The most useful number in AI performance marketing right now comes from an analysis of roughly 55,000 Meta campaigns: between May 2024 and May 2025, the cost to acquire a new customer through Advantage+ rose from about $257 to about $528, while reported return on ad spend stayed close to 4.52x. ROAS looked stable. Acquisition economics did not. If your dashboard says the automation is working and your finance team says growth is getting expensive, both are reading the same campaigns correctly.
Understanding why those two numbers moved in opposite directions is most of what you need to manage these systems well.
Why ROAS holds while CAC doubles
Automated bidding optimises toward the conversion event you give it, and it does so by finding the cheapest available conversions. Retargeting a warm audience is cheap. Acquiring someone who has never heard of you is expensive. Left alone, the system will lean toward the cheap conversions because that is what the objective rewards.
The result is a blended ROAS figure that stays healthy because it is increasingly made up of purchases that would have happened anyway, while the genuinely new customers inside that same account get more expensive to reach. The reported metric is not lying. It is answering a different question from the one you care about.
This is why new-customer acquisition reporting, where available, matters more than blended ROAS for any account trying to grow rather than harvest.
How much of your spend this already controls
The automated share of major platforms is now high enough that opting out is not really a strategy:
- Performance Max drives roughly 45% of conversions on Google Ads
- Around 78% of Google Ads spend now runs through Smart Bidding or Performance Max
- Advantage+ accounts for about 62% of Meta ecommerce spend
- Meta’s automated ad products are running at roughly a $60 billion annual rate
- Google reported $82.3 billion in advertising revenue in Q4 2025
Meta is also deprecating the older creation paths for Advantage+ Shopping and App campaigns around 19 May 2026, folding them into a single automated setup. If your team still builds campaigns the manual way, that route is closing rather than becoming a supported alternative.
Where automation genuinely outperforms, and where it does not
The data on this is less uniform than either the platforms or their critics suggest.
On the pro side, Optmyzr found that mature Performance Max campaigns returned about 616% compared with roughly 125% for new ones. That gap is the learning curve, and it is steep enough that judging PMax in its first month tells you almost nothing.
On the other side, in B2B accounts, Search campaigns returned around 553% against PMax at 436%. For considered purchases with long sales cycles and a small qualified audience, keyword intent still beats audience modelling. That is the clearest segmentation in the whole dataset: PMax rewards volume and broad addressable markets, Search rewards specificity.
Advantage+ has a structural requirement worth knowing before you commit budget. It needs roughly 50 conversions per week to exit its learning phase and perform. Below that threshold you are paying for a system that has not got enough signal to work, which is the most common reason small accounts report that Advantage+ underperformed for them. They were not wrong. They were under the floor.
Creative volume is now the main lever you control
Once bidding, placement, and audience selection are handled by the platform, what is left on your side is creative and measurement.
Accounts producing 20 or more ads per month report around 65% higher ROAS than accounts producing fewer. That is not a claim that more ads are inherently better. It is that automated systems need variation to test against, and starving them of creative options means they optimise within a tiny space.
This is the practical shift in the job. The skill that used to matter was bid and audience management. The skill that matters now is producing a high volume of distinct creative concepts fast enough to keep the system supplied. Our roundup of AI content marketing tools covers the production side of that, which is now the bottleneck in most paid accounts.
Managing target ROAS without breaking the campaign
Target ROAS is the main steering wheel you still have, and it is easy to use badly. A few mechanics that hold up in practice:
- Set your initial target within about 20% of your actual current ROAS. Setting an ambitious target on day one throttles delivery and the campaign never gathers enough data to improve.
- Raise it 10% to 15% at a time, no more often than every four weeks. Faster than that and you are re-triggering the learning phase repeatedly.
- Change one variable per cycle. Target, budget, and creative set all reset learning. Changing all three simultaneously means you learn nothing about which one mattered.
- Do not judge a change inside 14 days. Most of what looks like a result in week one is delivery volatility.
Above roughly $10,000 a month in spend, add holdout testing. Withhold a randomised share of your audience from advertising entirely and compare conversion rates. It costs you some reach and it is the only method that tells you whether the platform’s reported conversions are incremental or whether you are paying to reach people who were going to buy regardless. Every sophisticated advertiser above that spend level now runs some form of incrementality test, and the results are frequently uncomfortable.
The measurement problem coming from search
One more shift worth planning around. Google’s AI Mode has reached roughly 75 million daily active users, and approximately 93% of those sessions end without a click to any website.
That has two consequences for paid acquisition. First, the organic top-of-funnel that used to feed your retargeting pools is shrinking, which pushes more of the acquisition burden onto paid and contributes to exactly the CAC increase in the opening statistic. Second, attribution gets harder, because a user who researched you inside an AI answer and then searched your brand name directly looks like a free brand conversion when it was actually earned by content you can no longer measure. The dynamics behind that are covered in more depth in our comparison of GEO vs SEO.
A workable operating model
Putting it together, the structure that seems to work for accounts running automated paid acquisition in 2026:
- Run automated campaign types as the default, because the manual alternatives are being removed
- Report new-customer CAC separately from blended ROAS, always
- Keep Search campaigns alive for high-intent and B2B, where they still win
- Budget for creative volume rather than for campaign management hours
- Run a holdout test quarterly above $10k monthly spend
- Give any new automated campaign eight weeks before judging it, and make sure it clears the conversion volume floor first
For the tooling layer around this, our overview of AI marketing tools covers the software side, and AI marketing strategy takes the wider view across channels rather than just paid acquisition.
The thing to take away
AI performance marketing has not failed. It has succeeded at the objective it was given, which was efficient conversion, and efficient conversion turns out not to be the same thing as growth. The accounts doing well with it are the ones that changed what they measure, not the ones that fought the automation.
Check your new-customer acquisition cost against the same period last year. That one number will tell you more about your account than your ROAS column has for a while.
