How to Use AI in Marketing Campaigns, Stage by Stage

Most teams using AI in marketing campaigns point it at the cheapest job in the campaign. They ask a model for a landing page or a set of ads, and something comes back that looks finished. Looks is carrying that whole sentence. The page doesn’t know who is buying, what they’ve already tried, or which option is open in the next tab.

On Copy Chief’s Escape Velocity campaign, I gave AI a different job. Copy Chief is a membership platform for working copywriters, based in St. Petersburg, Florida. Escape Velocity was a Meta Ads campaign sending traffic to a mobile-first funnel. AI did the research. People made the calls. Escape Velocity returned 55.03x average return on ad spend, and its cost per acquisition fell from $42 to $22.

Below is the map I use for AI in marketing campaigns on client work: which stages get a model, which stay with a person, and how to tell them apart.

Where does AI in marketing campaigns actually help?

Short answer: AI in marketing campaigns helps most where the work is reading and sorting at volume, or turning out first drafts. It helps least where the work is judgment about one buyer, one offer and the budget behind it. Most campaigns mix the two, so the useful question isn’t “should we use AI?” It’s “which stage is this?”

Here’s how I split a campaign into seven stages, and who does the work at each one.

Read the table top to bottom and a pattern shows up. The strongest uses of AI in marketing campaigns sit early, in research and drafting. The stages where money moves (the budget, the offer, what the page promises) stay with someone who will answer for the result.

Why is a generated landing page the wrong first job?

The most common use of AI in marketing campaigns is the weakest one. A generated page is an average of every page like it. That’s not an insult to the model. It’s how the model works: it predicts the most likely next word from everything it has read, and “most likely” is another way of saying “most common.”

Averages rarely convert, because the buyer isn’t average. Picture a working copywriter deciding whether to spend rent money on a course. She’s at a specific stage, carrying a specific frustration, comparing the offer against something she already knows. A model can’t see any of that from a one-line prompt. It fills the gap with plausible copy.

Plausible copy is the most expensive kind. It passes the internal review, then fails quietly once it’s live. The sameness is measurable, too. In a study with 36 participants, people who brainstormed with ChatGPT produced ideas that were less distinct from one another than people using a different tool (Anderson, Shah and Kreminski, 2024). If your competitors prompt the same model with the same request, you all get the same page.

On Escape Velocity, the page design was never the constraint. Knowing what the page had to say was the constraint. So that’s where the AI went.

How do you use AI for customer research without trusting it blindly?

Customer research is where AI in marketing campaigns pays off first, if you start narrower than feels comfortable. Not “B2B companies.” Not “copywriters.” One stage, one pain, one kind of buyer you can picture. Research aimed at a broad label returns broad answers, and broad answers produce the same generic page you’d have got from the model anyway.

Then send the research outward, not inward. Asking a model what your buyers worry about is asking it to guess. Goli and Singh tested GPT-3.5 and GPT-4 as stand-in survey respondents. The models were less patient than real people when choosing between rewards, and the authors warn that asking the model directly can mislead, though it helps with ideas worth testing (Goli and Singh, 2024).

A second study, with 3,200 human participants and four language models, found the models misportrayed and flattened the groups they were asked to stand in for (Wang, Morgenstern and Dickerson, 2025). A synthetic buyer persona is a guess with a name on it.

What does work is pointing the model at things real buyers wrote. Timoshenko and Hauser compared customer needs found in online product reviews with needs found by professional interviewers. The reviews were at least as valuable a source, and machine learning made the review faster by filtering out repetitive and empty content (Timoshenko and Hauser, 2019). That’s the job I give AI in customer research: reading and sorting, not imagining.

On Copy Chief, my customer research ran on four inputs:

1. Pain mining in Perplexity. I used its research feature because each finding comes back with a source attached, so I can open the thread and read it myself.
2. A custom Python Reddit scraper. It pulled comments and fed them to a language model, which sorted each one into pains and desires.
3. Coach interviews. The coaches hear member problems firsthand, in a way no scraper does.
4. Member email outreach. Members answered in their own words, which is the raw material for headlines.

The model handled the volume. It never had the last word. A pain that showed up in the scraped comments but nowhere in the interviews or emails got a question mark, not a headline. Language models can write fluent text that their sources don’t support, which researchers call hallucination (Ji et al., 2023). So every finding has to point back to something a person actually wrote or said.

A rule I keep from that project: five real problems, each traceable to a source, beat fifty invented ones.

How does ad creative research turn pains into messages?

Ad creative research is the step between customer research and the ad itself. Most teams skip it, because the pains feel obvious once written down. They aren’t obvious until you’ve chosen one to lead with.

The method is plain. Each confirmed pain becomes one headline, written as close to the buyer’s own words as you can get. That headline goes on the ad, at the top of the landing page and in the follow-up email. One research pass feeds three places.

The match between ad and page isn’t decoration. People judge where to go next from the cues in front of them, which Pirolli and Card called information scent (Pirolli and Card, 1999).

AI’s part in ad creative research is the middle step. I give it the confirmed pains with their source quotes and ask for groupings, how often each pain appears, and four ways to phrase each one using only words from the quotes. Without that last limit, the model drifts back toward the phrasing it has seen on every other ad.

On Escape Velocity, the static and animated Meta creative was split into top-of-funnel and bottom-of-funnel sets, each built around the buyer’s pain points. Escape Velocity’s click-through rate went from 0.9% to 3.0%, and its landing page converted at 10.1%. One honest limit: the creative and the landing page changed in the same campaign, so I can’t tell you how much of the 55.03x return belongs to each.

That research-first approach is how I run paid social creative and media buying for clients.

What belongs in a campaign brief that AI helps write?

A campaign brief is the document everyone works from: the person writing ads, the person building the page and the person buying media. If the brief is vague, each of them fills the gaps with a guess, and the guesses don’t match.

AI is good at drafting a campaign brief from research notes, because that’s a summarizing job. It is not good at deciding what goes in it. Here’s the brief I use, with who owns each line:

• The one segment. One sentence naming the buyer and their stage. A person decides; AI can tidy the wording.
• Confirmed pains, with sources. Three to five, each linked to where it came from: a comment, an interview or a member email. AI drafts this from the research file.
• The lead promise. The single pain the campaign opens with, chosen during ad creative research. A person decides.
• Proof. Only results the client can stand behind, with the caveat attached. AI never supplies a number.
• The offer and the price. A person decides.
• Objections. What stops the buyer, in their words. AI drafts; a person checks.
• The result that counts. Purchases, booked calls or paid subscriptions, not likes. Decided before launch.
• Channels and audiences. Where the ads run and to whom. AI can summarize past platform data.

“The result that counts” is the line campaigns get wrong most often. On Copy Chief’s Membership Funnel Redesign, creative had been judged on engagement instead of subscriptions, so the best-performing ads were the ones least likely to sell. Rebuilt around paid subscriptions, the Membership Funnel Redesign returned 31x on ad spend with 215 paid subscriptions from August to October 2023. It ran to a deadline, so 31x belongs to those three months, not to the funnel forever.

In my experience, a model given a one-page campaign brief with sourced pains writes much closer drafts than a model given a product name and a tone word. The brief is the prompt.

Where do AI variants save the most time?

Variants are where AI in marketing campaigns saves real hours. In a preregistered experiment with 453 college-educated professionals, access to ChatGPT cut the time spent on writing tasks by 40% and raised rated quality by 18% (Noy and Zhang, 2023). Headlines, hooks, subject lines and short ad text are that kind of task.

Tailored versions help too. Across four studies with 1,788 participants, messages ChatGPT matched to a reader’s psychological profile were more persuasive than generic ones (Matz et al., 2024). In a B2B campaign, that usually means one version per role in the buying group: the owner, the finance lead and the person who’ll use the thing daily.

More tailoring isn’t automatically better, though. In a field experiment at an online travel firm, retargeted ads showing the exact product a visitor had viewed did worse on average than generic brand ads, until the visitor’s preferences had narrowed (Lambrecht and Tucker, 2013). Specific copy suits buyers close to deciding. Early buyers respond to the broader promise.

Here’s how I run variants:

1. Paste the whole campaign brief into the prompt, not just the product name.
2. Ask for 20 versions of one element at a time, each tied to one pain from your ad creative research.
3. Cut to three or four by hand. Delete any line a competitor could run unchanged.
4. Test those few against each other with real budget, judged on the result the brief named.

The gains aren’t even across people. In a study of 5,172 customer-support agents, an AI assistant raised productivity by 15% on average, but the most experienced agents saw small drops in quality (Brynjolfsson, Li and Raymond, 2025). That matches what I see: a senior marketer’s hours are better spent cutting and choosing than drafting.

Setting that production side up as a repeatable system is what my AI marketing systems work covers.

Where does AI in marketing campaigns not help?

Some of this is about risk. Some of it is that the model has nothing to work from. Here’s what I keep away from it on client campaigns:

• Choosing the offer. The price, the guarantee and the payment terms are business decisions with margins attached. A model doesn’t know your margins.
• Inventing proof. AI never writes a result, a testimonial or a statistic. If a number isn’t in the client’s own reporting, it doesn’t go in the ad.
• Standing in for the buyer. “Ask the AI what a CFO thinks” gives a confident answer with nothing behind it, which is what the Goli and Singh and Wang studies above describe.
• Regulated claims. Health, finance, legal and age-restricted categories need a person who knows the rules reading every line.
• Moving budget. A model can summarize what each ad set returned. Deciding which campaign gets next month’s money is a person’s call.
• Final approval. Someone whose name is on the result reads every word before it goes live.

There’s also a quieter failure. Using AI in marketing campaigns to make more of the wrong thing, faster, is the most common misuse I see. If the segment is too broad or the offer is weak, 20 polished variants just spend the budget sooner. Fix the campaign brief first.

The same goes for the build. Page structure, the order of proof and the form fields are design decisions, and the gap between the click and the checkout is where money leaks. That’s the part of funnel development I don’t hand over.

How do you measure whether AI in marketing campaigns is paying off?

You can’t measure the value of AI in marketing campaigns if you can’t measure the campaign. Before launch, every ad link needs a UTM tag, every form needs to carry those tags into the CRM, and the landing page each lead arrived on needs to be recorded. That’s the lead attribution method I set up on client sites.

Then compare like with like. Run an AI-assisted variant against a human-written one, changing only the copy, and judge both on the result your campaign brief named. If you change the research, the creative and the page all at once, as Escape Velocity did, you get a better campaign but no clean answer about which part did the work.

On the AI side, I track two numbers: hours from the start of customer research to an approved campaign brief, and the share of tested variants that started as AI drafts and won. If the first number drops and the second holds up, the model is earning its place. If time drops but nothing AI-drafted ever wins a test, it’s producing volume, not results.

What does an AI-assisted campaign cost, and how long does it take?

The tools are the cheap part of AI in marketing campaigns. Most of the cost is the hours spent on customer research, the campaign brief and the build, and those hours don’t shrink much because a model drafted some copy. What shrinks is the sorting and drafting in between.

On Copy Chief’s World of Financial Copy promotion, the funnel (ClickFunnels, ThriveCart and Keap) was built and launched in one week, and the World of Financial Copy funnel took 56 orders at a 4% conversion rate. Copy Chief’s Crossroads Summit event funnel launched in 14 days, and Crossroads Summit returned 10x on event ticket sales. In my experience the build is rarely the slow part. Agreeing on the message is.

These fixed quotes usually apply to a campaign like this:

• Landing Page Design: from CA$3,795, 1 to 2 weeks.
• Lead Source Reporting: CA$2,600, 2 weeks, so every lead arrives tagged with its campaign and ad.
• Paid social creative and media buying: a fixed quote after a call, because spend and scope vary too much to list.

Every quote is fixed before work starts, with two revision rounds per phase. The full list is on my pricing page and calculator, and the Copy Chief case study shows the funnel work in detail.

Want a second opinion on your next campaign brief?

Not a pitch. A 30-minute call. Bring the campaign you’re planning, or the one that’s live and flat, and I’ll tell you which stages I’d hand to AI, which I’d keep, and what I’d check first. When AI in marketing campaigns has meant faster drafts and the same results, the gap is usually upstream, in the customer research.

Book a 30-minute call with me.

Sources

Peer-reviewed research

• Anderson, B. R., Shah, J. H., and Kreminski, M. (2024). Homogenization effects of large language models on human creative ideation. Creativity and Cognition (C&C ’24), ACM, 413-425. https://doi.org/10.1145/3635636.3656204
• Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942. https://doi.org/10.1093/qje/qjae044
• Goli, A., and Singh, A. (2024). Frontiers: Can large language models capture human preferences? Marketing Science, 43(4), 709-722. https://doi.org/10.1287/mksc.2023.0306
• Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Madotto, A., and Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38. https://doi.org/10.1145/3571730
• Lambrecht, A., and Tucker, C. (2013). When does retargeting work? Information specificity in online advertising. Journal of Marketing Research, 50(5), 561-576. https://doi.org/10.1509/jmr.11.0503
• Matz, S. C., Teeny, J. D., Vaid, S. S., Peters, H., Harari, G. M., and Cerf, M. (2024). The potential of generative AI for personalized persuasion at scale. Scientific Reports, 14, 4692. https://doi.org/10.1038/s41598-024-53755-0
• Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192. https://doi.org/10.1126/science.adh2586
• Pirolli, P., and Card, S. (1999). Information foraging. Psychological Review, 106(4), 643-675. https://doi.org/10.1037/0033-295X.106.4.643
• Timoshenko, A., and Hauser, J. R. (2019). Identifying customer needs from user-generated content. Marketing Science, 38(1), 1-20. https://doi.org/10.1287/mksc.2018.1123
• Wang, A., Morgenstern, J., and Dickerson, J. P. (2025). Large language models that replace human participants can harmfully misportray and flatten identity groups. Nature Machine Intelligence, 7(3), 400-411. https://doi.org/10.1038/s42256-025-00986-z