Rightmove descriptions written by AI are converging on the same four adjectives

I opened Rightmove last Tuesday and searched a postcode I know well. Eight listings on the first page. Five of them used the word "stunning" in the first sentence. Three used "sought-after" in the second. Every single one ended with some variation of "ideal for families and professionals alike." The agents involved are not the same company. They are not using the same software. They have, however, all fed the same weak prompt into the same category of model, and the model has done exactly what it was trained to do: produce the most statistically average version of an estate agency listing.

This is the problem with AI-assisted listing copy right now. Not that it is bad writing. It is competent writing. It is so competent, so thoroughly averaged across ten years of Rightmove listings and property journalism, that it has become invisible. On a portal where the description is one of the few levers an agent actually controls, invisible is a serious commercial problem.

What actually changed when agents adopted AI copy tools

Before AI-assisted drafting, listing copy varied because humans vary. One negotiator wrote long, atmospheric descriptions. Another was terse and feature-led. The branch manager had a habit of opening with the street name. The quality was inconsistent, yes. But the output had fingerprints on it.

When agencies switched to AI-assisted drafting, they standardised the input (a prompt template, usually borrowed from a LinkedIn post or a supplier's onboarding guide) and pointed it at a general-purpose model. The model was never given any instruction about the agency's voice, its market position, or the specific character of the properties it typically sells. The prompt said something like: "Write a compelling Rightmove listing for a 3-bed semi in [area] with [features]. Tone: warm and professional."

Warm and professional is not a brand constraint. It is a genre instruction. Every model trained on property copy already knows what warm and professional sounds like in this context. It sounds like "stunning," "sought-after," "contemporary open-plan living," and "ideal for."

The variance did not collapse because the tool is poor. It collapsed because the brief was empty.

Why this matters more on Rightmove than anywhere else

On a portal, the agent has almost no visual differentiation. Photography matters, and floor plans matter, but beyond that, the description is the copy. It is the one place where an agent can signal that they understand this specific property and this specific buyer.

When that copy reads identically to the listing two doors down from a competing agent, the buyer's eye skips. Worse, the algorithmic behaviour of buyers on Rightmove means that listings with higher engagement metrics get surfaced more. A description that produces no pause, no read-through, no save, is quietly penalised in the feed ranking.

There is also a vendor trust issue that takes longer to surface but matters more. A vendor who reads their listing and thinks "this could be anyone's house" is not going to recommend that agent. They may not even consciously articulate why. They just feel that nothing about the marketing felt specific to them.

The brand problem in AI property listings is not new, but the Rightmove context makes it acute. You are competing in a grid. Sameness is the worst possible outcome.

What the model is actually doing

Large language models do not have opinions about your property. They have probability distributions over what words tend to follow other words in contexts that look like the one you have described. When you give a weak brief, the model defaults toward the centre of that distribution. The centre of the distribution for UK estate agency copy, trained on a decade of Rightmove and Zoopla listings, is "stunning," "sought-after," "contemporary," and "ideal for."

This is not a flaw. It is the model working correctly given the information it received. The flaw is in the workflow design upstream of the generation step.

The fix is also upstream. A different model will not solve this. A more expensive prompt will not solve this. What solves it is a brand constraint document that the model is given before it writes anything, and a review checkpoint before the copy goes to Rightmove.

The brand constraint document does not need to be long. It needs to answer four questions: What words do we never use? What is one thing that makes our listings different from a national corporate? What are the property types we specialise in, and what do buyers of those properties actually care about? What does our ideal listing description feel like to read?

That last question is the one most agencies skip. They can describe their brand in abstract terms. They cannot produce a worked example. Without a worked example, the model has nothing to anchor to.

The workflow that works

The current failing workflow looks like this: negotiator gathers property details, pastes into a prompt template, generates copy, makes minor edits, uploads to portal. The brand layer is missing entirely. The review checkpoint is "does this look roughly right."

A better workflow adds two steps and removes none.

listing_copy_workflow:
  steps:
    - id: brief_assembly
      owner: negotiator
      inputs:
        - property_details_form
        - vendor_conversation_notes
        - comparable_listings_to_avoid
      output: structured_property_brief
 
    - id: constrained_generation
      owner: ai_tool
      inputs:
        - structured_property_brief
        - brand_constraint_document
        - banned_adjectives_list
      prompt_note: >
        Brand constraint document and banned list are prepended to every
        generation call. Not optional. Not left to negotiator discretion.
      output: draft_listing_copy
 
    - id: differentiation_review
      owner: senior_negotiator_or_branch_manager
      check:
        - does_opening_sentence_reference_something_specific_to_this_property
        - are_any_banned_adjectives_present
        - could_this_description_apply_to_a_different_property_on_the_same_street
      output: approved_or_returned_for_revision
 
    - id: syndication
      owner: negotiator
      trigger: differentiation_review_approved
      destination: rightmove_and_other_portals

The banned adjectives list is worth treating as a standalone document. "Stunning," "sought-after," "contemporary," "ideal for," "spacious," "beautifully presented," and "must-see" should be on it by default. Not because they are wrong words, but because they are so overused on Rightmove that they carry no information. A buyer reading them learns nothing about the property. A model defaulting to them is a signal that the brief was too weak to push it toward anything specific.

The differentiation review checkpoint is the step most agencies will resist because it adds time. It adds approximately four minutes per listing. That is the right trade-off. A listing that reads like every other listing on the portal is not saving time. It is spending it badly.

Where to start

If your agency has already rolled out AI-assisted listing copy and you suspect the output has converged, the fastest diagnostic is a manual audit. Pull your last twenty Rightmove listings. Count how many times "stunning," "sought-after," or "ideal for" appear. If the answer is more than five across twenty listings, you have a brand constraint problem, not a volume problem.

The fix sequence:

1. Write the brand constraint document. One page. Four questions answered. One worked example of a listing you are proud of.
2. Build the banned adjectives list. Start with the seven above. Add any phrase that appears more than twice in your audit.
3. Prepend both documents to every generation call. Make this structural, not optional. The negotiator should not be deciding whether to include them.
4. Add the differentiation review checkpoint. Assign it to someone with authority to return copy. Not a vague "check before uploading" instruction.
5. Review the output after thirty listings and adjust the constraint document based on what is still slipping through.

This is not a technology project. It is a process design project that happens to involve a technology. The brief quality problem in AI workflows is the same problem whether you are writing property listings or marketing copy. The brief is the product. The model is the printer.

If your agency has adopted AI-assisted copy and the listings are starting to look like everyone else's, the issue is in the workflow design upstream of the tool. An AI Workflow Audit will identify exactly where the brand constraint layer is missing and what it needs to contain. The output is a concrete brief architecture, not a report that sits in a folder.