A listing written by a machine, and why it shows
Listing copy is the first task agencies hand to a machine. It is also where the result gives away missing material fastest.
Listing copy is the first task agencies hand to a machine. It is also where the result gives away missing material fastest.
It almost always starts here. A negotiator tries a consumer tool on one listing, finds the result acceptable, and the agency starts using it. Then, a few weeks in, someone notices that every listing reads the same, and that some contain things that are vaguely untrue.
The usual reaction is to blame the tool's quality. That is almost always wrong. Poor listing copy comes from thin input, not from a weak model.
What people typically supply: floor area, number of rooms, floor, neighbourhood, price. With that, no writer, human or otherwise, can produce anything but a generic listing. A negotiator writing from the same five lines, without having been to the property, produces exactly the same flat text.
What makes a listing good are the things only someone who has visited knows. Which way the living room actually faces in late afternoon. That the building re-roofed last year. That the school is three minutes on foot, not three minutes by car. That the vendor is relocating and is not in a hurry on price, which never appears in the listing but changes what you lead with.
The interesting part is that this information is rarely missing. It exists, scattered. In the negotiator's viewing notes, in the valuation report, in a message thread with the vendor, in the photographs themselves. It simply is not where you look when you sit down to write.
An automation that reads those sources produces text of a different nature from a tool you paste five fields into. That is not a difference of model, it is a difference of access to material.
A second gap between the individual experiment and a real installation. An agency has a way of writing, often unconscious but very stable. Some write in short, factual present tense. Others work the opening line and build atmosphere. Some never use an exclamation mark, others use them freely.
A generic tool has no reason to know that voice. It produces an average of the French property web, recognisable from the first paragraph. Calibrating on the agency's existing listings, before anything goes live, costs half a day and settles the problem once.
That leaves factual error, which is real and which matters. A wrong floor area or a wrong energy rating in a listing exposes the agency.
Which is why the loop always ends with a person. The negotiator receives a finished listing, reads it, corrects, approves. They save the writing and the keying, they keep responsibility for what goes out in the agency's name. That review takes two minutes against twenty to write, and it is the only thing that makes the whole arrangement tenable.
When an agency tells us it tried AI on its listings and was not convinced, we always ask what it fed in. The answer is almost invariably: the software fields. That was the problem, and changing tools does not fix it.