Listing descriptions were the first real estate task AI took over, and it is easy to see why. They are short, formulaic, and nobody enjoys writing them. The problem is that the tool is now common enough that buyers scrolling a search results page are reading five descriptions in a row that all sound like the same person wrote them. "Nestled in a sought-after neighborhood." "Boasts an open-concept layout." "A rare opportunity you won't want to miss."

When every listing sounds the same, the copy stops doing any work at all.

The problem is your input, not the model

Ask an AI to "write a listing description for a 3 bed 2 bath in Wash Park" and you will get generic copy, because that prompt contains no information a thousand other agents don't also have. The model fills the gap with the average of every listing it has ever seen. That average is exactly what you are trying to escape.

Feed it the three things only you know

Before you write the prompt, spend four minutes gathering: the specific detail you noticed at the walkthrough that is not in the MLS fields — the west-facing kitchen window, the fact the basement stays cool in August, the neighbor who runs the block party. Second, who the realistic buyer is, in one sentence: "a couple moving from a downtown condo who want a yard but not a commute." Third, the one objection you expect and how the house answers it: "small third bedroom, but the finished basement adds a real office."

Now the prompt has material. "Write a 120-word listing description for this property. Buyer profile: [X]. Lead with [the specific detail]. Address [the objection] honestly rather than hiding it. Do not use the words nestled, boasts, stunning, or opportunity."

Ban the vocabulary explicitly

That last instruction does more than anything else on the list. Models drift toward the phrases they have seen most, and real estate copy is one of the most repetitive corpora on the internet. Keep a running banned-words list in a note and paste it into every prompt. Add to it every time you catch a tic.

The honest caveat

AI will confidently write things about the property that are not true — square footage it inferred, a school district it guessed, a feature it assumed from a photo. Fair housing language is another live risk; a model describing an "ideal neighborhood for young families" has just written a compliance problem. Every description still needs your read before it posts, and that read is not optional.

The takeaway

Spend four minutes on input to save fifteen on output. The agents whose AI-written copy still sounds like them are not using a better tool — they are the ones who told it something it could not have known.