Since VEO 3 launched in May 2025, it has been possible to generate stunning ads with quality close to that of the biggest studios, simply from a prompt.
See the game-changer? An ad worthy of Nike or Decathlon, generated in 2 minutes, cheaper than a coffee (€0.20).
For example here with these Decathlon, Starbucks or Ikea ads:
And that is exactly what VEO 3, Google's video model, makes possible.
In this article, I show you how to use it to generate AI ads for yourself:
- with prompts optimized in JSON
- ultra-realistic scenes
- and a simple iteration loop to refine the result.
A closer look at the tool
You start from a prompt, you get an ad.
For this, we will use VEO 3 and you have several options to test it. I recommend going through Flow. You can also go through Krea or Leonardo, but the prompt length is limited and some features are missing.
By the way, in summer 2025, Flow let you test the tool for free for 1 month.
You will be able to generate complex scenes from text, images, or a mix of both.
The promise: create a video ad from a simple structured prompt.
The recommended format is a JSON prompt because you specify each parameter precisely:
- visual style: vintage, realistic, animated, …
- camera movement: zoom, tracking shot, …
- sound atmosphere: music, dialogue, silence
- objects to include in the scene
- final format: square, vertical, 4K, etc.
And it works! Some generated videos look like Nike or Apple ad spots.
Add to that native audio integration and dialogue you can drop in directly (English only, as of summer 2025), and you get breathtaking results.
Concrete steps to generate your ad
1. Find visual inspiration
You start from an existing video you like (e.g. IKEA, Nike, Patagonia…) or from an idea. Analyze the scene, the objects, the mood, and note the elements you want to reuse.
2. Create your prompt with ChatGPT
Give ChatGPT an instruction like:
“Here is a JSON prompt for an IKEA ad. Make me the same one, but for X, in the style Y”
ChatGPT generates a new structured prompt for you, in English (required for Flow). You can also include any corrections you want:
“Change this, place the objects on the sides, remove the visual duplicates…”
3. Test the prompt on Flow in the Fast version
You have several models on Flow. I advise you to use VEO 3 - Fast before switching to VEO 3 -Quality if you are satisfied with your prompt (cost x5).
1 generation = ~20 credits (out of 1,000)
4. Add a first frame to guide the result
You can import a realistic image created with ChatGPT's model or another image model as the first frame of the generated video. You can also add a last frame.
In summer 2025, Flow rejected photorealistic images of people as a first frame in some countries.
5. Refine the iterations
Fix the lighting, the inconsistent objects (reversed handlebars, strange textures, …). Use ChatGPT to help you tweak the prompt each time without starting over.
Remember to lock the result with first/last frames.
6. Edit the audio with Eleven Labs
You can even add dialogue in the prompt (English only as of summer 2025), and the audio is generated by default.
With Eleven Labs, you can generate sound effects or replace the AI voice with your own (or a cloned voice).
I explain this whole process in a video on my YouTube channel.
My experience
Just like when using a text AI model, the same prompt can generate different results. It is a bit frustrating because you do not have full control and it often takes many iterations if you have something precise in mind. Adding a first/last frame helps constrain the model (along with iterating on the prompt).
To avoid burning through your credits, I advise you to first iterate with VEO 3 Fast.
Once you are satisfied with the prompt, switch to the Quality version.
I created this GPT assistant that lets you easily generate your JSON prompt in 1 click, with your own constraints.
What's next?
In August 2025, generation was already impressive. But a few limits remained:
- objects sometimes poorly placed
- logos still poorly interpreted (except the most famous ones)
- dialogue in English only
But given the pace of progress, all of this will disappear very soon.
October 2026 update: Google launched Veo 3.1 in October 2025, with richer sound, now also generated from a first and last frame. Flow's pricing, checked in October 2026, has not changed: 20 credits per generation in Fast, 100 in Quality, out of the 1,000 monthly credits in Google AI Pro. I compare Veo 3.1 with other models in my AI video generator comparison.
We are heading toward a world where you test 10 versions of an ad per week, adapt the ad to each audience (TikTok style for Gen Z, more understated for LinkedIn), and iterate live from a landing page or an automatic A/B test.
The barrier to entry for quality videos is collapsing.
And iteration will become your best creative lever.
In short, video content creation is made simpler with AI, letting you build a true content machine. For a few cents, you can create a convincing, personalized ad that is testable right away.
Give it a try. Take a product, a staging idea, and play with the JSON prompt.
That's where the magic happens!
