From Performance Max's black-box bidding to Coca-Cola's uncanny AI holiday ad, marketing bet on AI years before it had reliable measurement to prove any of it works.
In December 2023, Toys "R" Us released a brand video generated almost entirely with OpenAI's Sora, produced with ad agency Native Foreign while the model was still in early access. It was, at the time, one of the most prominent uses of generative video in advertising — and it landed badly. Viewers described the toy-store founder's AI-rendered childhood as uncanny, waxy, emotionally hollow. A year later, Coca-Cola ran into the same wall with an AI-generated remake of its "Holidays Are Coming" truck ad, produced by Silverside AI, and got a similar backlash: technically impressive, aesthetically off, and read by a lot of viewers as a company being cheap with its most iconic asset.
Both incidents get cited constantly as cautionary tales, and they are. But they also obscure where AI in marketing has actually succeeded — quietly, unglamorously, in the parts of the ad stack nobody storyboards: bidding algorithms, audience segmentation, and copy testing. The generative-creative failures are visible because they're creative. The bidding and targeting successes are invisible by design — that's the whole pitch of automation. Marketing is a useful vertical to study precisely because it shows both ends of the AI adoption curve at once: mature, load-bearing automation in the parts of the job that are fundamentally statistical, and repeated public failure in the parts that are fundamentally about taste.
The least visible and most entrenched use of AI in marketing is automated bidding. Google's Performance Max (PMax), rolled out broadly in 2022, is the clearest example: advertisers hand over a budget, some creative assets, and a conversion goal, and Google's models decide in real time how to allocate spend across Search, YouTube, Display, Discover, Gmail, and Maps, and which combination of headlines, images, and video to serve to which user. Meta's answer is Advantage+ shopping campaigns, launched in 2023, which does the equivalent across Facebook and Instagram inventory — automating targeting, placement, and creative selection for e-commerce advertisers. By 2024, Meta was reporting that a majority of its advertisers' incremental ad spend was flowing through Advantage+ tooling.
The mechanism is genuinely just applied ML: multi-armed bandit-style exploration across creative and audience combinations, bid shading based on real-time auction dynamics, and conversion prediction models trained on the platform's own first-party data. None of this requires a language model — it's the same family of techniques that has powered ad exchanges for a decade, just with more inputs and tighter feedback loops.
The honest limitation, which advertisers complain about constantly, is opacity. PMax in particular became notorious for giving buyers almost no visibility into where their ads actually ran or which assets drove which conversions — the trade-off for automation is losing the placement-level reporting that manual campaign management used to provide. Google has incrementally added more reporting granularity in response to advertiser pressure, but the fundamental shape of the trade — cede control, get performance — hasn't changed. Agencies now treat auditing a PMax campaign's actual placements as its own subdiscipline.
Generative AI in ad creative split cleanly into two tracks that have had opposite outcomes. Language-level optimization has worked well and quietly for years. Persado, founded well before the current LLM wave, has spent over a decade building models that test and swap individual words and phrases in ad copy, subject lines, and CTAs against measured lift — a customer like JPMorgan Chase uses it to A/B-test marketing language at a granularity no human copywriter would bother with. Jasper AI, which raised a $125 million Series A in 2022 at a $1.5 billion valuation, took the more general LLM-copywriting angle: drafting ad copy, product descriptions, and social posts at volume for brand marketing teams. Neither of these generates anything a customer directly associates with "AI" — the output looks like normal marketing copy because that's the point.
Generative video and image, by contrast, keep producing visible, attributable failures, because the whole value proposition is replacing an expensive, high-craft human process — and viewers are unusually good at detecting when that substitution happened. The Coca-Cola and Toys "R" Us cases aren't outliers; they're the predictable result of using a technology whose failure mode is "almost right but subtly wrong" for a use case where taste is the entire deliverable. Brands are still using generative image and video tools heavily — for storyboarding, for rapid concepting, for localizing a shoot into ten languages instead of reshooting — but the finished, hero-asset use case remains high-risk enough that most large brands treat it as a legal and reputational review item, not just a creative one. There's also a substantiation problem the FTC cares about: generative copy has a documented tendency to invent product claims, and an ad claim that isn't true is a regulatory problem regardless of who or what wrote it.
Dynamic creative optimization (DCO) — automatically assembling different versions of an ad from a library of image, copy, and offer components based on who's viewing it — is one of the oldest applied-ML use cases in adtech, offered commercially for over a decade by vendors like Smartly.io and Celtra. It's mature, well-understood, and not particularly controversial. What's changed is the input side: the signal that DCO and targeting models depend on has been shrinking for years, and that erosion has nothing to do with the sophistication of the models themselves.
Apple's App Tracking Transparency, which shipped with iOS 14.5 in 2021, made cross-app tracking opt-in rather than opt-out; Meta estimated it cost the company roughly $10 billion in 2022 ad revenue alone. Google spent years threatening to deprecate third-party cookies in Chrome on a similar timeline, then in July 2024 reversed course, opting instead for a user-choice prompt rather than a default removal — which is arguably worse for advertisers, since it replaces a predictable migration with an unpredictable, user-dependent one. The net effect is that the personalization models have gotten better at exploiting whatever signal remains, while the total signal available has gone down. That's not a story about AI capability; it's a story about a shrinking data supply chain that better models can't fully compensate for.
This is where marketing's AI story gets genuinely underappreciated. As cookie-based, click-level attribution has degraded, the industry has been pushed back toward media mix modeling (MMM) — a statistical approach, decades old, that estimates channel effectiveness from aggregate spend and outcome data rather than individual-user tracking. Both Google and Meta have released open-source MMM frameworks in the last two years (Google's Meridian and Meta's Robyn), explicitly positioning them as privacy-resilient alternatives to the attribution models that powered the 2010s. The pitch is that Bayesian MMM, done well, can recover much of what granular tracking used to provide.
The catch is that MMM is statistically hungry and interpretively slippery. It needs a long, clean history of spend and outcome data to produce stable estimates, which is a bad fit for fast-moving digital campaigns; it's also easy to build an MMM that confirms whatever channel allocation the team already wanted. Incrementality testing — deliberately holding out a control group and measuring the actual lift — remains the only rigorous ground truth, and most brands run far too few of these tests to properly calibrate the models sitting on top of them. This is, honestly, the least mature of the four application areas below, and the one where the real research problem in marketing AI is sitting largely unaddressed.
| Application area | Core technique | Maturity / adoption | Where it actually breaks |
|---|---|---|---|
| Programmatic bidding & campaign automation | Bandit algorithms, real-time bid prediction | High — default mode for major platforms | Opacity: advertisers lose placement-level visibility |
| Generative ad copy & language testing | NLP-based phrase testing, LLM drafting | High for copy, low for finished creative | Video/image still triggers visible "AI-made" backlash |
| Dynamic creative optimization / personalization | Rules + ML-driven asset assembly | Mature technique, degraded by shrinking signal | ATT and cookie changes shrink usable targeting data |
| Measurement & attribution | Media mix modeling, incrementality testing | Early / actively re-forming | MMM needs data and testing rigor most teams don't have |
Most AI regulation aimed at marketing so far hasn't come from AI-specific law — it's come from existing consumer-protection and election law being applied to AI outputs. The FTC finalized a rule in August 2024 banning fake reviews and testimonials, explicitly covering AI-generated ones, closing off one of the more obvious abuse paths (bot-generated five-star reviews, fabricated influencer endorsements). Separately, a wave of U.S. states passed laws in 2024 requiring disclosure when political ads use AI-generated content — a narrower, faster-moving front than anything in general AI policy, because election law already had disclosure infrastructure to bolt this onto. Marketers should read this pattern correctly: the near-term legal risk isn't a novel "AI ad law," it's existing truth-in-advertising and election-disclosure statutes being applied, with real enforcement teeth, to AI-assisted output.
AI in marketing succeeded first and most durably in the places where the job was already statistical — bidding, targeting, phrase-level copy testing — and struggled visibly in the places where the job is fundamentally about taste and trust, like hero creative. The industry's harder problem right now isn't generating more content faster; it's that the measurement infrastructure needed to know whether any of this actually moves revenue has been quietly falling apart for four years, and MMM is a real but immature patch, not a full replacement. If you're evaluating an AI marketing tool in 2026, the sharper question isn't "can it generate creative" — nearly everything can now — it's "can it tell me, with a control group and a straight face, whether the creative worked."