above test, one can argue that, statistically speaking, a bit more testing is needed for a rock-solid case.

Still, such notable differences in results and pretty high significance make a good case for adopting machine learning.

In fact, rather than waiting longer for a few tests to get perfect significant results, you should launch more tests so that collectively more data is collected across more areas.

Yes, that means taking a greater leap of faith.

However, that will also mean gathering data faster across more areas and ultimately building a more solid case than if fewer isolated tests were running.

Recommended RSA Best Practices

On a related note, to accelerate machine learning, do provide a healthy number of ad components so that machine learning can quickly test various combinations, decide what works, and hone in on what does.

A minimum of three headlines and two descriptions are needed (i.e., six ad combinations).

Just by adding one more headline or description, one would go to 9 variations, giving the AI multivariate technology 50 percent more ads versions to try.

At current reporting stage, Google Ads only provides impression counts only for each headline and description. One will not know what element exactly contributed to what extent to the performance.

As much as that is frustrating, this is also not something one can control. And that is the trade off with AI – ROI and CVR gains come with a trade-off for control and visibility into “how the sausage is made.”

It is also important to revisit ad extensions in this process.

With the need to come up with more ad text fields, there is a risk of repeating what is already running in ad extensions: snippets, callouts, sitelinks.

This is an opportunity to level up your copy strategy and rethink more generally ad messaging for search.

Rethink You Ad Messaging SEM Strategy

Until now, extensions were often treated as add-ons, with the primary focus on the ad text.

Longer descriptions over the years were used as chances to stuff more and more things making them richer but also more cumbersome.

Machine learning now gives a chance to adopt a more strategic approach which aligns better with a brand’s wider marketing objectives.

Below are suggested guidelines on how to streamline messaging to make learning more actionable and easier to integrate with wider organizational objectives

  • Headlines 1 for Brand or Product reference.
  • Headlines 2 or 3 for key call to action or brand positioning.
  • Description text for general statements highlighting benefits for the target audience. Ideally, each variation will target a different user group such as price conscious users, those valuing a certain experience, etc.
  • Maximize ad extensions for specific product features and attributes.

And then sit back relax of fasten your seat belt whichever is appropriate for the team.

Let AI do its magic.

Personally, I am in the “fasten your seat belt category,” though in the passenger.

As much as it is important to closely monitor results, RSA machine learning needs a proper chance to work and it is important to let it be in the driving seat.

More Resources: