--- title: "How to Recreate an Ad You Like With AI (Using Your Product)" description: "Recreate an ad you like with AI and your product photo: two copy-paste prompts for ChatGPT or Nano Banana, a copy sheet, and fixes when it looks off." canonical: "https://marketprompter.com/blog/recreate-an-ad-with-ai" author: "John P. Jochem" datePublished: "2026-09-13" dateModified: "2026-09-13" --- # How to Recreate an Ad You Like With AI (Using Your Product) Decide whether the ad has a slot for your product or only an idea to reuse, start from a photo the model can trust, rewrite the copy yourself, then brief the model on fidelity, integration, and what does not come across. Two short prompts, one for each job. By [John P. Jochem](https://marketprompter.com/about) · 2026-09-13 · 15 min read You found an ad that works. Clean layout, good light, a product sitting in it like it belongs there, copy that makes the offer obvious. You want that ad with your product in it. So you screenshot it, open an image tool, and type "make an ad like this for my product." What comes back is close enough to be annoying and wrong in every way that matters: your product is redrawn, their brand name is still in the corner, and the headline is something the model invented. Recreating an ad you like with AI works when you treat it like a small construction project. Nobody on the crew knows the obvious unless you say it: what stays, what goes, what your product looks like, what the words are, what "done" means. Say those, and the same tool gives you an ad you can run. **Reference + Product photo = Result** ![Reference ad: a fashion poster with large rounded lettering filled with sky, a figure in a puffer jacket and layered jeans leaping through a cloudy blue sky, and a white script tagline](https://marketprompter.com/images/recreate-an-ad-with-ai/reference-fashion-ad.webp) Reference ![Product photo: a blue drink can with tall yellow lettering standing beside its yellow box on a white background](https://marketprompter.com/images/recreate-an-ad-with-ai/drink-can-product-photo.webp) Product photo ![Recreated ad: the blue drink can standing on a concrete ledge against a cloudy blue sky, under large rounded lettering filled with sky and a white script tagline](https://marketprompter.com/images/recreate-an-ad-with-ai/drink-can-recreated-ad.webp) Result ## First, decide which of two jobs this is **The ad has a slot your product can fill.** There is a clear advertised product doing something yours could do: sitting on a surface, held in a hand, next to props, floating on a color field. This is a swap. Keep the ad, replace their product with yours, remove what belongs to them. Most of the work is integration, making your product look photographed into that scene rather than placed on top of it. **The ad does not transfer cleanly.** What you liked is the idea, the energy, the palette, the way the type sits, but there is no sensible slot. Their bottle is mid-splash and yours is a wool blanket. The ad is a face and a tagline. The product is on fire, or worn, or bent. This is not a swap. You recreate the central idea in a form your product can physically be in, with the same composition, light, and finish, changing only what has to change. Squint at the reference and ask: is there one product in there doing something mine could do? Yes is a swap. No is a recreate. Decide before you prompt. A model told to swap into a slot that does not exist will invent one, and a model told to reimagine an ad that only needed a swap will throw away the parts you liked. ## Get a product photo the model can trust The model keeps what it sees, so anything wrong with the photo is wrong in every output, and anything extra in the photo tends to come along. - One product, plain background, cropped close. No hands, no props, no lid sitting beside it unless you want a lid in the ad. - The label readable. Small label text is the thing most likely to come back garbled, and a sharp label gives the model something to preserve instead of guess. - The angle you want in the ad. The model can only show your product accurately from the angle it was photographed; whatever the photo does not show, it invents. If the reference product faces the camera at eye level, shoot yours that way. - The state you sell it in. Cap on, box closed, standing up. A clean phone photo meets all four. A professional photo with a hand in it does not. If the only photo you have is cluttered or dark, the studio prompts in [AI prompts for product photography](/blog/ai-prompts-product-photography) can give you a clean shot from it; check the label against the real product before you use that shot here. ## Find an ad your product can live in [Meta's Ad Library](https://www.facebook.com/ads/library/) is public, searchable by brand, and shows when each ad started running. An ad that has been running for months is one somebody keeps paying for. Your own past ads that performed come with the fewest rights questions. Screenshots you saved because they stopped you are the third source, and the one most people already have. An easy reference has one hero product in a clear slot, lighting and a palette your product would look right in, text in its own zones, and a situation your product can be in without cheating. A hard reference is one where the appeal is a face, a celebrity, a logo, a typographic stunt, or a physical action your product cannot do. Hard references are fine, but they are recreate jobs, not swaps. Before you upload the reference, crop out what you already know has to go: a watermark, a caption bar, a corner logo. Anything you remove now is something the model cannot leak later. ## Rewrite the copy yourself, one line at a time This is the step people skip, and skipping it is where invented headlines and leftover brand copy come from. Hand the model a reference full of words and say "make it for my product," and you are asking an image model to read the ad, work out which words are the brand's, decide which claims are true for you, write replacements that fit the layout, and render them, while also doing the image. It will invent a headline, keep a line that was theirs, or turn their claim into yours. Too many jobs at once. So take the writing away from it. List every piece of text the reference has as an ad: headline, subhead, body lines, call to action, brand name, handle, specs, price. Skip text printed on their product; that leaves with the product. For each line, write what it becomes in your ad at roughly the same length so it still fits, or mark it to be removed. Only facts that are true for you. Their brand and handle become yours or go. Their review, rating, or offer goes; do not write a softer version of it. A review of their product is not a review of yours, and presenting it as one is what the [FTC's rule on fake reviews](https://www.ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commission-announces-final-rule-banning-fake-reviews-testimonials) prohibits. ```text COPY SHEET headline: "[theirs]" → "[yours, similar length]" subhead: "[theirs]" → "[yours]" or (remove) body: "[theirs]" → "[yours]" cta: "[theirs]" → "[yours]" brand: "[theirs]" → "[your brand]" handle: "[theirs]" → "[@yours]" or (remove) spec: "[theirs]" → "[a spec true for you]" or (remove) price: "[theirs]" → (remove) unless you have the same offer ``` The sheet goes into the prompt. The model's only writing job becomes placing your words where theirs were, in the same size and style, and leaving cleared areas clean. Writing the lines is quicker when your product facts and brand voice are already written down; [how to use AI for copywriting](/blog/ai-copywriting-beginner-to-advanced) covers building both. ## Which AI image model to use The job needs a model that edits the images you give it, follows a long set of instructions, and keeps a product's details intact. Two current models do all three well, and both are strong everyday image models too. **GPT Image 2.5, in ChatGPT.** This is OpenAI's newest image model, released in September 2026 as ChatGPT Images 2.5 and available on every ChatGPT plan. It keeps the subject of a reference photo more faithfully than earlier versions, which is what keeps your product looking like your product, and it follows a run of edits in the same chat more reliably, which is what follow-up fixes depend on. In ChatGPT, attach both images to one message and paste the brief. OpenAI notes it can still place text imprecisely in tightly structured layouts, so check where the headline landed before anything else. [OpenAI's announcement](https://openai.com/index/introducing-chatgpt-images-2-5/) has the details. **Nano Banana Pro, in Gemini.** This is Google's Gemini 3 Pro Image model, which Google positions as its choice for the most complex visual tasks. It works through a prompt before it draws, which helps with a brief that has several parts, and it renders text inside images well, which matters when a headline has to look designed rather than typed on. In the Gemini app, choose Create images and the Thinking model. Free accounts get a limited number of these images before Gemini switches to an older model, so check which model made an image before you judge it. [Google's guide](https://blog.google/products-and-platforms/products/gemini/where-to-use-nano-banana-pro/) lists where else it is available. Use whichever you already have. Both let you keep editing the result in the same chat by describing a change. If one keeps changing your label after a couple of tries, run the same brief in the other. ## Brief it like a construction crew A good brief has six parts. Each one exists because of a specific way the output goes wrong without it. **Which job it is.** Right after naming the images, say whether this is a swap or a recreate. Otherwise the model picks, and it can pick wrong. **Who is the authority for what.** The reference is the authority for the idea, composition, camera, scene, lighting, color, and where the type sits. Your photo is the authority for what your product looks like. Without that split, the model averages the two products and you get a bottle that is neither. **Fidelity.** Name what must stay the same: shape, proportions, packaging, colors, marks, the lettering and layout of the label, the angle, the state. Then say "add nothing the photo does not show." Models complete products from memory of similar products; a cap, a badge, a second flavor appear because the model has seen them on things like yours. **Integration.** Name the physical checks: perspective, scale, contact with the surface, light direction, shadow, reflections. "Pasted" is what mismatched light and a missing contact shadow look like. Then tell it what to sacrifice when the reference fights your product: drop the prop or the pose, not the realism. **What does not come across.** List it: their product or a lookalike, logo, marks, watermark, handle, brand name, prices, reviews, claims, and people's faces. The model does not know that the palette is fine to take and the logo is not until you draw the line. **What to do when it is unsure.** If the reference product is doing something yours cannot, your product takes its place in a way it physically can: a drink being poured becomes your product standing where the glass was, not your product being poured. One product by default, even if the reference shows three. If the ad needs a person, ask for a new one who cannot be identified, and keep the product a separate object. Then the copy sheet, the format, and at most one requested change. Two prompts below, one for each job. They are shorter than you would expect, because the sheet and the photo are doing the heavy lifting. ### Prompt for a swap ```text Image 1 is a reference ad. Image 2 is my product: [product] by [brand]. Make one finished [4:5 / 1:1 / 9:16] ad by editing Image 1: replace the advertised product with my product and keep everything else worth keeping. My product: render the same real product shown in Image 2, not a lookalike. Keep its shape, proportions, packaging, colors, marks, and the lettering and layout of its label exactly as photographed, and add nothing it does not show. Show it only from the angle in Image 2, in the same state. One product. Make it real: match the reference's perspective, scale, contact with the surface, light direction, shadow, and reflections, so my product looks photographed in the scene, not placed on it. If a prop, pose, or effect from the reference would put my product in an impossible position, drop that element. Do not carry over the reference's product or anything else belonging to their brand like logo, marks, watermark, handle, brand name, prices, reviews, claims, or people's faces. Text: change the ad's words exactly as listed and write nothing else. Put each new line where the original sits, in the same size and style. Where a line says (remove), leave that area clean. [copy sheet] Requested change: [one thing, or none]. ``` ### Prompt for a recreate ```text Image 1 is a reference ad. Image 2 is my product: [product] by [brand]. Make one finished [4:5 / 1:1 / 9:16] ad for my product that recreates the central idea of Image 1 in a form my product can actually be in. Keep the reference's composition, camera and crop, lighting, palette, mood, and the size and placement of its type. Do not copy its objects literally. Adapt the action: whatever the reference product is doing becomes something my product can physically do in the same spot and at the same scale. If it cannot do anything like it, let it rest or float naturally where the reference product was. One product. My product: render the same real product shown in Image 2, not a lookalike. Keep its shape, proportions, packaging, colors, marks, and the lettering and layout of its label exactly as photographed, and add nothing it does not show. Show it only from the angle in Image 2, in the same state. Make it real: match perspective, scale, contact, light direction, shadow, and reflections, so my product looks photographed in the scene. If the ad needs a person, create a new one who cannot be identified, and keep my product a separate object. Do not carry over the reference's product or anything else belonging to their brand like logo, marks, watermark, handle, brand name, prices, reviews, claims, or people's faces. Text: change the ad's words exactly as listed and write nothing else. Put each new line where the original sits, in the same size and style. Where a line says (remove), leave that area clean. [copy sheet] Requested change: [one thing, or none]. ``` If the reference has no copy of its own, replace the text section with one line: "The reference has no ad copy. Add no words." ## Then change one thing at a time Generate two or three takes and pick the best. From there, do not rewrite the brief. Stay in the same chat and ask for one change in a plain sentence: - "Same ad, but move the headline to the top." - "Same ad, warmer light." - "Same ad, change the background to cream." - "Same ad, make the product a little larger in the frame." One change per run. If you want two, that is two runs. The reason is not tidiness; it is that when you change three things and the ad gets worse, you cannot tell which one did it. If the text keeps coming back garbled after a couple of tries, take it out of the model's hands: ask it to "leave the text areas clean and write no words," then add the words in a design tool, or put them in the ad's text fields instead of on the image. That also makes testing different headlines cheaper. ## What goes wrong, and how to fix it Most problems are one detail, and a detail is fixed with a short follow-up in the same chat: say what is wrong, where it is, and what it should be. Start a new chat with the brief only when the whole ad has gone wrong. **Text on your product changed or turned to nonsense.** For a line you can read, name it and quote it: "The label should say '[the exact words on your label]', exactly as in Image 2." Fine print, like an ingredient panel, is harder, because these models draw very small text poorly and more wording rarely helps. Start from a sharper, closer product photo and the highest resolution the tool offers. If the fine print still matters at the size people will see the ad, place your real product photo over the product in a design tool. **Their words, logo, or marks are still there.** Point at what survived and say what happens to it: - "Remove the logo in the [top left corner] and continue the background behind it." - "Replace '[their line]' with '[your line]', in the same lettering." Before the next run, crop what you can out of the reference and check that every line of their copy is on your sheet. **Your product changed state.** Lid off, lit, bent, because the reference product was. Say the state you sell it in: "Keep my product closed and upright, exactly as in Image 2." If that state is the whole idea of the reference, the job is a recreate, not a swap. **It looks placed on the scene, not in it.** Something really photographed in a place shares that place's light and touches its surfaces. A pasted product gives itself away in one of four ways, and each has a one-line follow-up: - No shadow where it meets the surface: "Add a soft contact shadow where my product meets the surface." - Light from a different direction than the rest of the scene: "Light my product from the same direction as the rest of the image." - A different color of light, like a cool studio shot dropped into warm sunlight: "Match my product's colors and highlights to the scene's light." - Edges crisper than everything around them: "Blend my product's edges into the image's focus and grain." If two follow-ups have not fixed it, try a reference whose light is closer to your product photo's. **You got more than one product.** Models tend to repeat what they see. A reference with several people or products in it, or a product photo that also shows the box, can come back with two or three copies of your product, or with a box you did not want. Fix it in one line: "Show my product once, and remove the box." Next time, crop your product photo to the single item you want in the ad. **The model redesigned the ad instead of adapting it.** You asked for their ad with your product and got a clean, generic one: a new layout, a stock background, different lettering. Pull it back: "Go back to Image 1's layout, background, and lettering, and change only the product and the words on my copy sheet." If it keeps drifting, start a new chat with the brief. Every follow-up edits the last image, so a wrong turn carries into the next one. **Your words do not fit.** Your line is longer than theirs, so the lettering shrinks, breaks in an odd place, or runs into the product. Shorten it toward the original's length, or say how it should sit: "Set '[your line]' on two lines, in the same lettering, clear of my product." ## What crosses the line Most of this is common sense, and you have more room than it might seem. A layout, a lighting setup, a palette, and the energy of a type treatment are ideas, and [copyright does not protect ideas](https://www.copyright.gov/help/faq/faq-protect.html). The limits come from the things that make an ad theirs. This is general information about US law, not legal advice. **The obvious lines:** - **Their logo, name, or product.** Anything that identifies their brand stays out, including a product or package that looks like theirs. Trademark law turns on confusion: if a customer could think the ad is theirs, or that they approved it, it has crossed the line. - **Real people.** A recognizable person has to have agreed to appear in your ad. Using someone's face to sell something without permission is what right-of-publicity laws cover. - **Their slogan and their proof.** A slogan can be a trademark. Their reviews, ratings, and results describe their product, not yours, and presenting them as yours is deceptive advertising. **The less obvious lines:** - **A product look that is theirs without the logo.** Some brands are recognizable from the product itself: a signature color on a signature package, a distinctive bottle shape. That look can be protected as trade dress, so make sure your product in the ad still looks like yours, not like theirs. - **Lookalikes.** A person who closely resembles a celebrity can suggest an endorsement that does not exist, which is its own legal problem. A new person who cannot be identified avoids it. - **Owning what you make.** The US Copyright Office's position is that [AI output is protected only where a person determined enough of its expressive elements](https://www.copyright.gov/newsnet/2025/1060.html), and writing prompts alone does not count. A competitor may be free to reuse parts of an AI-generated ad. Your own product photo is still yours. In practice, the ad platform is where this shows up first. Meta's ad standards [do not allow ads that infringe someone's copyright or trademark](https://transparency.meta.com/policies/ad-standards/intellectual-property-infringement/third-party-infringement/), and an ad can be rejected or removed after the rights holder reports it. The simplest way to stay clear: if a customer could mistake your ad for theirs, change the palette and where your product sits until they could not. Ad Library shows you what is running; it does not give you the rights to any of it. ## Check before it runs - The product is yours: label lettering, shape, colors, proportions, state, and a shadow that grounds it. - Nothing of theirs survived: no product lookalike, face, logo, handle, or fragment of their words. - Every word in the ad is on your copy sheet and true for your product. - It is the format you will run, with the text clear of where the interface sits. - At a squint it looks like your ad in their style, not their ad with your product in it. ## Going deeper The judgment behind image direction, including how to read a reference, brief a scene, and change one variable at a time, is what [Module 3: AI Creative](/modules/ai-creative) covers. For more on AI ad creative, see the [AI ad creative topic](/topics/ai-ad-creative). ## Frequently asked questions **Which image model should I use?** GPT Image 2.5 in ChatGPT or Nano Banana Pro in Gemini. Both edit the images you give them, follow long instructions, and are built to keep a product's details. Use the one you already have; if it keeps changing your label, run the same brief in the other. **Why not let the model write the copy?** Because it is already doing the image, and copy is a second job with its own rules: which words are the brand's, which claims are true for you, what fits the layout. Given both at once, it invents a headline or keeps a line that was theirs. The copy sheet takes five minutes and removes the most common failure. **Is it legal to recreate someone else's ad with AI?** Borrowing an ad's approach is generally fine: its layout, lighting, palette, and type energy are ideas, and copyright does not protect ideas. What crosses the line is anything that makes the ad theirs: their logo or name, a product that looks like theirs, real people, their slogan or reviews, or a result customers would mistake for their ad. This is general information about US law, not legal advice. **Can I use an ad from the Ad Library as a reference?** Yes, for its structure, lighting, palette, and type. Not for its product, people, logo, or words, and not so closely that a customer would mistake your ad for theirs. **How many outputs should I make?** Two or three from the brief, then one-change variants from the best. Twenty at once tells you nothing, because you cannot see which change did what. ## Go deeper on image direction Module 3 covers briefing a scene, using references well, and iterating one variable at a time. The AI ad creative topic collects the related guides. - [Open Module 3: AI Creative](/modules/ai-creative) - [Browse AI ad creative](/topics/ai-ad-creative) Module 1 is free · No card required.