Before the automation of Google Ads, there was the manual way. In a sense, the latter is still available, but making such adjustments might be ill-advised in well-developed accounts with good past performance. The reason for this is that it might reduce the efficiency of smart bidding algorithms. By the way, you should know that most common bid adjustments are outright ignored when using almost all of the smart bidding strategies, such as Target CPA or Maximize Conversions.
In this text, we will take a look at automation and the types of algorithms. To be honest, you probably won’t find many actionable insights to boost your campaigns starting tomorrow. But, if you like PPC and if you find the topic of PPC automation intriguing, you might enjoy the next 5 to 10 minutes. 🙂
Google Ads (PPC) — Your Good Assistant
PPC algorithms all are about using the input to estimate output. Some experts say that the goal of algo is to get more conversions, which is true, but it’s a limited view in my opinion. The goal of the algorithm is to leverage machine learning to improve itself in real time, based on real-world data, shortening the gap between the predicted outcomes and the actual outcome. This is why I don’t use manual CPC anymore at all. It’s focused on the keyword (exact match, anybody?) and modern smart bidding strategies are focused on the people. This is why I also stopped caring if my audience of 24-29-year-old males liked the blue or the red CTA button on the landing page.
Google knows where people are going. Google knows where the user is in the funnel because it knows that the user visited 3 competing websites, how many pages they visited, and whether they watched a YouTube video about the product they are considering buying. In a sense, Google Ads algorithms are prediction models based on user behavior data. This is why I don’t think CPC is the key metric in performance-based prospecting campaigns (in many cases, an increase in CPC is a good sign), and that is why the quality score shown to us in the platform interface is not a primary metric and it isn’t used in auctions (if you don’t believe me, believe Google’s official help doc available HERE). To be fair, some form of Quality Score is used in the auction, but that is not public domain.
I’ve been in the PPC for only 2 years, so I haven’t really worked in the AdWords era, but from what I do know, many decisions and actions had a bit of a time lag. To be fair, the internet from 2016, consequentially advertising, e-commerce, and lead generation, was vastly different from what we consider standard today. Even if we used scripts, they would adjust bids once a day, maybe even every hour. But that was all far from real-time.
Google today uses thousands of individual signals, ranging from gender, location, online behavior habits, seasonality, and the influence of real-world patterns such as weather, elections, etc. to predict the best possible bid, and consequently the outcome of the auction. All this data is collected, processed, and applied for each search on Google’s search engine page, in real time. We mustn’t forget that Google is all about relevancy. Money – yes, ads – yes, but if the users are not happy, they’ll go elsewhere, and Google will end up like Yahoo and Altavista (You can read more on this topic on medium.com, University of Maryland, and Feedough). No human can achieve this task successfully. No human can process that much data in milliseconds. Only the machine can do it. And how does it manage all this?
Toddler & AI
Let’s start from the beginning. In the early days, people were programming computers to perform specific tasks. This is similar to how we teach toddlers some basic life skills. When I was potty-training my daughter, the teaching process went like this: when you think you need to pee, tell daddy. Then we walk to the bathroom, then we take the potty, then we drop the pants, then we sit on the potty, then we do the business. I had to teach her each step in a particular order.
It was a pleasure watching her learning process about the garbage can. At first, I instructed her to throw the garbage in the garbage can. The next stage was when she started asking me about many things, can she dispose of them in the garbage can. It included everything from actual trash, pencil-sharpening waste, eaten apple, milk cartons, egg shells, dirty plates, ill-behaving toys, etc. Naturally, more than once I had to take wrong “inputs” from the garbage (data exclusion). It was my job to instruct her on the principles and core concepts of what we throw away and how.
Over time, she got the gist of it and was mostly able to handle garbage on her own, occasionally asking me for some out-of-the-ordinary things or seeking confirmation. Now, she is almost five, and she can make real-time decisions based on the changing variables, such as throwing kitchen towels in the garbage but not the toilet and also taking into account the frequency of said action and self-moderation if we’re using the last one, and daddy forgot to buy a new roll. 🙂
And just to confirm my analogy that toddlers and young kids are like algorithms – when they finally learn something, you think to yourself: “OK, that’s done”, they stop doing it for whatever reason, and you need to figure out why.
Back to the topic, a key aspect came with machine learning – when the machine was able to teach itself how to perform a task in a given situation. It’s no coincidence that chess was used for this. A famous example of machine learning in this scenario is Google’s AlphaZero (DeepMind team) winning over Stockfish 8 (find more interesting facts on PNAS website). AlphaZero was self-taught and learned only by playing against itself. It took 4 hours to learn the game.
Patrick Gilbert in his book “Join or Die: Digital Advertising in the age of Automation” draws an interesting parallel with Harry Potter and Sorcerer’s Stone, where the three protagonists each solve the tasks in the three rooms. However, none of them would be able to solve all three, or any task outside of their “specialty” for that matter. However, machine learning algorithms would.
In the next chapter, I will take a look at several types of algorithms and machine learning principles.
Want to learn more? Check out our other blog posts:
- How ChatGPT Will Impact Google Search
- Keeping Your PPC Accounts Secure: Tips to Protect Your Business
- Why You Should Have Access to the Facebook Audience Interests (Full List)
We created Royku to train marketers in data-driven marketing.