Google Search Ads: I Fired the AI (Twice) and Won

Artificial intelligence (also called machine learning) has played an essential role in PPC since I began in the field in 2010. In the last few years it has permeated everything in large, obvious ways and small and subtle ways too. One of the key roles of a digital marketing expert now is to know when and how to employ AI, what to feed it so it can learn as quickly as possible, how to monitor it when platforms severely limit visibility, and so on. AI makes a powerful servant, if you use it right and don’t bow to it as your master.

Two parallel experiences of mine in the past year, one at the day job and one with a moonlighting client, have illustrated another key role of the PPC expert: knowing when to fire the AI and go back to doing some part of the work manually — in this case, manual keyword bidding.

I’m not opposed to AI. From almost the beginning, I used one of its manifestions in Google Ads (then AdWords) with excellent effect. We worked carefully and diligently to estimate the lifetime value of a customer, then used those results (continually analyzed and often updated) to guide our PPC spend. We gave our ad campaigns an appropriate CPA (cost per acquisition) target and turned it loose. Within 18 months or so, we were profitably spending over $200k per month, orders of magnitude more than before.

Two Accounts with Similar Targets

More recently, with one of my moonlighting clients, we gave our busiest e-commerce search and smart shopping (later, Performance Max) campaigns automated targets that weren’t available back then: return on ad spend (ROAS). Depending on the campaign, among other things, we might set the target ROAS at 2.0, meaning our revenue doubled our ad spend. If we wanted to pull more traffic to a certain campaign for a while, we’d lower the ROAS target temporarily to 1.5 or even 1.2.

On the day job, we had very productive branded search campaigns (with search keywords including the brand name). We might have used ROAS targets, but ROAS already was reliably at or near our target ROAS (10.0) anyway, and then it doubled during the pandemic. So I set impression share targets instead. I wanted top-position ad impressions at least 97% of the time when someone searched on anything resembling one of our branded keywords. (I don’t set such targets at 100%, because the last two or three percent can get outrageously expensive.)

Both of these targeting strategies rely heavily on Google’s AI. And both were quite successful for months or even years. Which brings us to today’s tale.

New Trouble, Part 1

Last year at the moonlighting client, over several months, in our most important search campaign and to a lesser degree in our most important shopping ads campaign, we saw the cost per click (CPC) double, then triple, then nearly quadruple. (As a performance target CPC is far inferior to CPA or ROAS, but it can be an important indicator.) The math is unforgiving: CPA and ROAS plunged proportionally, as did clicks and impression share (IS), simply because, all else being equal, at a higher CPC the same budget gets fewer clicks and a lower impression share.

Over and over again, I analyzed our own performance and our competitors’. I looked at everything I could think of. Yes, competition had increased, but not that much. No, we weren’t leaking money on irrelevant keywords. And so on. I had developed a pretty good bag of PPC tricks, and I knew a lot of ways to study performance, but nothing helped.

I consulted with some fairly good Google Ads reps. (They exist.) They were puzzled too.

To succeed for very long in digital marketing, you have to learn continually. I’m okay with that. In recent years, with revolutionary changes in the major platforms, the learning curve has been noticeably steeper. I’m generally up for that challenge. But this problem had me beating my head against the wall, not quite literally.

I had also developed pretty good instincts, I think. Those help a lot with lead generation campaigns and with low-traffic campaigns and accounts generally, where actionable data can be sparse. In this case my instincts stubbornly insisted that none of the usual suspects, especially increased competition, were enough to account for our troubles.

Back to Basics

When I take over an account, what I do in search campaigns is a lot like starting fresh. I go back to the basics: the most obvious, workhorse keywords; careful attention to keyword bids and performance; and so on. But I’d never had to do that with my own campaigns, my own accounts.

Maybe it was inspiration or desperation or both. It dawned on me one day that the only possible solution I hadn’t tried here — in an account I had created years before and managed ever since — was going back even further, to the most basic basics, so to speak.

So I fired the AI — I stopped targeting ROAS — and resumed bidding on keywords manually. Yes, I was deluged with hints, suggestions, and outright warnings that I was losing clicks, that I was missing out on revenue, and so on. But within a couple of weeks the results were clear, and within an month they were even clearer. Our CPC went back down, close to what we had previously regarded as normal. Our clicks and impression share recovered. And our ROAS climbed above our target and stayed there.

New Trouble, Part 2

Several months later, on the day job, in those branded search campaigns which had met my impression share targets and soared above our ROAS goals, I realized I was seeing similiar decay. For a while, it could have been e-commerce tailing off after the peak of pandemic lockdowns and lavish relief checks, enhanced by competition. But eventually it was clear that there was more going on. Our impression share — our key performance indicator (KPI) in this case — was tailing off because our CPC kept climbing. We increased — then finally doubled — the spend, but we couldn’t get back to our target or even close.

I studied our own performance carefully, and I studied the competition. I had access to sales performance numbers beyond digital marketing, so I weighed those in the balance too. The longer I watched and the more I analyzed, the more it looked like what I had seen (and, perhaps foolishly, half-forgotten) several months earlier.

Why Not Twice?

I fired the AI again — removed the impression share targets — and went back to manual keyword bids. This initiatied the same ongoing sequence of automated threats, warnings, suggestions, and invitations to surrender to the AI. (I’m paraphrasing, and none were as ominous as Locutus of Borg intoning, “You will be assimilated.”) But here are some real before-and-after numbers, separated by a very few weeks:

• In the week of January 30, 2023, CPC peaked at $4.89. That’s when I fired the AI. Two weeks later, CPC was down to $1.40.
• In the same period, search impression share climbed from 73.6% to 94.3%.
• Conversion value (revenue) climbed from $6,255 to $7,420 that first week. By the end of February, it was $14,085 for the week.
• Spend fell from $1,787 to $557 in those two weeks.
• ROAS (“conv. value/cost”) climbed from 3.49 to 11.58.

If you must go back and review these numbers to digest them, do that now. They’re excellent.

Since then, the numbers have stayed in these happier ranges, which has freed up some budget to try new things.

I Can’t Explain Why

To summarize: in the space of two weeks, firing the AI by switching from an IS or ROAS target to manual keyword bidding had these results:

• CPC decreased more than 71%.
• Search impression share increased 28%.
• Revenue increased more than 18% in the first week and 125% in a month.
• Spend decreased 69% (in those two weeks).
• ROAS more than tripled.

I cannot explain why the AI worked well for months or years then gradually but dramatically decayed. I’m not suggesting Google caused this intentionally. Google mid-level analysts are often pretty good, and the first time this happened, they couldn’t explain it either. The second time, I didn’t ask them.

I can say that in two very similar cases in the span of several months, I fired the part of the Google Ads AI, and depressed PPC performance immediately soared.

If there’s a moral to this story, and I think there is, it’s something like this. You have to know when to trust the AI, and how far, and in what ways. If it stops pulling its weight, you may have to show it the door, ignore all the system’s protests, and trust yourself and your quaint, old-fashioned PPC skills instead.

It worked for me — twice — both it terms of costs vs. revenue and in terms of my professional confidence. The latter was briefly shaken, when I couldn’t find a solution. Until I did.

Your results may vary, but good luck!

Photo credit: Abby Savage on Unsplash