Try this with your last month of ads. Pull every one of them up, the statics, the video thumbnails, the carousel cards, the ones your nephew made in an app on a Sunday, and lay them out side by side on one screen. Then cover the logos with your hand.
For a lot of small businesses running AI-made creative right now, what comes back is twenty ads that each looked fine when they were approved and look like twenty different companies when they sit together. One is warm and grainy. One is glossy and blue. One has a bold condensed headline, the next a thin script, the next something that looks like a greeting card. The product is photographed from a slightly different angle in a slightly different light in every frame, and in two of them it has picked up a lid it doesn't have.
None of those ads is bad on its own. That's the problem. Nobody approves the batch. They approve one ad at a time, and one ad at a time, everything passes.
AI made ad volume nearly free, and that is genuinely good news. It also made it trivially easy to publish a stream of ads that share nothing but a logo in the corner, which quietly throws away the thing volume was supposed to build. This article is about keeping the look while the machine makes the volume: why AI drifts, what should change from ad to ad and what never should, and the small, boring system that holds it all together.
Every Generation Starts From Zero
A designer who has made your last forty ads carries your brand around in their head. They remember the shade of green that looks right and the one that looks cheap. They know the product is always shot from slightly above. They don't have to be told.
An image model has no such memory. Every request starts from a blank page, and when you don't tell it something, it fills the gap with whatever is most typical for that kind of image. We've written before about how that pull toward the typical flattens your ideas into everyone else's. The same force works on your look. Leave the lighting unspecified and you get the model's default lighting. Leave the type unspecified and you get whatever type it associates with "sale ad." Each image resolves toward the middle of the internet, and a slightly different middle every time.
Then add the ways a small business actually produces ads now:
- More than one tool. One batch from a chatbot, one from a design app, one from a video generator, each with its own house style baked in.
- More than one person. You, a freelancer, the office manager, a family member with a good eye. Each one prompts differently.
- Words doing a job words are bad at. "Clean and modern" means something different for every person and every model. Jamie Gannon, an AI creative director, put it plainly on the How I AI podcast: a picture is the better language, because "a picture to an LLM is worth a thousand words."
- The platform editing the result. Even after you approve an ad, Meta may crop, expand, animate or rewrite it. We covered how much Meta changes after you hit publish in its own piece.
Even the tools admit it. In Canva's own tutorial on keeping a brand consistent with AI, the presenter generates a set of social posts and says, before fixing anything, that "the AI-generated designs have different layouts and styles." That is the default. Consistency is not something the tools do for you. It's something you set up and hold.
Meta Asked for Variety. It Did Not Ask for Strangers.
Here is where owners get genuinely confused, because the advice seems to point two ways at once.
Meta tells advertisers to diversify their creative. In its own guidance, diversification means different concepts built around different reasons people buy: one customer wants the jacket because it's stylish, another because it's built for the mountain, so you make ads for both. It also points you to its generative AI tools to produce more assets at scale. That advice is right. We've made the case ourselves that you find winners by testing genuinely different ideas, not by guessing.
Meanwhile, the case for consistency is just as right. Recognition compounds. The tenth time someone sees an ad that looks like you, it feels familiar, and familiar is halfway to trusted. We made that argument in full in Looking Expensive Has Nothing to Do With Your Budget, and it still holds.
The two only collide if you treat "the ad" as one thing that either varies or doesn't. It isn't. Every ad has two layers, and they obey different rules.
The idea layer should vary. The angle, the customer motivation, the hook, the person or scene in the frame, the format, the offer. This is what Meta means by diversity, and this is where testing lives.
The identity layer should hold. Your colors, your typefaces, the way your images are lit and color-graded, how the product looks, the framing you return to, the voice in the copy. This is what a customer recognizes before they read a word.
The failure with AI is almost never too much variety in ideas. It's variety leaking into the identity layer, because nobody told the machine which layer was which.
Show the Machine Your Brand. Do Not Describe It.
The fix is not a longer prompt. It's a reference pack: a small, curated set of files that goes into every single request, from every person, in every tool. Think of it as the designer's memory, written down, so the machine can borrow it.
A useful pack for a small business has four parts.
1. Ten to twenty images that are unmistakably you. Not your favorite photos. Your most representative ones: past ads that performed and looked right, real photos of your work, your storefront, your team, shot in the light you want everything to share. These are what you upload as style references. A model matches a picture far better than it follows an adjective.
2. Your actual product, from several angles. If what you sell is a physical thing, give the model real photos of it from the front, the side, above and in hand, under more than one kind of light. Curious Refuge, which teaches AI advertising production, is blunt about why: unless a product is fairly generic, getting the same label, text and logo to come back correctly "can be very challenging." Several honest angles is how you narrow the gap. (Where AI still gets your product wrong, and why that matters so much more on a product page, is its own subject, and we've covered where that quietly costs you.)
3. One page of rules, written as dos and don'ts. Your two or three brand colors as exact codes. Your one or two typefaces. How images are lit ("warm window light, never flash"). How the product is framed ("always slightly from above, never floating"). What never appears: stock-photo handshakes, a gradient background, a font that isn't yours. Canva's tutorial builds exactly this into its brand kits, with dos and don'ts sitting next to the assets, so anyone who opens them stays on brand. You don't need Canva to do it. You need one page.
4. The saved setup. Which tool, which reference images, which standing instructions. Written down in one place. Mike Futia of SCALE AI showed how far this goes when he briefed an image model on a three-piece campaign: an ad, an email header and a packaging concept, with an explicit instruction that all three share one campaign identity, the same color, headline and angle, so that anyone who saw them laid out would "immediately recognize them as one campaign." It worked because the consistency was a written requirement of the request, not a hope.
Gannon described the payoff of doing this properly. For one brand image set she produced, she estimated that 90 to 100 percent of the images came from the same setup: the same references, the same saved style profile, the same final prompt. That's the whole trick. Consistency doesn't come from being careful on each ad. It comes from making every ad start from the same place.
If you already hand briefs to a freelancer or an agency, the pack sits alongside the brief, not inside it. The brief says what this ad is for. We've written a whole piece on what a brief needs to carry. The pack says what every ad looks like. One changes every time. The other almost never does.
One Loud Reference Will Take Over the Whole Batch
Here's the part nobody warns you about. A reference pack is not a pile. Every image in it pulls on every image the model makes, and the loudest one pulls hardest.
Gannon gave a perfect example. She built a set of style references for a brand look, and the results kept coming back too green. The cause was a single reference photo: a close-up where half the eye was painted bright green eyeshadow. It was the most striking thing in any of her references, so the model treated it as the most important. Every generation inherited it. She didn't fight it with more instructions. She pulled that one image out, and the whole batch came back into line.
Small businesses do this constantly without noticing. You include the one photo with the dramatic sunset, and suddenly every ad glows orange. You include the holiday ad with the red background, and red creeps into July. You include a photo where a customer is wearing a bright yellow jacket, and yellow starts showing up in places it has no business being.
So curate the pack like an editor:
- Look for anything that shouts. A color, a prop or a lighting condition that appears once but dominates the frame will dominate the output. If it isn't part of your identity, take it out, even if it's a great photo.
- Keep the references similar to each other. Gannon's observation was that the more consistent the reference set, the more consistent the results. Ten images that agree beat thirty that argue.
- Rebuild the pack from your winners. Once a batch comes out right, the best images from it become the new references. After a few rounds, the pack is made of images that are already on-brand and already proven to work with the machine. It gets tighter every time you use it.
Judge the Batch, Not the Ad
Everything so far happens before you generate. This part happens after, and it's the part most owners skip, because their approval process was designed for a world where ads arrived one at a time.
When you're reviewing AI output, stop looking at ads individually. Put the whole batch, plus your last few live ads, into one grid on one screen. Then run the logo test from the top of this article at batch scale. Cover the logos. Does it read as one business? Which ones break the pattern?
This takes real honesty, and Gannon named the habit it requires. Faced with images that were individually striking, her question was whether they actually matched the look she was building, and her answer to herself was blunt: "truthfully, it does not." A one-off image can be beautiful and still wrong for the set.
Be willing to cut the ad you like. That's the hardest discipline in the whole process. An AI tool will occasionally hand you a genuinely great image in the wrong look, and the temptation is to run it anyway because it's great. Run it and you've taught every customer who sees it that your brand looks like that too. If the idea is strong, keep the idea and regenerate it inside your identity layer. The idea is the valuable part. The borrowed look is not.
A few quick checks that catch most drift:
- Find the odd one out. Give yourself two seconds to point at the tile that doesn't belong. If you can find one that fast, so can your customer, and it comes out or gets regenerated.
- Check the product in every tile. Same shape, same label, same lid, same color. Any tile where it changed is not a style issue. It's a truth issue, and it comes out.
- Read one headline from each ad aloud in a row. If they sound like different companies wrote them, they did, as far as your customer is concerned.
Your Voice Drifts Too
Everything above is about pictures, but the same drift happens in the words, and it's easier to miss because bland copy doesn't look broken.
Ask an AI tool to write an ad and you get its default voice: fluent, reasonable, a little too polished, full of phrases like "elevate your experience" and "unlock the difference." Anthony Camacho, who writes about ad copy for online stores, describes the pattern well: AI copy tends to tell instead of show, so it comes out logical and surface-level when it should make someone feel something. Across twenty ads, that default voice slowly replaces yours. Your customers stop hearing the business they know.
Give the voice its own reference pack. It's smaller than the visual one:
- Five of your best real pieces of writing. Past ads that worked, the email customers actually replied to, the way you describe the business when someone asks at a barbecue.
- Your customers' own words. Phrases lifted straight from reviews and messages. They're the most on-brand copy you have, and you didn't write them.
- A short banned list. The words that are never you. Most businesses can name ten in a minute once they start.
- Two or three rules of voice. "Short sentences." "We say 'you,' not 'customers.'" "No exclamation points."
Paste it into every copy request, just like the visual pack goes into every image request. The goal isn't to stop AI from writing your drafts. It's to make sure the drafts sound like somebody who works there.
Whoever Holds the Reference Pack Holds the Brand
One last point, and it matters more the more people touch your ads.
Traditionally, when a small business paid someone for creative, it got back finished files. The knowledge of how to make more of them stayed with the vendor. If they left, or you switched, the look walked out the door with them, and the next person started from zero, which, with AI in the loop, now means starting from the model's average.
Gannon described a different model that's worth copying no matter who makes your ads. She does the upfront work of defining the look, then hands the client the setup itself: the reference images, the style profile, the final prompt, packaged so the client can keep generating on their own. The deliverable isn't just the images. It's the recipe.
Make that your standard, whoever you work with:
- The pack belongs to you. Keep it somewhere you control, a shared folder with your name on it, not inside a freelancer's account.
- Ask every vendor for the setup, not just the files. Which references, which tool, which standing instructions produced this.
- Everyone uses the same pack. Your in-house person, your agency and your nephew all start from the same references. That single rule removes most of the drift before it happens.
- Review it every quarter. Swap in the newest winners, remove anything that started pulling the look somewhere you didn't want, and write down what changed.
Recognition Is the Asset
It's easy to measure what AI saved you on production, because the number is right there on the invoice you didn't have to pay. It's much harder to see what you lose when every ad looks like a different company, because the cost doesn't show up anywhere. It shows up as a customer who has seen you nine times and still doesn't recognize you on the tenth.
The volume is a gift. Take it. Test more ideas than you ever could before. Just decide, once and on purpose, which part of every ad is allowed to change and which part never does, and then make the machine start from your brand every time instead of from its own average.
That split between the idea that changes and the identity that holds is how we run creative production for our clients: AI does the volume, and a person owns the look. If you'd rather have someone build your reference pack and keep every batch on-brand, that's the work we do. If you'd rather build it yourself, everything above is the whole method.
