If you’ve ever set up a Google Performance Max (PMax) campaign, you probably know the feeling. You upload your creatives, set your budget, click “Launch,” and then just watch the numbers, trying to figure out what the system is actually doing.
But imagine that you can look under the hood of this “black box.” Imagine that you understand how PMax artificial intelligence thinks — just as we learn to formulate queries to ChatGPT correctly in order to get more accurate answers.
In this article, we will examine four key principles of AI in advertising that Google does not mention in its documentation. This is not a list of official recommendations, but rather a practical model of thinking. It will help you stop simply observing your campaign and start purposefully training your advertising algorithm.
How many calls and sales will I get by ordering contextual advertising from you?
For example, instead of wondering why CTR dropped on the third day, you can predict how the system will respond to audience changes or the addition of new images. It’s like going from being a passenger to being a pilot — technically, the autopilot is still working, but now you understand its logic.
How does AI see your ad?
To understand the logic behind artificial intelligence in PMax, it is worth forgetting about complex marketing terms. In fact, the system thinks in a very simple way — in the format of binary tags:
1 — conversion has occurred. This is “good.” Each such unit has its own value attached to it.
Key point: AI does not understand your business realities — margins, costs, seasonal demand. Its inherent goal is to collect as many “ones” with the highest aggregate value as possible. Restrictions such as Target ROAS, which you set, are external constraints, not the internal motivation of the algorithm.
Imagine a dog that has been trained to fetch balls. It will fetch all balls in a row — big, small, dirty, someone else’s. It doesn’t care whether you need that particular ball. Its task is to fetch as many as possible.
Machine learning works the same way in advertising. Without clear ROAS restrictions, a campaign can generate dozens of conversions that look great in reports but actually eat into your profits. The system simply performs a basic task — it collects “ones with the highest value,” regardless of whether it makes sense for your business.
Practical tip: always set Target ROAS or Target CPA from day one of your campaign. This isn’t an option for experienced users — it’s a mandatory leash for the algorithm, which will otherwise optimize according to its own rules, not yours.
PMax does not start from scratch, but uses previous experience
Many advertisers believe that the new Performance Max campaign is a tabula rasa, a blank slate. They say that the system is “blind” and has to learn from scratch using your data. In reality, things work differently.
Each new campaign is based on what is known as global training — Google’s enormous knowledge base, formed from billions of auctions, clicks, and user behavior patterns from around the world. It’s like hiring a new sales manager who has been working in the industry for 10 years. Yes, they don’t know the specifics of your business, but they are familiar with the general market patterns.
Counterintuitive conclusion: Your new campaign is never “dumb.” It already has a reasonable starting hypothesis about who to show ads to—even if the account has zero history.
This debunks the popular myth about the need for “warm-up.” Some people advise first launching regular ad campaigns, collecting data, and only then switching to PMax. But thanks to training, the system already knows:
Buyers of expensive electronics in Germany usually compare prices for several days before making a purchase.
Customers looking for inexpensive goods in Ukraine tend to make impulsive decisions.
Mobile device users convert differently than desktop users.
Practical example: imagine that you are launching an online sports nutrition store. Even without any previous conversions in your account, the Google Ads algorithm already “understands” that the typical protein buyer is a 20-35-year-old man who is interested in fitness, watches YouTube channels about training, and often makes purchases in the evening after work.
Performance Max is not a primitive advertising tool. Yes, it is essentially a huge calculator. But it is a calculator that performs thousands of operations for you in real time: analyzes audiences, tests creatives, distributes your budget across channels, and adjusts bids at each auction.
Tip: Don’t spend weeks “warming up” your account before launching PMax. It’s better to invest this time in high-quality creatives and proper conversion settings — this will help the algorithm adapt global knowledge to your specific business faster.
Your settings are a “prompt” for artificial intelligence
If you have worked with ChatGPT, you know that the quality of the answer directly depends on the quality of the query. A vague prompt will result in a vague answer. A clear, detailed prompt will result in an accurate answer.
With Performance Max, everything works the same way. Your settings in the ad account are not just technical parameters. They are a complete prompt for AI, a set of instructions that determine the behavior of the algorithm. And the more accurately you formulate this prompt, the better results you will get.
Here are the important elements of your “advertising promta”:
Conversion goals and their value
This is the main instruction for the system. You literally tell the algorithm: “This is what I consider success, and this is how much this success is worth to me.” If you have set up purchase tracking with the transfer of the actual value of orders, the system will understand that an order for $5,000 is more valuable than an order for $500. If all conversions have the same value, the algorithm will simply optimize by quantity, ignoring your actual revenue.
How many calls and sales will I get by ordering contextual advertising from you?
These are your business parameters, a kind of filter. You tell the system: “Find me conversions, but only those that fit these profitability indicators.” Without this filter, the algorithm will participate in any auctions where it sees a chance for conversion — even if the cost per click eats up all your profits.
Product feed
For machine learning, a Google feed is not just a list of products. It is a dictionary that the system uses to describe your offer. Compare the two approaches:
Bad feed: “Men’s black sneakers.”
Good feed: “Nike Air Max 90 men’s sneakers, black, leather, size 42, for running, item NM90-BLK-42.”
The first option is like asking ChatGPT to “write something about marketing.” The second is like giving it a detailed technical task with all the input data. What’s more, during the auction, the system compares your feed with those of your competitors. A richer description = higher quality rating = advantage in impressions.
Campaign structure and geotargeting
This is not just a convenient way to organize your account. It is a strategic tool called task decomposition in machine learning. Instead of one big task, “sell everything to everyone,” you create several smaller ones: “sell electronics in Kyiv,” “sell clothes in Lviv.” Each campaign becomes a separate learning container with its own clean data set. The result is more accurate predictions and more stable optimization.
Practical example: imagine an online store that sells both budget accessories and premium equipment in a single PMax campaign. The algorithm receives conflicting signals: sometimes the conversion is $5, and sometimes it is $1,500. Buyer behavior is radically different, but the system tries to find an “average” pattern. The result is mediocre performance for both categories. Divide them into separate campaigns, and each model will learn from relevant data.
Tip: Treat PMax settings as a brief for a contractor. The more detailed you are in describing what you want, who you are selling to, and what exactly you are offering, the less likely the algorithm is to “think up” something on its own.
Performance Max creates a personalized AI model specifically for your store.
Google’s global knowledge base is just the starting point. The real magic happens during local fine-tuning. This is the process where the general model adapts to the unique data of your specific advertising account.
Imagine that you have just hired an experienced salesperson from another company. They know the market in general, understand buyer psychology, and are skilled in sales techniques. But they need time to learn about your specific product range, your pricing policy, and your typical audience. In a month or two, he will already know what product to offer to a specific customer even before they voice their needs.
Fine-tuning in PMax works on the same principle. The system analyzes thousands of your unique “ones” and “zeros” to understand patterns:
Which products are most often purchased together?
At what time is your audience most active?
Which creatives inspire the most trust in your customers?
Which devices generate the most profitable orders?
If you communicate with your own profile in ChatGPT or Claude for a long time, you will notice that the system begins to adapt to your style, remembers the context, and offers more relevant answers. The same thing happens with personalized advertising on Google — only instead of text, the algorithm studies the behavior of your customers.
Why is this important in practice? Here’s the answer to a common question: “Why does my competitor’s campaign with the same settings work better?” The fact is that each advertising account develops its own unique AI model over time. Two stores with identical products, budgets, and Target ROAS will have different results because their models were trained on different data.
Practical example: two online stores sell the same Nike sneakers. The first one has been operating for three years, has thousands of conversions, and a detailed history of buyer behavior. The second has just launched. Even with identical settings, the first store will get better results — its machine learning model already “knows” that sneaker buyers in this region usually compare three or four options, buy mainly in the evening, and often return for socks a week later.
That is why Google recommends collecting 30–50 conversions before evaluating the effectiveness of a campaign. This is the minimum sample size required for the algorithm to start building a personalized model for your business. Until then, the system relies primarily on global data rather than your specific characteristics.
Tip: Don’t make any radical changes to your campaign until you’ve gotten at least 50 conversions. Every significant adjustment — changing your bidding strategy, adding new ad groups, drastically changing your budget — restarts the learning process. You’re essentially forcing the algorithm to start fine-tuning all over again, losing all the experience it has accumulated.
Conclusion
Performance Max is not a magical “black box” whose results are impossible to predict. It is a complex AI system for advertising that can and should be consciously trained. Once you understand how it works, you cease to be a passive observer and become an active trainer of the algorithm.
Instead of asking, “Why did PMax do that?” you start asking another question: “What do I need to change so that he acts differently?”
Your tools for this training:
Conversion settings — so that the system understands what you consider to be a success.
The quality of the product feed — so that the algorithm knows exactly what you are selling.
Campaign structure — so that each model learns from clean, relevant data.
Clear business goals (Target ROAS/CPA) — so that the system searches for profitable conversions rather than just any conversions.
Each change in settings is a new prompt that adjusts the model’s behavior. You no longer struggle with the algorithm, but guide it in the right direction.
Practical challenge: now that you understand how AI thinks, ask yourself this question: which one prompt in your advertising account will you change today? Perhaps it will be feed detailing, separating campaigns by category, or adjusting your target ROAS. Start small and see how the system responds to your new instructions.
