Real Estate Marketing Graphics: The AI Prompt Playbook
An agent I have designed for since 2019 sent me a message in March that I think about a lot. She said: my photos are fine, my listings are fine, and my marketing looks exactly like the eleven other agents in my zip code. She was right. I pulled up her feed and three competitors and honestly could not tell them apart. Same script font. Same beige. Same house photo with a translucent white bar across it.
Real estate marketing AI prompts fixed that for her faster than I expected. But not in the way most people assume, and there is a line here you really do not want to cross. So let me start with that part.
The one thing you should never point AI at
Do not use AI to change the property.
I am going to be blunt because this stopped being a taste question and became a legal one. California's AB 723 took effect on January 1, 2026. If a broker or salesperson uses a digitally altered image in an ad or promotional piece, the listing needs a disclosure statement placed on or right next to that image, plus a link, URL, or QR code pointing to the original unaltered photo. Removing a neighbor's roofline. Dropping in a green lawn. Virtually staging an empty room. All of it counts.
Brightness, contrast, color correction, cropping, lens distortion fixes: those are generally fine, because they do not change what the house actually is. And rules vary by state and by MLS, so check yours before you touch anything.
So what is left? Everything around the listing. The flyer. The open house sign. The market update graphic. The just listed announcement. The neighborhood guide. The illustrated header on your monthly email. None of that is a photo of the house, so none of it carries that risk, and all of it is where agents look identical to each other anyway.
That is the whole play. Leave the property alone. Redesign everything else.
What AI is actually good at here
Text on graphics used to be the punchline. It is not anymore. ChatGPT's image models render short headline text reliably now, and the newer image model pushes character-level accuracy to near perfect on typical Latin type. I still proof every word before it goes anywhere near a printer, because near perfect is not perfect and a misspelled street name is a bad day. But the era of asking for a flyer and getting alphabet soup is over.
Where it saves the most time is the first draft. My old flyer process was about 45 minutes: pull a layout, set the type, adjust the hierarchy, export, send, get notes. Now I generate three layout directions in under 10 minutes, pick one, and spend the rest of the time on the part that actually matters, which is making it hers and not generic.
Here is the open house flyer prompt I use as a starting point. Notice how much of it is layout instruction rather than mood words. That is deliberate.
Swap the bracketed parts and you have a flyer system rather than a one-off. Craftsman, mid-century, new build colonial, downtown loft: the illustration changes, the skeleton does not.
Why your prompts keep returning generic work
Most people are prompting this wrong, and it is almost always the same mistake. They describe a feeling instead of a layout.
Watch what happens with these two.
Weak: "Make a modern, professional real estate flyer for a beautiful home, clean and elegant."
Strong: "Create a 1080 x 1080 announcement graphic. Deep navy field, cream type, one accent at #C9A227. Eyebrow text top left, oversized two-word headline in the middle third, one supporting sentence at 40% of the headline size below it. Bottom 120 pixels left empty. Flat vector, no photographs, no house icons."
The first one gives the model nothing to decide with, so it returns the statistical average of every real estate flyer ever made. Which is exactly the beige thing your competitor posted yesterday. The second constrains size, palette, hierarchy, type scale, and negative space, so there is only a narrow band of outputs it can land in.
Ask yourself this before you hit enter: could a junior designer build this from my description without asking me a single question? If no, the prompt is not done.
The market update graphic that actually gets saved
Monthly market updates are the most-posted and least-read graphic in the entire industry. Nine stats crammed into a grid, six-point type, nobody stops.
One number. That is the fix. Pick the single stat that would make a homeowner in that zip code sit up, make it enormous, and let one sentence explain it.
That last line matters more than it looks. Always reserve a clear zone for your logo instead of asking the model to invent one. Generated logos are where these graphics fall apart, and a real mark placed on top reads as yours instantly. If you do not have one yet, the logo and branding prompts are a decent place to work one out, and the signature logo prompt is the one agents pick most often, since a clean handwritten name mark suits personal real estate branding better than a house-shaped icon does.
Making six graphics look like one brand
This is where agents usually give up. One graphic looks great. The next one comes back with different margins, different type weight, different everything, and now the feed looks like four people run it.
The fix is to stop asking for graphics and start asking for a system.
Define the rules once, in writing, then generate against them one template at a time. Same palette, same type pairing, same margins, same clear zone. When the model has the constraints written out in front of it, template four still matches template one.
Then you drop your real listing photo into that empty photo well in whatever tool you already use. The photo stays untouched, which is the point. The design around it does the work.
Most of the prompts I write for client campaigns end up in the marketing and ads collection, since a just listed graphic and a small business promo graphic are structurally the same problem wearing different clothes.
Frequently Asked Questions
Can I use AI to stage or clean up my actual listing photos?
Not without disclosing it. In California, AB 723 has required a disclosure statement next to any digitally altered listing image plus a link to the unaltered original since January 1, 2026, and other markets and MLS systems have their own rules. Basic color and crop corrections are usually fine. Anything that changes what the property looks like is not something to do quietly.
Does ChatGPT render text on graphics accurately now?
Yes, far better than it did two years ago. Short headlines, dates and prices come out clean most of the time. Proof every character anyway before printing, especially street names and phone numbers.
What size should I ask for?
Always state exact pixel dimensions or print size, plus the aspect ratio. "1080 x 1080" and "4:5 vertical" and "8.5 x 11 print" all steer the layout. Leaving it out is how you end up with a composition you cannot crop into anything usable.
Do I still need a design tool?
For the final 10%, yes. Logo placement, exact brand fonts and the real listing photo all go on top afterwards. AI gets you a strong layout fast. You finish it.
Will this work for commercial listings too?
It works better, actually. Commercial marketing leans on data, floor area figures and location context rather than lifestyle photography, and those are exactly the kinds of graphics that respond well to a tight layout brief.
Where to start this week
Pick your next listing and build one thing: the open house flyer. Use the prompt above, swap in the house style and your accent color, and give yourself 20 minutes. Do not try to rebuild your whole brand on a Tuesday.
Then look at it next to the last flyer you sent out. If it does not look meaningfully more like you and less like everyone else in your zip code, the prompt needed more constraints, not more adjectives.
If you want a head start, everything in the shop came out of real client work, including the ones I originally wrote for agents. They are structured exactly like the examples here: layout first, mood second.
