AI for Better Ad Creative: 3 Steps to Better Results

Struggling to combat ad fatigue without burning out your design team? Wonder how your small brand can generate the ad volume Meta's algorithm now requires?

In this article, you'll discover a three-step system for using AI to produce high-quality image and video ad creative at scale.

This article was co-created by Fraser Cottrell and Michael Stelzner. For more about Fraser, scroll to the end of this article.

Why Most Marketers Are Wrong About AI Ad Creative

Two misconceptions stop most marketers from using AI tools for ad creative before they start.

The first is that using AI is lazy. Fraser Cottrell, CEO of Fraggell, a direct-to-consumer ad creative agency, pushes back. Getting AI to produce what you actually want during the creative process requires significant effort.

The second misconception is that AI produces low-quality creative. AI is only as good as the context and instructions you give it. In truth, current models produce images nearly indistinguishable from professional photographs. Video models still have room to improve, but for static-image ads, quality is no longer the barrier.

The upside when AI and ad creative work together: it levels the playing field.

For e-commerce brands that previously had to pay for a studio shoot or hire a freelancer to get decent product images, that's a fundamental shift. Product images that once cost hundreds or thousands of dollars to produce can now be generated for a couple of cents.

Finally, advertisers need genuinely different ad variations, and AI makes that volume achievable without a massive team or a huge production budget. This matters now, because Meta's Andromeda update ended the practice of running hundreds of slight variations of the same ad. The platform now treats those as a single creative.

None of this works, though, without one foundational step: training generative AI on your brand on who your customers are, what your brand stands for, and what a great ad looks like.

The goal of the three steps below is to systematically build that context before you ever begin using AI tools for ad creation.

#1: Build Your Brand Knowledge Base With Deep Research

The first thing Fraser does when any new brand comes to work with Fraggell is run a deep research session in ChatGPT or Gemini. Deep research is a function available in LLMs that prompts the AI to thoroughly browse the internet before responding with a well-researched document rather than a quick answer.

The objective is to create a comprehensive external view of your brand. Fraser instructs the AI to find out who's buying the product and why, but also, just as importantly, why people who encountered it chose not to buy. He wants to know what people are saying about the product on Reddit, the most common complaints and issues, and where the customer base is geographically concentrated. These specific angles surface the objections and pain points that make an ad creative resonate with real buyers.

Rather than writing the deep research prompt himself, Fraser uses a voice dictation tool called Whisper Flow to dictate his request to Claude, asking it to write a deep research prompt for Google Gemini and specifying which areas to cover.

Write me a deep research prompt for Google Gemini.
The goal is to build a comprehensive external profile of [Brand Name], which sells [brief product description].
The prompt should instruct Gemini to research who is buying this product and why, as well as who encountered it and chose not to buy, and what stopped them.
I want it to pull what people are saying on Reddit, identify the most common complaints and issues, surface the biggest objections and pain points, and find where the customer base is geographically concentrated.
Format it so Gemini knows to browse the internet thoroughly before responding with a well-researched document, not a quick answer.

For the research itself, Fraser most often uses Gemini. Gemini, built on Google's data infrastructure, tends to return larger, more thorough reports and runs faster. ChatGPT performs well but can hallucinate.

The output from Gemini is a substantial document about your brand. It can be potentially several pages long and will take a few minutes to generate. Once you have it, review it for accuracy before you use it as a training source.

Fraser's technique for verifying it quickly: paste the document into Claude and ask Claude to ask you questions about it to confirm the information is correct. Claude works through the document with questions, and you can clarify or correct on the spot.

I'm going to share a research document about my brand.
I need to verify that the information in it is accurate before I use it.
Please read through it and ask me questions — one at a time — to confirm the key facts, claims, and characterizations are correct.
If I tell you something is wrong or needs clarification, note the correction.
Once we've worked through it, give me a summary of any changes or flags from our review.
Here's the document: [paste document].

This is also the moment to add what the internet can't know. The deep research captures publicly available information such as reviews, forum discussions, and social mentions. But you hold knowledge the AI can't find: proprietary customer insights, internal data, the nuances of how your product actually works, and what your best customers say on calls.

Fraser's instruction is to manually feed that information back into the document to fill out what's missing.

The final version should blend the AI's external research with everything only you know.

#2: Train a Claude Project on Your Brand

Once the deep research document is reviewed and accurate, load it into a Claude Project — a dedicated workspace within Claude with its own persistent memory and context, completely separate from your regular Claude conversations. Think of it as a fresh AI instance that retains only what you've given it. Every conversation within the project draws from that knowledge base, and nothing bleeds in from your other Claude chats.

The starting contents Fraser loads into every project are:

Deep Research Document: This is the foundation built in Step 1, reviewed and corrected.

Customer Reviews and Testimonials: Export reviews directly from your store, pull them from a third-party review platform, or export them from a spreadsheet as a CSV file. This voice-of-customer language is some of the most valuable creative input you can give the AI.

An Internal Brand Document: Fraggell has a roughly ten-page internal document that every new team member receives — covering what the agency stands for, who they are, what they do, and what constitutes a good ad. Fraser feeds this same document into every client's Claude project. The definition of a good ad is entirely brand-specific, and making it explicit gives the AI a benchmark to aim for rather than guessing.

Past Ad Performance Data With Visual Context: Export your performance data from Facebook Ad Manager and add it to the project along with visual context.

Fraser takes the top ten performing ads from the past quarter and uses a tool called Poppy, which physically watches short-form video (rather than just transcribing it) and therefore understands visual elements, pacing, and on-screen action. He feeds each video into Poppy along with the performance data and asks it to summarize what made the ad perform the way it did and identify key visual elements.

Those learnings go into the Claude project alongside the numbers.

Pro Tip: As an alternative, Gemini has built-in vision and can watch a downloaded video directly. Upload the video to Gemini and ask it to describe the creative assets, hooks, and anything that stands out visually. Fraser updates this top-ad analysis quarterly.

#3: Use AI for Ad Creative

With a trained Claude project in place, you can start producing creative. The workflow differs between image and video ads, but the Claude project is central to both.

When you're ready to create the image, Fraser recommends the hybrid approach: generate the image with AI and add the text manually. This way, if you want to test different copy against the same visual, you don't have to regenerate the image every time.

Generate Ad Copy With Claude

Start by opening your Claude project and asking it to generate copy with a prompt like:

Using the knowledge you have, write me some static ad headlines.
Ask me any questions you need in order to complete this task.

Claude will return a set of headlines drawn from the brand knowledge in the project, potentially pulling in customer language from the reviews you uploaded. From that set, identify two you like and two you actively dislike. Tell Claude to give you more like the ones you want and less like the ones you don't — and explain why.

Because conversations within a Claude project compound over time, this feedback trains the AI's future responses. It learns your preferences progressively with each session.

Brainstorm and Generate Ad Images and Prompts With Claude

Once you have strong copy, ask Claude to brainstorm image concepts to accompany it. Give Claude your ideas and let it respond with its own. The more specific the brief, the more targeted the output.

If you already have a creative direction in mind — a specific message you want to hit and a target persona — bring that into the brief.  For a sports hydration product, that might mean telling Claude:

I want an image targeting marathon runners with the message that they need to be properly hydrated before they run, so they should buy [product].

If you find an ad from another brand that you admire, you can paste or upload it to Claude and ask it to rewrite the concept for your brand.

Should the image feature the product on a clean background? Drag a photo of your product into Claude, so it knows exactly what it's working with, then describe the shot you want in plain language, such as:

Give me a prompt for Nano Banana to use for a professional studio product shot on a purple background, with soft lighting that looks real.

Keep the description fairly brief and invite Claude to ask questions if needed. Claude will write a formatted prompt for your image generation tool that references the product you've provided.

Read through the prompt to confirm it aligns with your direction before taking it to the next step.

Generate Ad Images With Nano Banana in Gemini

Fraggell's current go-to for image generation is Nano Banana 2 Pro, available within Gemini. When using Gemini directly via the chat interface, you'll receive one image per request. Via the API, through third-party platforms like Arcads or Higgsfield, you can generate three or four variations at once.

A few tips Fraser has learned from working inside Gemini:

When saving an image from Gemini, use the download arrow inside the interface to get the full 4K version. Right-clicking and saving gives you a lower-resolution copy. If you're working via the API, you can pass it specific camera commands, such as zoom out or change the camera angle, and the platform will handle the adjustment cleanly without requiring you to rewrite the entire prompt. If an initial image isn't right, paste it back into Claude with a description of what needs to change. Claude updates the prompt, and you run the revised version through Gemini. Once you have an image you're happy with, don't touch it. Going back to tweak even a small detail risks the AI going in a completely different direction and ruining what you had. If the image needs a structural adjustment, like changing the aspect ratio or extending a canvas edge, switch to Photoshop's generative fill instead. It modifies only the targeted area without regenerating the whole image.

Once you have an image you're satisfied with, bring it into your favorite image editing tool and layer the copy on top.

Generate Video Ad Script Drafts With Claude

While Fraggell doesn't produce fully AI-generated videos for clients because the quality of current video-generation tools isn't yet up to the agency's standards, he says AI makes a significant difference in scripting and ideation.

The process uses the same Claude project.

Describe the video concept to Claude: who's in it, the general scenario, and the target length. For a running brand, that might look like:

I have a video idea: a UGC creator running a marathon while talking to the camera about these new shoes.
Based on your knowledge, write me a 30-second UGC script.

Claude will return a timestamped, detailed script. If you only need a starting framework, ask for just a paragraph of script rather than full scene-by-scene direction.

The AI-written script isn't a finished product. Human writers understand nuance, conversational tone, and what resonates with other humans in ways AI still struggles to replicate. But AI is excellent at generating a first draft fast, which gets a creative person 30% of the way to the finish line in a fraction of the time.

A copywriter can take that draft, pull the two lines that work, and build from there.

Other Notes From This Episode

Connect with Michael Stelzner @Stelzner on Facebook and @Mike_Stelzner on X. Watch this interview and other exclusive content from Social Media Examiner on YouTube.

Where to subscribe: Apple Podcasts | Spotify | YouTube Music | YouTube | Amazon Music | RSS

✋🏽 If you enjoyed this episode of the AI Explored podcast, please head over to Apple Podcasts, leave a rating, write a review, and subscribe.

---

Stay Up-to-Date: Get New Marketing Articles Delivered to You!

Don't miss out on upcoming social media marketing insights and strategies! Sign up to receive notifications when we publish new articles on Social Media Examiner. Our expertly crafted content will help you stay ahead of the curve and drive results for your business. Click the link below to sign up now and receive our annual report!