The Ups & Downs of AI
Artificial Intelligence and the bidding strategies that come from it are wonderful tools for growth. Every large-scale account should experiment with automated bidding and strive to incorporate it into search marketing. However, no technology is 100% perfect. AI bidding presents its own set of problems, which we’ll explore in this guide. We’re hoping this will help you avoid common mistakes and take better precautions when managing campaigns that use automated bidding.
The Faults of AI Bidding
There are a few things that AI bidding doesn’t do very well. Some of them are completely intentional, believe it or not. For example, AI bidding is prone to overbidding and overpaying in certain scenarios. We can easily understand why if we think of it from Google’s perspective. Their interest is in advertisers paying more on a per click basis. That’s how Google generates more revenue.
In this guide, we will cover another issue that Google and various other tech companies are continuously trying to improve.
We’re talking about the ability of AI to adapt to drastic change.
The Ability of AI to Adapt to Change
Artificial Intelligence is a rapidly evolving technology. We went from things like image recognition, which was considered “science fiction” 10 years ago, to AI experts claiming that AI will surpass human intelligence and be fully independent within our lifetimes.
As we approach the year 2020, AI has almost reached the point where it can optimize PPC campaigns completely on its own! All it needs is a CPA or ROAS goal and it’s good to go, right? Well, not exactly. It is highly recommended to train your campaigns under manual bidding and then transfer to AI after you’ve done some of the heavy lifting yourself.
Additional concerns arise when you want to make a BIG change, whether that be a drastic decrease in budget or even worse, a drastic increase. Changing your efficiency target too drastically can have negative outcomes as well. Website changes can also play a major role in your campaign’s performance.
We’ll cover these cases here and focus on what happens when you make drastic website changes. We have a good case study to show you a recent issue we experienced that had some unexpectedly negative results. Luckily, reading this guide can help you avoid similar outcomes.
Let’s suppose you’re running some campaigns under an automated bidding strategy and over the holidays you want to run a special promotion and waiver your application fee for the next 3 days. You want to start spending double your daily budget, but just for those 3 days. Well, the artificial intelligence won’t be too happy about it. As it learns and grows it looks at roughly 14 to 30-day time frames and optimizes based on the data it finds in these periods. So, if you put AI up to the task of doubling spend for only 3 days, it won’t be able to do much.
You have 1 primary option, and that is to increase the budget option from X to 2X. The AI will simply not know what to do. It will try to increase spend as much as possible, but 2 things will happen simultaneously: Due to the change being so drastic, the AI will not be able to adapt fast enough to utilize the 2X budget. Instead, it will underspend. The second thing is that the AI will drastically increase your CPL, as it is trying to spend more money. The AI will raise bids inefficiently and your CPCs will go up. If you had given the AI enough time and tasked it with an incremental increase of, let’s say, 10% per every few days, over the course of, let’s say, 2 weeks, perhaps it would have done a good job. Going from X to 2X, however, will be a disaster. Then it gets worse. The 3-day promo is now over, your CPL is up and it is time to reduce the budgets back to their original values. The AI was just starting to try to achieve balance, and now you want it do something completely different.
Will the AI just revert back to how it was doing before? Nope. It does not remember that anymore. It now remembers those 3 days, where you wanted to spend 2X and will therefore have to start optimizing back down. We’re sure Google is working on solving this problem but as of now, this is what we’ve experienced over and over again.
Efficiency Goal Changes
Much like drastic budget changes, we highly recommend that you do not attempt to make drastic efficiency goal changes to your AI campaigns. If you have a CPL goal of $30 and the AI is comfortably achieving that, you can’t just randomly decide that you want to get a $20 CPL. To be fair, you’d have a similar outcome if a human tried to do the same thing. Trying to get cheaper leads won’t work, which is common knowledge in PPC. But if you try and change your AI goal, there is no going back. Once the AI goes off balance, you will have to retrain it manually and set it back to your $30 goal. We just have to be aware of this and only make small incremental changes to our efficiency targets.
Same thing goes for making a drastic change the other way around. Even if you decide to drastically increase your goals (a rare case), if the change is too sudden, the AI won’t adapt well. In this case, a human can do much better if, say, you’re looking to increase your CPL while spending significantly more money.
AI & Website Changes
Perhaps the least known AI-related issue is that running AI campaigns while making changes to your website can drastically impact your performance. You see, AI’s reactions are based on many, many variables. Google put this number at over 80 million! One of these variables is, of course, CVR-based performance. If your website change leads to a change in CVR, the way your automated bidding strategy works will be impacted as well. If that change is too drastic, it can lead to disastrous results.
We recently found this out the hard way. The client changed their lead form by adding 4 extra fields. Usually not a big deal, but this lowered CVR just enough that the AI went completely crazy to the point of dysfunction. CPL ballooned by 3-4 times and was performing way outside of the set goal, despite previously being very efficient. The AI could not adapt fast enough to the change in performance and it started making very inefficient optimizations. At first, we waited a few days in hope of recovery, but we kept burning through cash. The only choice we had left was to pull the plug on the robot and take back manual control.
The Results – Conv. Decrease
The Results – CPL Increase
The Results Explained & Damage Control
The numbers speak for themselves. After the website change was executed on 10.24.19, we can see from the screenshots that the conversions of these campaigns decreased by 76% and CPL increased 210% on a WoW basis. The AI could not adapt to the change in CVR that was caused by the website change. Its optimizations kept on getting worse and worse, which led to a good chunk of wasted spend.
We reverted to manual CPC as the AI was beyond recovery. If you experience a similar situation, you should quickly do the same. However, there are things you can do to prevent this from happening in the first place if you’re running AI and want to make changes to your website:
- Slowly implement website changes. Don’t launch everything at once but just a few sections at a time, if that is an option.
- A/B test either through “experiments” in AdWords or split traffic at the URL level.
- Keep the old version of your website live and lead only a few campaigns to the new version to start. See how this will impact performance and decide on what to do next.
- Finally, you can decide to pre-emptively switch to manual CPC prior to drastic changes in order to avoid wasting budget.
In conclusion, we can say that AI has become an important part of our lives and is essential for any large scale PPC account. We’re not longer competitive unless we utilize it in one form or another. The time of man vs machine is upon us, and as humans, we must adapt.
However, giving AI all the control and being fully reliant on it is no good either. We must be aware of its weaknesses and supplement it with our own knowledge and expertise when necessary. We’re still the guides for our machine friends because at this point, AI still can’t do it all.