What Does Generative AI Actually Change in Marketing? One Systematic Review Has the Answer

A friend of mine who works in brand marketing posted something on her WeChat Moments recently.

She said her team had used AI this year to write three thousand pieces of product copy and generate two thousand images, saving roughly four hundred thousand yuan in outsourcing fees. But her last line was: "Sure, we saved money — but I feel less and less sure of myself."

I didn't rush to like the post. Because that "less sure" feeling is, in 2026, a near-universal knot in nearly every marketer's stomach.

Generative AI moving into marketing has been the hype for three years now. But the louder it gets, the less anyone can actually explain: what has it really changed about marketing? Is it just a tool upgrade, or has the underlying logic shifted? Which gains are real, deployable dividends — and which are just demo-era fireworks?

The good news is, someone has done the math for us.

A Review That Sweats the Details

A systematic literature review published in 2026 in a peer-reviewed computer science journal set out to do exactly this.

So what is a systematic literature review?

Put simply: the researchers didn't go by gut feeling or cherry-pick cases. They set a strict set of criteria, pulled in every study that met them, and then synthesized the findings. This one used the PRISMA method — a recognized procedure in evidence-based medicine and the social sciences, designed to stop you from only selecting the papers that support your conclusion.

They dug through three databases: ACM Digital Library, IEEE Xplore, and Scopus. The time window was wide — 2018 to 2025, a full eight years. And what they ended up with were peer-reviewed, serious studies — not some vendor's white paper.

The advantage of this approach is that the conclusions aren't a startup's PR piece. They're the consensus that shows up again and again across more than a thousand papers.

So what is that consensus?

I'll pull out the three most worth discussing.

First: Text and Images Genuinely Save Money

The first hard conclusion the review reaches: generative AI has made concrete progress in three areas — text creation, image generation, and multimodal advertising.

What do I mean by concrete?

It has moved past the "can it write at all" stage and into the "it writes about as well as a human, and ten times faster" stage. Product descriptions, social-media captions, email subject lines, poster visuals — work that used to require a schedule, a brief, and three rounds of revisions — now produces a first draft in minutes.

On the ledger, that boils down to one line: the marginal cost of creative work has been pushed close to zero.

A creative director used to oversee maybe five pieces of copy a day. Now one person with an AI toolset can do fifty. The extra capacity isn't used to fire people — it's used to cover more SKUs, run more A/B tests, and reach more long-tail scenarios.

On this point, that friend's four hundred thousand yuan in savings already backs it up.

Second: The Flip Side of Saving Money Is Responsibility

But the review didn't stop there. It immediately lays out three hurdles.

The first hurdle: data privacy. What you feed the model is users' behavioral data, chat logs, and purchase preferences. Where that data comes from, where it's stored, and who's allowed to use it — every link in the chain is a place you can step on a landmine. The EU's GDPR, California's CCPA, and China's Personal Information Protection Law (PIPL) don't go soft when it comes to fines.

The second hurdle: model bias. AI learns from historical data, and historical data carries the biases of the past. If every successful ad case in your training set is built on a male perspective, the new ads the AI produces will most likely also be built on a male perspective. This isn't a technical bug — it's values baked into the algorithm.

The third hurdle: compliance. Every country's advertising laws, every industry self-regulation code, and every platform's content rules are all scrambling to keep up with AI. A line of copy you generate with AI today might get pulled down tomorrow for tripping some new regulation.

These three hurdles, at root, come down to a single line: the money you save will ultimately have to be spent on "how to use it right."

Third: What Really Decides Whether AI Can Take Root Isn't the Technology

This is the part of the whole review I most wanted to talk to you about.

When the researchers explained why some companies can actually put AI to work and others can't, they borrowed two classic theories: the Diffusion of Innovations theory and the Technology Acceptance Model (TAM).

It sounds academic. But translated into plain language, it's one sentence:

Whether an organization uses AI comes down to "is it worth it" and "do we dare."

"Is it worth it" is the core of the Technology Acceptance Model. If an employee feels this thing can help me do my job better, they'll use it; if they feel it adds to my workload and threatens my job, they'll resist.

"Do we dare" is an extension of the Diffusion of Innovations theory. Whether the company culture encourages trial and error, whether leaders are willing to pay for an AI failure, and whether someone inside is advocating for it — these organizational factors matter more than the model's parameter count.

And the two can pull against each other. For instance, a sales director might think AI-written proposals are clearly worth it, but if the compliance department says this can't go out externally, he won't dare use it. The tug-of-war between perceived value and the ethical framework decides how deep a new tool can put down roots inside an organization.

So if you look at the companies that have genuinely put generative AI to work, almost all of them did the same thing: they solved the organizational and ethical problems first, and only then talked about technology. Get the order backwards, and no matter how advanced the tool is, it's just window dressing.

What This Review Doesn't Say

It's a good review, but it has its limits.

It looks back at research from 2018 to 2025 — those eight years. In that time, generative AI went from GPT-2's plausible-sounding nonsense all the way to today, when it can write code, make videos, and run multimodal ad placements. Things are changing so fast that even the researchers themselves admit the conclusions may only stay fresh for six months.

It also can't answer the question every marketer cares about most: should I, today, get on board or not?

This review can't give you that answer — because it depends on the scale of your business, your data assets, your team's digital foundation, and that one person in your company who's willing to make the call.

But I can pull one line out of this review for you:

Generative AI in marketing: the technology is no longer the bottleneck. Organization and ethics are.

If you're still tangled up in "does AI write well enough," you might be asking the wrong question. The right one is: is my team ready to work alongside AI? Can my compliance process keep up with the speed of content production? Can the way I handle user data survive an audit?

If even one of those three questions makes you hesitate, it means you're not ready yet.

Finally

That friend who posted on Moments later direct-messaged me: tell me, was saving that four hundred thousand worth it?

I thought about it for a long time, then sent her one line back:

The money saved is a certain, small gain. What you take on is an uncertain, large risk. Figure out the risk first — only then does the gain count as a real gain.

She replied "got it" — and then said nothing more.

I still don't know whether she really got it, or whether she was just annoyed at me for throwing cold water on her. But this question is worth every person currently using AI for marketing stopping to think about carefully.