Social Ads After ATT: Creative Volume, Honest Measurement and Where AI Fits

The short version. On paid social, the creative now does much of the targeting and experiments now do the measuring. Generative AI makes ads cheap to produce. It does not make them good, and it does not tell you which ones caused a sale.

What changed, and what it cost

In April 2021, Apple released iOS 14.5 with App Tracking Transparency (ATT). Apps must ask for the user’s permission before tracking their data across other companies’ apps or websites [1]. Apple defines tracking as linking data collected in your app with data from other companies’ apps, websites or offline properties for targeted advertising or advertising measurement [2].

The effect on advertisers was large. In February 2022, Meta told investors it expected the iOS changes to be a headwind “on the order of $10 billion” to its 2022 revenue. It named e-commerce as a vertical with a meaningful slowdown in growth. It described two problems: targeting became less accurate, which raised the cost of driving outcomes, and measuring those outcomes became harder, with less granular conversion data arriving on a delayed basis [3].

For a store owner, that means two things at once. Customer acquisition cost (CAC) went up. The dashboard became less reliable. You have to manage both.

I have grown sales through paid digital marketing while keeping customer acquisition cost (CAC) tightly controlled. What follows is how I think about creative, measurement and AI in that kind of work.

The creative is the targeting now

With less user-level signal, ad platforms lean on broad audiences and their own machine learning to find buyers, with less reliance on targeting [3]. The ad itself does more of the selecting. A hook written for new parents finds new parents. That makes creative volume and creative variety the main levers you still control.

Separate two kinds of test:

• Concept tests: a different angle, problem, offer, format or creator. These find new pockets of demand.
• Variation tests: a new hook, caption, thumbnail or length on a proven concept. These extend its life.

A common mistake is running dozens of variations of one concept and calling it creative testing. That tunes one idea. It never finds the second one.

The budget arithmetic of creative testing

Judging a concept on purchases takes money. Illustrative numbers: you want at least 20 purchases before calling a concept, and your CAC is about $25. That is $500 per concept. Ten new concepts a week is $5,000 a week of test budget before you scale anything.

Twenty purchases is a floor, not proof. On a count of 20, ordinary random variation spans roughly 12 to 31 purchases at 95% confidence. So a measured $25 CAC could plausibly sit anywhere from about $16 to $41.

Use cheap early signals to kill obvious losers: hook rate (the share of viewers who keep watching past the first few seconds), click-through rate and cost per add-to-cart. Use purchases and contribution margin to pick winners.

This matters for AI too. Creative volume is capped by test budget, not by how fast you can make ads.

Measurement after ATT: three layers

No single number tells you what is working. I use three layers, each answering a different question.

Platform attribution is fast and directional. After ATT it is partly modelled. Meta itself described work on closing an “underreporting gap” for iOS web conversions [3]. Treat it as a steering signal, not a profit statement.

Lift tests hold out a random part of the audience, or a set of regions, from ads and compare sales. This is the closest thing to proof, and it matters because observational shortcuts are unreliable. In 15 large US advertising experiments at Facebook, common observational methods often failed to recover the effect the experiments measured [4]. A follow-up study across 663 experiments found large errors in non-experimental estimates of ad lift [5]. Both studies used the platform’s own experiment data.

Marketing mix modelling (MMM) uses aggregate spend and sales over time, so it needs no user-level tracking. It is slow and noisy, but it sees every channel at once. Google made its open-source MMM, Meridian, available to everyone in January 2025, and it can take lift-test results as priors [6]. That is the right pattern: experiments calibrate the model, and the model fills the gaps between experiments.

Underneath all three, keep one number no attribution model can argue with: blended CAC, total marketing spend divided by total new customers. That is why I treat it as the ceiling. The bank account does not use last-click attribution.

Where generative AI helps

Generative AI is now built into the ad tools. In October 2023, Meta rolled out background generation, image expansion and text variations to all advertisers. In Meta’s survey of early testers, half estimated the tools would save them five or more hours a week [7].

Where I see real value:

• Format work: resizing and reframing one asset for feed, stories and short video.
• Variations on proven concepts: hooks, captions and thumbnails, generated in volume and then tested properly.
• Drafting: scripts, creator briefs and first-pass copy for a human to edit.
• Analysis: tagging every ad by angle, format, hook and offer, so you learn which attributes win across hundreds of ads, not just which single ad won.

The last one is the most underused. A winning ad is a result. A winning attribute is a lesson you can reuse.

Where it hurts

• Sameness. If competitors use the same tools on similar prompts, ads converge. Converged ads stop selecting distinct audiences.
• Invented claims. Generated copy can state benefits your product does not have. The advertiser is responsible for every claim, so a human checks each one against evidence before it runs.
• Fake social proof. In August 2024, the US Federal Trade Commission (FTC) announced a final rule banning fake reviews and testimonials, including reviews that claim to come from someone who does not exist, such as AI-generated fake reviews [8]. Anyone selling to US consumers should treat generated testimonials as off limits.
• False winners. More variants mean more chance wins. Test 50 variants that truly do nothing at 95% confidence and you should still expect 2 or 3 to look like winners. Retest before you scale.

The test I apply to every AI tool, here and in my work at Iteranium, is simple: what does it do to acquisition cost, conversion and margin?

Mistakes I see most often

• Reading platform return on ad spend as incremental profit.
• Calling 20 variations of one idea “creative testing”.
• Scaling a creative after a handful of purchases.
• Letting generated copy go live without a claim check.
• Never running a holdout, so nobody knows what the biggest channel actually adds.

What I would do this week

1. Write your breakeven CAC next to your blended CAC for the last 30 days.
2. Split the creative plan into concepts and variations, and budget each concept at target purchases times CAC.
3. Book one lift or geo holdout test on your largest channel.
4. Use generative tools for resizing and variations, and route every claim through a human check.
5. Audit where every review and testimonial on your ads and store came from.

References

1. Apple Inc., “iOS 14.5 delivers Unlock iPhone with Apple Watch, more diverse Siri voice options, and new privacy controls,” Apple Newsroom, Apr. 26, 2021.
2. Apple Inc., “User privacy and data use,” App Store, Apple Developer.
3. Meta Platforms, Inc., “Meta Q4 2021 earnings call transcript,” Feb. 2, 2022.
4. B. R. Gordon, F. Zettelmeyer, N. Bhargava, and D. Chapsky, “A comparison of approaches to advertising measurement: Evidence from big field experiments at Facebook,” Marketing Science, vol. 38, no. 2, pp. 193-225, 2019.
5. B. R. Gordon, R. Moakler, and F. Zettelmeyer, “Close enough? A large-scale exploration of non-experimental approaches to advertising measurement,” Marketing Science, vol. 42, no. 4, pp. 768-793, 2023.
6. H. Nair, “Meridian is now available to everyone,” Google Ads & Commerce Blog, Jan. 29, 2025.
7. Meta, “Generative AI features for ads coming to all advertisers,” Meta for Business, Oct. 4, 2023.
8. Federal Trade Commission, “Federal Trade Commission announces final rule banning fake reviews and testimonials,” press release, Aug. 14, 2024.