My name is Vyacheslav Ignatiev. I am a PPC specialist at 1jamagency.com. We are engaged in managing paid traffic in eCommerce. We have a lot of expertise in various niches. I will try to clarify the question of opting in or out of automated bidding strategies for your business.
If you have experience using automated bidding strategies, feel free to share it in the comments to make the picture more complete.
Automated bidding strategy is artificial intelligence. However, the word intelligence can be misleading. AI cannot think and analyze like a human, its learning process is more akin to dog training. A dog hears "Sit down!", it sits down and gets a treat. Repeat it 50 times, and a reflex is formed: upon hearing "Sit down!" the dog feels the need to sit down, and it knows it will feel good.
Automated bidding strategies in Google Ads work in much the same way, only the training signals are different. A signal could be any factor that can help achieve a goal: a device, a location, a search query, seasonality.
If George from Washington purchased something from us, then the algorithm makes a neural connection between George, Washington and the purchase. Then it will try to bring Washington residents with the same interests as George to the site, and the more often this hypothesis works, the stronger the neural connection becomes.
If the algorithm receives insufficient signals to learn from, it will not be able to work effectively. For an online books and gifts store, we tested automated bidding strategies on one of the advertising campaigns with a monthly budget of about $1000, and here's what came out of it.
Stage 1: optimization for microconversions. First, we set up an automated bidding strategy with optimization for microconversion action "Viewing 5 pages", which consistently returned 30+ conversions per week to the algorithm. In 5 weeks, we managed to reach 370% ROAS with a target of 833%.
Stage 2: optimization for orders. Next, we tasked the algorithm with bringing orders, which unlike microconversions added up to just over 10 per week. This was not enough for training and reaching stable indicators, therefore ROAS decreased by 152%.
Lesson Learned: Automated bidding strategies are not worth launching unless we expect 50+ conversions per month. Google's recommendations say that 30+ conversions per month should be enough, but in our experience this is not the case. Running automated bidding strategies on a small number of conversions is likely to just drain the budget.
In this project we applied an automated bidding strategy with "warm-up", which helped to gradually reduce CPL. We were not sure if we would get enough conversions to optimize for conversions, so we started with optimizing for clicks.
December-May: optimize for clicks. We started with an automated bidding strategy with optimization for Maximum Clicks. This helped us to verify on large traffic that the offering is in demand, and it brings sales. As a result, we got 67 conversions for $46.
May-September: we apply "Maximum Conversions" optimization strategy. Next, we tasked the algorithm with bringing as many conversions as possible. We deliberately abstained from limiting the cost of conversion into leads (CPL) so as to see the maximum possible number of leads that can be reached with this automated bidding strategy within the allocated budget. As a result, we got more conversions, and the cost per lead decreased: 101 orders for $27.
September-February: Optimization by Target Cost per Lead. Now we told the algorithm, "Do the same, but now CPL should not be higher than $25." In such a situation, the algorithm begins to act more aggressively and raises the bids within the target cost per lead (CPL). The result was 191 conversions for $20.
Lesson Learned: If we gradually move optimization lower down the sales funnel (from clicks to sales), we can reduce risks, gain more control over the situation and give the algorithm enough data to train on.
Working with automated bidding strategies has many advantages: they take over the routine, analyze data inaccessible by a specialist and run through huge data streams in literally seconds.
But there are also disadvantages, one of which is unpredictability. We came across this with an online store for school supplies when we switched from manual settings to automated bidding strategies:
Lesson Learned: There are no magic pills :) You should always be ready for a scenario where everything breaks down and you need to find other ways to achieve the goals. Moreover, the attempts to find the cause and remedy the situation may lead to nothing.
Another way to feed more conversions to an automated bidding strategy and thus speed up training is to combine several campaigns into one portfolio. With excessive fragmentation of campaigns, each automated bidding strategy will live in its own small world and do the same thing, but with fewer resources. So, the number of conversions may not be enough for optimization. If the campaigns are bundled into a portfolio, the algorithm will get a bigger budget to operate with, it will be able to analyze more positive and negative scenarios, draw more conclusions and apply them to all the goals of the ad campaign at once.
Understanding all this, we developed a strategy for an online store for vanity phone numbers. We started with a lot of advertising campaigns configured manually: 1 campaign = 1 type of user queries. Then, to scale the project, we used a portfolio bidding strategy with target cost-per-action (tCPA) optimization. From the very first month, we managed to improve on the past results: the number of orders grew, ROAS grew.
Lesson Learned: If the products have similar advertising parameters (audience, target cost per order, profit margin, price), use portoflio bidding strategies. With small volumes, this will help to surpass the threshold of 50 conversions per month for training, but with large ones, it will allow the algorithm to wallow in feedback signals and learn quickly.
In this case, we applied the tactic of gradual transition between optimization goals. At first, an automated bidding strategy with optimization for microconversions worked, which gave 15+ conversions per week. For the tenth week of training, we switched to optimization for placing orders on the website.
Everything seemed to be sensible, and the results were expected to be better. But there is a nuance.
This happened in February-April 2022, when electronics sellers faced a rush demand: prices spiked amid complete uncertainty about the exchange rate and future imports — in short, it was better to buy now than to be sorry later.
And these were the conditions in which our algorithm was learning. Automated bidding strategies are able to adjust for demand seasonality, but they certainly were not programmed to adjust for such a unique crisis.
Therefore, when demand fell, the algorithm started to learn the behavior of the audience anew. We made the task more complicated by switching to optimization for leads, which meant that the algorithm now had fewer conversions to learn from. As a result, orders dropped from 15-25 to 4-12 per week.
Lesson Learned: Automated bidding strategy learns best in a stable environment. It is able to use familiar tools, carefully test hypotheses and improve results. If something changes drastically, do not expect a quick reaction: the algorithm will use the past experience as long as it can, and only then will it slowly begin to learn how to work in the new conditions.
If you operate in an unstable environment, it is better to stay on manual settings or "help" the automated bidding strategy by adjusting the budgets and the target cost per lead. Advertising systems have algorithms that take into account seasonality — but we are talking only about "traditional seasonality". They can predict growth of demand for equipment before the New Year, but they do not understand anything about sudden changes in the macroeconomic situation.
When we started working with this sports goods store, most of the costs were sustained for a campaign which included such search phrases as "category + brand", "brand + model" generated by the template.
The campaign was managed manually and consisted of 89 ad groups. Only in 18 of them the costs were more than $125. The results were so-so:
One option was to simply disable this campaign, but we decided to first test it with an automated bidding strategy with optimization for adding goods to the cart. In just a week, we noticed changes: inefficient search phrases began to disappear from the top of costs, the total cost of the advertising campaign decreased, revenue increased by 4.5 times, ROAS increased 21 times — up to 1111%.
Lesson Learned: Automated bidding strategy algorithms are great at analyzing large data sets. It is difficult for a specialist running manual settings to manage 89 ad groups in real time: they are likely to launch them and then update them only every few days. By contrast, an automated bidding strategy continuously redistributes the budget between groups, and already in the second month, we were left with only efficient search phrases.
This case is not from eCommerce, but from services. However, it is too revealing not to share.
At the start, we usually set up automated bidding strategies with optimization for microconversions so that the algorithm receives enough data for training. This was also the case with this project: "Target CPL" strategy with optimization for the "Viewing 2 pages" target gave strong results. The growth compared to the manual settings was visible immediately: there were 2 times more leads, and their cost was 3 times lower.
Lesson Learned: Optimization for microconversions can improve orders and costs. To make this work, we need to analyze and choose the correct microconversion which most correlates with the target action. For example, it may turn out that the orders are placed by:
Then it makes sense to set up the automated bidding strategy to attract people who will spend 2+ minutes on the site.
Here is our algorithm for finding the microconversion that correlates with the target action the most ↓
The microconversion to be applied at the launch of the automated bidding strategy should be selected carefully. This way you will not just train the algorithm at your cost, you will get it trained and get acceptable results as well.
The main KPI for the project is the target return on advertising spend (ROAS). With ROAS of 500%, advertising investment breaks even, while with a lower value it becomes loss-making. The optimal balance of sales and costs is achieved with a target ROAS of 677%.
The history of the project saw several attempts to launch smart banners. Each time they attracted a lot of traffic and raised conversions, but the ROAS was below 500%, which is why they had to be turned off.
Stage 1. In July 2021, we optimized the feed keeping only the top product categories and launched an automated bidding strategy with optimization for clicks. We received only 1 sale and 40% ROAS.
Stage 2. In August, we set up optimization for target CPA with "Transaction" as the target action. In September, we changed the action to "Add to basket". The number of orders increased slightly, but the target ROAS was still low.
Stage 3. We set up optimization for ROAS — and the campaign took off! Since then, we have stable revenue for this campaign with ROAS above 1000%.
Lesson Learned: If a campaign does not work with any strategy, optimization for ROAS often helps out. However, the algorithm needs to be properly warmed up: start with optimization for clicks, then switch to optimization for conversions, and only after it is trained, set up optimization for ROAS.
From the cases above, it may seem that the simpler the optimization, the better. The algorithm has a lot of data, it brings the audience to the site and visitors somehow end up buying something on their own. But no, it's not that simple.
For the online car tuning parts store, we started with the automated bidding strategy "Maximize clicks" — the final ROAS was 500%. It should be understood that with such a goal, the algorithm brings in the most "clicking" customers — and those are not always the people most interested in buying and paying.
Therefore, after several months of training, when we reached a stable number and cost of orders, we switched optimization from maximum clicks to placing orders on the site. As a result, the two factors converged: the algorithm was already trained enough on clicks and it was refocused to bring sales. As a result, ROAS rose to 714%.
Lesson Learned: Traffic-based strategies prioritize just that — attracting traffic. They can solve performance tasks by inertia, but they can rarely do it effectively. Therefore, they are suitable only for the first stages of training the algorithm. After that, the optimization should be switched to the target action.
We launched automated bidding strategies only when we really expected them to be effective, because we already had a rough understanding of the situations where they should work. In most cases, our expectations came true. That said, there would have been much more failed cases if we did not follow the basic contraindications to launching automated bidding strategies. These are:
The data volume is the biggest factor in effectiveness of automated bidding strategies. If it is small, there is no point in even enabling them. Google recommends enabling conversion-based automated bidding strategies if there are more than 30 conversions per month. However, in our experience, it needs double that.
The results are needed immediately. Automated bidding strategy will achieve optimal results only in a few weeks, sometimes months. Before that, it needs to learn. If you need cheap leads quickly, automated bidding strategies are not the best option.
You have a low margin Even after training, the algorithm may "act funny": if increasing the cost per lead by 10-20% for you means operating at break-even or at a loss, do not risk it.
A complex, unique, expensive product. Automated bidding strategies work well for advertising something typical and in high demand: clothing, electronics, massive online courses, books, etc. If you have a product with long-cycle sales or low demand, the algorithm simply doesn't have enough data to learn on.