AI ad creatives for paid media: how to generate variations and test what sells

To use AI creatives in paid media, start from clear hypotheses about what makes your customer buy, use AI to generate variations that change one thing at a time and judge each variation by the qualified leads and sales it brings in, not by clicks. AI solves the production bottleneck. Deciding what to test and what's worth keeping is still marketing's job.

This article shows how to set up that process. It's part of the guide to AI in marketing for B2B companies.

Why have creatives become so important in campaigns?

Ad platforms now handle much of the targeting on their own. Instead of choosing every detail of the audience, the advertiser defines the goal and the algorithm looks for the people most likely to respond. In this scenario, the creative is part of the targeting: the message and the image help define who gets interested.

That has a direct effect. A creative that speaks to the right buyer tends to attract the right buyer. A generic creative attracts anyone. That's why testing creatives has become one of the most important levers in paid media, and why AI, which makes producing variations cheap, gets so much attention.

The point many companies forget: testing more only helps if you read the test well. For B2B sellers, that means reading results past the lead. The B2B media context is in paid media for B2B.

The parts of a creative you can test

Before generating any variation, break the ad into parts. Each part is a possible hypothesis.

Change one part at a time. If the hook, the image and the offer all change together, the result won't tell you what worked.

How to generate and test creatives with AI, step by step

1. Build a message bank with the sales team

The best source of ad angles is the sales conversation. Ask the people who sell: why customers close, what makes them hesitate, what words they use to describe the problem. Record the answers in a simple document. That material is what will feed the AI.

Without that bank, AI generates plausible ads that look like any competitor's.

2. Write the hypotheses

A hypothesis has three parts: what changes, why we think it will work and how we'll measure it. Example: "Swapping the generic hook for a question about leads going cold on WhatsApp should attract sales managers with that problem. We measure by cost per qualified lead."

3. Generate the variations with AI

With the message bank and the hypothesis, ask AI for variations of only the part being tested. A useful prompt includes: the audience, the pain, the angle, the brand's tone, what can't be promised and the format. For images, describe the setting, the style and the brand's required elements.

Generate more options than you'll use and pick the best. AI is fast; curation is the job.

4. Review before uploading

Every piece goes through a person who checks:

• Promise. Does the ad say something the company delivers? Does it have numbers or results without proof?
• Brand. Are colors, logo, tone and language on brand?
• Image. Is there wrong text in the image, a product that doesn't exist, distorted hands or faces?
• People. Does the image suggest a real customer or a testimonial? If so, don't use it.
• Platform policy. Exaggerated promises, comparisons and certain topics are restricted under each platform's ad rules.

5. Structure the test

Few variations at a time, enough budget for each one to generate a signal and time for the algorithm to exit the learning phase. If the campaign generates few conversions, test fewer things at once. Platforms also offer features that generate and combine variations automatically. They can be useful, as long as you know what's going live and can read the results. The testing logic applies to the whole funnel, as we show in A/B testing in the sales funnel.

6. Read results in terms of sales

This is the difference between a useful creative test and a misleading one. Read in layers:

1. Attention signals: click-through rate, video retention. They show whether the piece grabs attention.
2. Interest signals: cost per lead, landing page conversion rate. They show whether the call to action works.
3. Sales signals: cost per qualified lead, per opportunity and per customer. They show whether the piece attracts people who buy.

Decide based on the layer closest to the sale that has enough volume. A creative with cheap clicks and zero opportunities lost, even if it looks like a winner in the platform dashboard.

For that, the lead has to reach the CRM with the source ad recorded. And, if possible, the sale needs to flow back to the platform, as we explain in optimizing campaigns for sales.

7. Record what you learned

Keep a simple log: hypothesis, variations, result in each layer and decision. Over time, that log becomes the best briefing for AI, because it shows what works with your customer.

Want to test AI creatives by reading results in sales, not clicks? See Nexo's AI for marketing

Example of a test round

An illustrative example, without numbers, to show the logic. A company that sells management software for clinics runs an ad with the hook "Organize your clinic with a single system." The ad generates leads, but sales complains: lots of small players, no budget, just curious.

The team talks to the sales reps and finds out that the customers who close usually have a specific problem: they lose patients because of scheduling failures. That becomes the hypothesis. With AI, they generate three hooks about that problem, keeping the same image, the same offer and the same page.

When they read the results, one of the new hooks has more expensive clicks than the original ad. It looks worse. But when they cross-reference with the CRM, leads from that hook become opportunities more often, and the cost per opportunity is lower. It becomes the new control, and the next round tests the offer.

Notice what made the difference: the hypothesis came from sales, AI sped up production and the decision was made with sales data.

AI-generated image or real photo?

It depends on what the ad needs to prove. Generated images work well for illustrating a concept, creating scenarios and testing visual styles. To show the product, the team or the result of a service, a real photo is usually more honest and more convincing.

A practical rule: if the image could lead someone to believe something untrue about the company, use a real photo or change the image.

Common mistakes

• Generating dozens of variations without a hypothesis.
• Changing several parts of the ad at the same time.
• Picking the winner by the cheapest click.
• Uploading a piece without reviewing promise and brand.
• Using an image that looks like a real customer or testimonial.
• Not recording the source ad in the CRM.
• Ending the test before there's enough signal.

Conclusion

AI removed the limit on creative production. The limit now lies in the quality of the hypotheses and in how results are read. For B2B marketing, that means starting from what sales hears from customers, testing one thing at a time and deciding by what becomes opportunities and sales.

With the source ad recorded in the CRM and sales fed back to the platforms, each test round makes the next one better. Without that, AI only helps you burn through the budget faster.

Want to know which ads are turning into sales? Take Nexo's free assessment.
Take the free assessment