The listing description that ChatGPT writes on the first prompt is fine. It is also legally risky, generically pleasant, and identifiable as AI-written within about ten seconds by any agent in your market who has used the same tool. The agents producing genuinely strong AI-assisted listings are not running a one-shot prompt. They are running a four-stage chain that captures the sensory detail, drafts the copy, audits it for fair-housing exposure, and adapts it for each portal.
The NAR 2025 Technology Survey found 46% of agents using AI-generated content. Most of that content reads identically across competitors. The chain below is what separates the listings that win the click from the ones that get scrolled past.
Capture the sensory detail before you sit at the laptop
The single biggest reason AI-generated listings sound generic is that the input was generic. Agents type “four bedroom, two bathroom, open plan, deck, double garage” and wonder why the output reads like every other listing on the portal. The output reads that way because the input did.
Before you write anything, walk the property with your phone in voice memo mode and narrate what you see. Forty-five seconds is enough. Talk about the way the light falls in the kitchen at 4pm, the school you can hear from the back deck, the quirky built-in bench in the laundry that the family clearly uses every morning, the mature tree that frames the front gate. Then drop the voice memo into ChatGPT or Claude and ask for a transcription with this:
“Transcribe the following voice memo from a property inspection. Then organise the content into three buckets: factual features (hard data: room counts, materials, layout), sensory and emotional details (light, sound, the feel of the space, signs of how the family uses it), and selling angles (the things you would mention to a buyer that aren’t in the spec sheet). Do not write a listing yet. List each bucket separately.”
You now have something to write from. The model will produce dramatically better copy from this than from a feature checklist, because human-written sensory detail is exactly the part it cannot invent reliably from numbers.
The draft prompt that produces something usable
With the bucketed notes in hand, the listing prompt becomes specific. The phrase “warm and engaging tone” is doing no work in the prompt that everyone uses. Replace it with concrete instructions about structure and constraints.
“Draft a listing description for [ADDRESS], a [BEDROOMS]-bedroom [BATHROOMS]-bathroom home in [SUBURB/CITY]. Use the bucketed notes below as your only source material. Structure: a one-line headline that captures the single most distinctive thing about this home, a 90-word opening paragraph that puts the reader inside the home, a 70-word middle paragraph on the layout and standout rooms, a 60-word closing paragraph on the lifestyle and location, then a bullet list of factual features (no more than 8 bullets, ordered by what a buyer would care about most). Tone: confident, specific, no real-estate clichés (no ‘breathtaking’, ‘must-see’, ‘won’t last’). If you find yourself writing a generic adjective, replace it with a specific detail from the notes instead. Notes: [PASTE BUCKETED NOTES].”
The instruction to replace generic adjectives with specific details is the thing that lifts the output above the generic baseline. The model will sometimes still slip in “stunning” or “spacious.” Strike them in the edit, the rest is usable.
The audit prompt that nobody is running yet
Here is the failure mode most agents do not realise they are creating. AI listing tools, including the major ones, will cheerfully generate phrases that violate fair housing law in the United States and equivalent legislation in other markets. “Family-friendly neighbourhood,” “perfect for a young couple,” “great schools nearby” (depending on jurisdiction), and “walkable for active retirees” are all phrases an AI will produce by default and a regulator will treat as discriminatory.
The audit prompt to run on every single AI-generated listing before it goes anywhere near a portal:
“Review the listing copy below for fair housing compliance and risk language. Flag every phrase that: (1) implies a preference for or against a protected class (familial status, age, race, religion, national origin, disability, sex), (2) describes the neighbourhood in a way that could be read as steering, (3) makes a factual claim that would need to be substantiated (school ratings, crime statistics, future appreciation, square footage if not measured), or (4) uses superlatives that could be challenged (‘largest’, ‘best’, ‘only’). For each flagged phrase, suggest a compliant rewrite that keeps the marketing intent. Return the result as a table with columns: original phrase, risk type, suggested rewrite. Listing copy: [PASTE].”
Run this once and you will see how many phrases the first-draft model produced that you would have published. Most agents who run this prompt for the first time find at least three flags in a 250-word listing. The audit takes about ninety seconds. The exposure it removes is significant.
Adapting for each portal
Different portals have different conventions. Realtor.com and Zillow tolerate longer descriptions and reward keyword density. Trade Me in New Zealand and realestate.com.au reward a strong opening line because more of the description is hidden behind a click. Facebook and Instagram want emotion and a hook. The same listing copy is wrong for all of them.
“I have one master listing description below. Adapt it into four versions: a 220-word version for the main listing portal that leads with the strongest factual draw, a 90-word Facebook version that opens with a question and ends with a soft CTA, a 60-word Instagram caption that opens with a hook and ends with ‘open home Saturday’ or ‘message for details’, and a 45-word text message version for past clients in the area. All four versions must keep the same factual claims, the same compliance posture (no flagged phrases), and recognisably the same voice. Master copy: [PASTE].”
You now have four versions in under a minute. Verify the factual details one more time before publishing. The model occasionally rounds a square-footage figure or shifts a feature between versions, and the verification step is what catches it.
A counterintuitive observation about what buyers actually read
The agents who track listing analytics on their own websites tend to find the same pattern: buyers read the headline, the first sentence of the opening paragraph, and the bullet list. The middle prose paragraphs get skimmed. This means the time you spend perfecting the middle is time that mostly nobody sees, while the headline and the bullets carry most of the weight. AI is genuinely good at headlines and bullets, and that is where you should spend the most effort iterating. Run the headline prompt three times, pick the strongest, and use the time you save to walk one more property.
The hours of writing time AI saves are real. The hour you used to spend on the middle paragraph is now an hour you can spend on the listing photos, the open home script, or the seller call. That reallocation is the actual edge.