According to PPC Land’s report, Advantage+ creative features can generate up to 150 variations from the advertiser’s assets, and Meta claims a 22% ROAS increase, a claim practitioners dispute. With the new terms of 30 October, the brand answers for whatever AI produces. The practical answer is a brand guide for AI and an approval process.
What 150 variations means in practice
The report notes that Advantage+ creative features can generate up to 150 variations from the assets the advertiser supplies: combinations of images, texts, formats and adjustments. Not all will be shown, and the platform chooses what to distribute based on performance. The advertiser rarely sees them all.
The critical point is this: the greater the number of variations, the lower the chance anyone has seen them one by one. If the rule is “approve everything that goes out”, the rule becomes impractical. That’s why you need to decide where AI may act and where it may not.
A brand guide for AI, on one page
Write what AI can and can’t do with your brand: tone of voice (formal or close), forbidden words, promises never made, accepted price formats, image rules (no overlaid text, no generated people, no changes to the product) and permitted languages. For Portugal, define whether the Portuguese must be European and whether automatic translations are accepted.
This guide serves the AI, the manager, the agency and whoever approves. The more concrete, the easier it is to check whether a variation complies. A vague guide (“keep the brand image”) protects nobody.
Where AI may act
Low to medium risk
- Image framing and format
- Variations in element order
- Headlines with version approval
High risk: turn off
- Automatic translations
- Generated product images
- New promises, prices and deadlines
Where AI can act with less risk
There are low-risk zones: image framing variations, format adjustments (square, vertical), small changes to element order. There are medium-risk zones: rewriting headlines and descriptions. And there are high-risk zones: translating, generating product images, adding promises, prices or deadlines, and changing the click destination.
A simple policy: low risk, turn on with sampling; medium risk, turn on with approval of versions; high risk, turn off. In regulated sectors, move everything up one level of caution. What matters is that the classification exists and is written down.
How to sample and approve without seeing all 150
Instead of viewing all, define a sample: for example, view ten published ads weekly, chosen at random, in the main formats (feed, stories, reels). Compare them with the brand guide. If you find two or more infractions in a sample, turn the feature off until you understand the cause.
Keep screenshots and record the date. Sampling doesn’t eliminate risk, but it gives you a defensible method: you can say you check regularly, with written criteria, and that you react when you find problems. That counts, both in client relationships and in any legal conversation.
Nobody approves a hundred and fifty variations one by one. So you set rules.
Testing the +22% ROAS promise
Meta claims a 22% ROAS increase from these features, but the report says practitioners dispute it. The right attitude is to test: one ad with AI creative and one without, with the same budget, audience and period, assessing cost per purchase, revenue and margin, not just clicks.
If the AI version wins consistently and complies with the brand guide, keep it with sampling. If not, the decision is easy. In any case, the advantage should not come at the cost of messages the company can’t fulfil.
The role of human creative
AI combines and varies; human creative decides what the brand wants to say. Invest in the base assets: real photographs, short videos showing the product in use, clear and specific texts. Input quality determines the quality of the variations.
A good practice is to create three to five distinct “concepts”, each with its own idea, and let AI vary within each concept. That way diversity comes from strategy and not chance, and it’s easier to see what really works and why.
Questions for your ad manager
- Which AI creative features are active in my campaigns? 2. Is there a brand guide for AI, in writing? 3. How do you sample published ads? 4. Have you tested with and without AI in my account? 5. What happens when a variation breaks the guide?
A supplier with these answers ready has a process. One who answers “Meta optimises by itself” is delegating to a machine a decision that is yours. In Portugal, where the ad speaks European Portuguese and knows local context, that care weighs even more.
Summary: control isn't distrust
Setting rules for AI isn’t being against technology. It’s the same logic as any approval process: whoever publishes in the brand’s name should know what they publish. With a brand guide, risk classification, sampling and tests, AI helps without taking the decision away from whoever should make it. And, from 30 October, responsibility is indeed yours.
Frequently asked questions
Are all 150 variations shown?
The report says up to 150 can be generated; the platform chooses what to distribute. The advertiser rarely sees them all, which is why sampling matters.
Is the 22% ROAS increase true?
It’s a Meta claim, disputed by practitioners according to PPC Land. Confirm in your account with a test.
Should I turn everything off?
Not necessarily. Classify by risk and decide per feature. In sensitive sectors, lean towards turning more off.