How Estate Agents Use AI to Write Property Descriptions

Estate agents use AI to write property descriptions by giving it a verified fact sheet (rooms, measurements, tenure, condition and features seen at the valuation) plus the agency's house style, then asking for a portal description, key-feature bullets and a social caption. Every claim is checked against the sheet before the vendor approves. Drafting drops to minutes per listing.

The fact sheet is the whole method. Without one, AI invents flattering detail: a "south-facing garden" from a sunny photo, "recently renovated" from a new kitchen, "within walking distance of" amenities it has guessed at. In property, a wrong description is more than embarrassing, because advertising and consumer-protection rules in most markets treat misleading listings seriously, and a buyer who relied on a false claim has a genuine grievance against the agency.

Follow me on Instagram@sagnikteaches

Build the fact sheet at the valuation, not at the desk

The best time to capture facts is while you are standing in the property with the vendor. A short template on a phone or tablet takes five minutes more at the valuation and saves a callback later. Mark every item with where it came from, because "vendor says" and "seen" deserve different wording.

Connect on LinkedInSagnik Bhattacharya

PROPERTY FACT SHEET
Ref:                       Type:                 Tenure:
Charges (service, association fees, other):
Rooms (name, measurements from floorplan):
Heating and hot water:     Glazing:              Energy rating:
Parking:                   Outside space (size, aspect CONFIRMED how?):
Works done (what, when, paperwork seen? Y/N):
Known issues disclosed by vendor:
Standout features (SEEN / VENDOR SAYS):
What the vendor loves about living here (their words):
Do not mention:

Filled in for an illustrative three-bedroom semi-detached house, the parts that matter most look like this:

Subscribe on YouTube@codingliquids

Tenure: Freehold (vendor says; title to confirm)
Outside space: rear garden approx 14m; aspect west (compass app, SEEN)
Works done: kitchen replaced 2023, paperwork seen Y; boiler 2019,
  service record seen Y; loft insulation "a few years ago" (VENDOR SAYS)
Standout features: open-plan kitchen-diner with bifold doors (SEEN);
  quiet road (VENDOR SAYS)
What the vendor loves: "evening sun on the patio, and the kids walk
  to school in five minutes"
Do not mention: vendor relocating for work; neighbour dispute over
  fence (resolved)

Two lines in that sheet would cause trouble if the AI met them without the labels: "quiet road" is the vendor's opinion, and the school walk is a vendor claim that also hints at who the house suits. Both need careful handling, covered below. The "Do not mention" line matters as much as the rest; AI will otherwise weave in any detail you give it.

A prompt that writes in your house style

Save this once where the whole office can reach it, in a shared Claude or ChatGPT Project or a Gemini Gem (becoming a skill from November 2026), with two or three of your agency's best past descriptions as examples, so every negotiator starts from the same instructions.

You write property descriptions for [agency]. Use ONLY the facts in the
fact sheet below. Never add features, measurements, directions, amenities
or condition claims that aren't on the sheet.
- Items marked VENDOR SAYS must be attributed ("the vendor tells us") or
  left out. Items marked SEEN can be stated plainly.
- Ignore everything under "Do not mention".
- Lead with the strongest SEEN feature, not the location.
- Plain, confident English. No exclamation marks.
- Never use: nestled, boasting, deceptively spacious, must be viewed,
  stunning, dream home, sought-after, perfect for.
- Describe the property, never the kind of person who should buy it.
Write:
1. Portal description, [150-250] words
2. Eight key-feature bullets, each under ten words
3. A two-sentence social caption
Then list any fact you were unsure how to use.
Fact sheet: [paste]

The final instruction is the useful one. Asking the model to list what it was unsure about surfaces the ambiguous items before a buyer does.

What the first draft gets wrong, and the fixes

An illustrative opening from the prompt above, using the filled-in sheet:

A beautifully presented three-bedroom semi-detached home on a quiet
road, with a south-westerly rear garden that enjoys the evening sun.
The recently renovated kitchen-diner opens through bifold doors onto
the patio, making it ideal for families...
Unsure: loft insulation date; whether to mention the school.

Four fixes before anyone else reads it:

• "South-westerly" is wrong; the sheet says west, confirmed with a compass. Models round aspects towards the flattering direction.
• "Recently renovated" overstates a 2023 kitchen replacement. Say "a kitchen fitted in 2023", which is both accurate and more convincing.
• "On a quiet road" is a vendor opinion stated as fact. Remove it, or attribute it.
• "Ideal for families" breaks the prompt's own rule: prompts make slips rarer, not impossible. Replace it with the fact behind it: three bedrooms and a garden.

The "Unsure" line did its job: it flagged the school claim, which stays out.

The key-feature bullets from the same run need a separate pass, because compressing a sentence into six words strips out the qualifiers that made it safe. The illustrative bullets that came back:

- Three double bedrooms
- South-west facing rear garden
- Recently renovated kitchen-diner
- Bifold doors to the patio
- Boiler 2019 with service history
- Close to local schools
- Quiet residential road
- No onward chain

Five of the eight need changing. The sheet never said the bedrooms were doubles, and the floorplan shows the third at 2.4 metres wide. The aspect is wrong again. The school and the quiet road were attributed to the vendor in the prose, but the attribution vanished in the bullets. "No onward chain" appears nowhere on the sheet: the model filled a slot it expects every listing to have. The corrected set is shorter and entirely defensible: three bedrooms; west-facing rear garden, approx 14m; kitchen-diner fitted 2023; bifold doors to the patio; boiler fitted 2019, service record seen. Add one line to the prompt so it doesn't recur: "Bullets follow the same rules as the prose. If a fact needs 'the vendor tells us', leave it out of the bullets."

One fact sheet, five pieces of copy

Generate them in one conversation so they stay consistent, but check each separately; the social caption is where hype creeps back in. Ask for three caption options rather than one, and pick the plainest.

The brochure's area paragraph deserves a word of its own. AI knows very little that is reliable about a particular street, and what it does "know" may be years out of date: the closed café, the bus route that changed, the park that was built over. Write the area paragraph yourself from local knowledge, or keep a short, checked paragraph per neighbourhood in your shared Project and let the AI adapt it. Negotiators know their patch far better than any model does, and buyers can tell.

The applicant email alert is the most underused output. If your CRM stores what each applicant asked for (a garden, parking, a home office), paste those criteria in with the fact sheet and ask for one sentence explaining the match: "You mentioned wanting space to work from home; the third bedroom is currently set up as a study with a fibre connection." Applicants read a matched alert far more carefully than a generic one, and the sentence is only as accurate as the fact sheet behind it.

Photos: let the AI describe them, never let it decide

Modern assistants can look at listing photos and suggest room captions, spot rooms you haven't photographed, or describe a view. That's useful as a completeness check. They are poor at judging materials, scale and condition. A realistic mistake: an assistant describes "solid oak flooring throughout the ground floor" from photos of good-quality laminate. Nobody at the agency catches it, a buyer asks about it at the second viewing, and the negotiator spends the next week rebuilding trust with a buyer who now doubts everything else in the listing.

Keep photos as supporting material, never as the source of facts. If you use AI to alter images rather than describe them, disclosure rules change; virtual staging costs and disclosure covers that separately.

The claims check before anything goes live

Every listing gets a two-minute read against the fact sheet, looking specifically for these words, each of which needs evidence on the sheet:

• Time words: new, recently, newly, modern. Replace with the year.
• Direction words: south-facing, sunny, light-filled. Only with a confirmed aspect.
• Distance words: walking distance, close to, moments from. Only with a real distance you have checked, and never implying a school place.
• Potential words: scope to extend, development potential, loft conversion potential. These imply planning outcomes nobody can promise.
• Status words: no onward chain, vacant possession. Confirm with the vendor in writing.
• Condition words: immaculate, turnkey, no work needed. Opinions that buyers' surveyors test.

Then add what the rules in your market require you to include, such as tenure, charges and known material issues, which AI will happily leave out because they don't sound appealing. For the full list of ways AI listings mislead, see AI listing mistakes that can mislead buyers; the broader method in fact-checking AI marketing copy works for brochures and social posts too.

Wording that can discriminate without meaning to

Listing copy that describes the ideal buyer or tenant, rather than the property, can breach equality and fair-housing rules in many places, and AI produces it constantly because it is common in the text it learned from. A before and after:

• Before: "Perfect for a young professional couple or a growing family, with great schools nearby and a church at the end of the road."
• After: "Three bedrooms and a 14-metre garden. The high street is a 10-minute walk."

The rewrite describes space and distance, both facts, and lets buyers decide whether the house suits them. Put "describe the property, never the buyer" in the prompt, and look for it in the check anyway.

Leaseholds, probate sales and tired houses need different instructions

The prompt above is tuned for a well-kept family house. Three kinds of listing trip it up, and each needs a line or two added for that job.

A leasehold flat. The facts buyers most need are the ones AI treats as dull and pushes to the end, or drops. On the sheet they look like this: "Tenure: leasehold, 118 years remaining (lease seen). Service charge: [amount] a year (latest statement seen). Ground rent: [amount] a year. Pets: lease permits with freeholder's consent (VENDOR SAYS)." Add "State tenure, years remaining and every charge in the first 60 words, and again in the bullets" to the prompt. Then check that the figures in the copy match the statement, not last year's figure the vendor quoted from memory.

A dated house sold by executors. AI reaches for euphemism here, and buyers see through all of it. A before and after:

• Before: "A blank canvas offering the discerning buyer a rare opportunity to create their dream home, in a much sought-after location."
• After: "A three-bedroom detached house in need of full modernisation. The kitchen and bathroom are original, there is no central heating, and the property is offered with vacant possession."

The honest version brings the right viewers, the ones with a budget for works, and saves the negotiator from showing it to buyers who leave in five minutes. Confirm "vacant possession" with the executors in writing before it appears anywhere.

A listing that has gone quiet. After eight weeks with few viewings, a fresh description with a different lead feature can help. Paste the fact sheet again rather than the old description, or any error nobody caught gets carried into the new version. An illustrative prompt and opening:

Rewrite the portal description from this fact sheet, leading with a
different SEEN feature from the current version. Keep every figure
identical. Don't mention the price or any change to it.

Output (illustrative): "Bifold doors run the width of the kitchen-diner,
opening onto a west-facing patio and a 14-metre garden that catches
the evening sun..."

"Catches the evening sun" is fine here, because west is confirmed. Check the rewrite against the banned-words list anyway: a fresh draft is where "stunning" tends to reappear.

What the vendor sees before it goes live

Vendor approval is your second line of defence, but only if you ask the right question. "Are you happy with this?" gets a yes from a vendor who has skimmed for compliments. Ask them to confirm facts instead:

Before we publish, please check the facts in the attached description:
room sizes, the year the kitchen was fitted, the garden's direction,
and anything we've said about works or condition. Reply with any
corrections, or "facts confirmed". We'll publish once you confirm.

A written "facts confirmed" is worth keeping on file with the fact sheet. It won't transfer responsibility for accuracy away from the agency, but it shows a proper process, and it regularly catches genuine errors: the extension that was actually built in 2015, not 2018, or the boiler the vendor thought was replaced but was only serviced. When a vendor asks for changes that add claims ("say it's a quiet road, everyone knows it is"), the same rule applies as with the AI: it goes in as the vendor's description or not at all.

Keeping every negotiator on the same prompt

The fastest way to lose the benefit is for each negotiator to keep a private version of the prompt. Within a month there are four house styles, and nobody knows which one dropped the banned-words list. Give the prompt an owner, usually the sales manager or whoever writes best, and keep it in one shared Project or document.

Keep a simple corrections log alongside it: the date, the listing reference, what the AI got wrong and what the fix was. After a few weeks the log shows patterns worth fixing in the prompt itself, such as a habit of rounding aspects, or of calling any extension "spacious". New staff read the log in their first week and learn the agency's standards faster than from any style guide.

A few illustrative rows from a log's first month, and the prompt change each one led to:

Week  Ref     What the AI got wrong                   Prompt change
1     S-2231  "South-facing" (sheet: south-east)      Never round an aspect
1     S-2236  Called a 2.1m box room a bedroom        Use floorplan room names only
2     S-2240  "No onward chain" (not on sheet)        Status words only from sheet
3     S-2244  "Minutes from the station" (a guess)    Distances only if on sheet

To check the method is working, pull ten live listings at the end of each month and count claims that have no support on the fact sheet. Say the first month finds 18 across ten listings, just under two each. If the count isn't falling by the third month, look at the fact sheets before the prompt: the usual cause is negotiators skipping the SEEN and VENDOR SAYS labels, which leaves the AI unable to tell a checked fact from a hopeful one.

Turnaround for a two-branch agency

For an illustrative two-branch sales agency listing about 40 properties a month, with four negotiators:

At 40 listings a month, moving from about 35 minutes to about 15 minutes per listing frees roughly 13 hours a month across the office, before counting the callbacks avoided by capturing facts properly first. Tool cost: ChatGPT Business at $20 a seat a month on annual billing (two-seat minimum) or Claude Team at the same price gives four negotiators a shared Project, business data terms and no training on your content by default. Check first whether your agency CRM already has a description writer; if it does, test whether it respects your fact sheet and banned words before paying for anything new.

Keep the prompt and examples in one shared place and review them monthly: add any cliché that slipped through, and swap in a recent description the vendors particularly liked. When your valuation process feeds the fact sheet, preparing valuation appointments with AI is the natural step before this one.

Questions agents ask about AI-written listings

Do I need to tell buyers a description was written with AI?

Generally what matters is whether the description is accurate, not who typed it, but check your trade body's guidance and any portal terms. Where AI has changed images, such as virtual staging, disclosure is a different matter and usually expected. Many agencies simply note in their terms that marketing copy is prepared with software assistance and checked by staff.

Can AI write a description from the photos alone?

It can produce something readable, but it will guess at finishes, sizes, orientation and condition, and some of those guesses will be wrong. Use photos to help it describe rooms you have already recorded on the fact sheet, not as the source of facts. If a feature appears only in a photo and not on your sheet, check it before it goes in the copy.

Does the same approach work for lettings listings?

Yes, with a different fact sheet. Lettings copy needs availability date, furnished status, deposit and holding-deposit terms, what bills are included, pet and smoking policy, and any restrictions from the landlord. Tenants skim for these facts first, so ask the AI to put them in the opening lines and keep the lifestyle description short.

Which AI tool is best for property descriptions?

Any of the main assistants, such as ChatGPT, Claude or Gemini, writes good listing copy when given a proper fact sheet and style guide; the input matters far more than the tool. Check first whether your agency CRM has a built-in description writer, and test whether it lets you control the style and limits it to your recorded facts.

Further reads

• How to Automate Buyer and Vendor Follow-Up With AI — Once the listing is live, automate the follow-up it generates.
• Can an AI Chatbot Book Valuations for Estate Agents Out of Hours? — Capture more valuations to write descriptions for, out of hours.
• How to Build a Shared Prompt Library for Your Team — Keep every negotiator on the same house-style prompt.
• How to Spot Bias and Stereotypes in AI-Written Content — A wider check for assumptions hidden in AI-written copy.
• Can ChatGPT Read PDFs, Spreadsheets and Photos? What Breaks — What AI reliably reads from floorplans and photos, and what breaks.
• How to Write an AI Disclosure Statement for Your Website — Wording for telling clients how your agency uses AI.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: vendor pricing pages for ChatGPT Business and Claude Team; general advertising-standards and fair-housing principles as applied to property marketing (check your own trade body's guidance).