17 AI Writing Tells That Are Killing Real Estate Copy (And How to Strip Them)
Vendors and buyers are starting to spot AI-written copy. The "delve into the intricate tapestry of coastal living" voice that's showing up in listing descriptions, vendor reports, and agent social posts has become a tell. Every time it shows up under your name, it costs you.
When every agent's copy sounds the same, there's no differentiation. And copy that reads as obviously AI-generated reads as low-effort. That's the worst signal you can send a vendor who's deciding who to list with.
The good news is that AI tells are predictable. There are 17 of them that I have found. Once you know what to look for, you can either edit them out yourself or build the check into your AI workflow so they never make it into your final content.
This article is the field guide. The 17 patterns below are taken verbatim from the rule set I built to humanise AI-generated copy at Bash & Co. At the end, you'll find three ways to bake the rules into your own AI tools, plus a copy block with the full skill write-up if you want to install it in Claude.
Why this matters for real estate
Agents are leaning on AI in three main places right now:
Listing descriptions for Trade Me and realestate.co.nz.
Vendor reports and market updates.
Social posts and personal-brand content on LinkedIn and Instagram.
Each one is a trust touchpoint. A vendor reads your CMA and decides whether to work with you. A buyer reads your listing description and decides whether to book the open home.
If your copy gives off AI tells in any of those moments, the message you've sent is that you couldn't be bothered to write it yourself.
The 17 patterns
1. Inflated significance
Puffing up a subject with generic claims about its importance or role in a broader movement.
Before: The opening of the cafe marked a pivotal moment in the suburb's culinary landscape.
After: The cafe opened in March.
2. Promotional tone
Travel-brochure admiration, especially around heritage, community, legacy.
Before: A vibrant community with a rich cultural heritage and a significant place in the region.
After: A suburb of 4,200 people, settled in 1862.
3. Superficial analyses
Phrases that gesture toward meaning without delivering it. Empty-calorie insight.
Before: It's not just about efficiency. It's about transformation.
After: The new system cuts processing time by 40%.
4. Vague attributions
"Experts say," "many scholars agree," "it is widely considered," with no one named.
Before: Many experts consider this approach groundbreaking.
After: Name the expert, or cut the claim.
5. Em dash overuse
Em dashes used for punchy emphasis where a comma, full stop, or parentheses would work.
Before: The result—surprisingly effective—changed everything.
After: The result was surprisingly effective.
6. Rule of three
Compulsive triplets in adjectives, benefits, or list items.
Before: Fast, reliable, and scalable.
After: Fast and reliable. (Or name a specific benefit.)
7. AI vocabulary
Cursed words that LLMs overuse. Core list: delve, intricate, tapestry, pivotal, underscore, landscape (metaphorical), foster, testament, enhance, crucial, realm, robust, seamless, holistic, elevate, navigate (metaphorical), leverage, paramount, cutting-edge.
Before: We delve into the intricate tapestry of modern commerce.
After: Here's how modern commerce works.
8. Negative parallelisms (and the negation stack)
"It's not X, it's Y" constructions, plus stacked variants like "Not X. Not Y. Just Z." or "It's not just X — it's Y." AI reaches for rhetorical contrast even when the assertion alone would land harder. Listing what something isn't before saying what it is reads as performative scaffolding.
Before: It's not just a checkbox — it's a brand asset.
After: Treat it as a brand asset.
Caveat: intentional emphatic versions with all-caps stress ("The fix is NOT small. But it is steady.") read as deliberate rhetoric, not AI default. Leave them alone when used sparingly and the writer is making a real point.
9. Excessive conjunctive phrases
"Moreover," "Furthermore," "In addition," "Notably," "Importantly" used as padding.
Before: Moreover, this approach is effective. Furthermore, it scales easily.
After: This approach is effective. It scales easily.
10. Editorialising
Slipping interpretation or opinion into what should be neutral description.
Before: The company's bold decision to pivot fundamentally reshaped the industry.
After: The company pivoted in 2022.
11. Overgeneralising opinions
One or two sources' views framed as widely held.
Before: Widely regarded as one of the most influential figures in the field.
After: Name the source; drop the framing.
12. False ranges
"From X to Y" when there's no real spectrum, just two loosely related things. Before: From intimate gatherings to global movements. After: Drop the construction; name what you actually mean.
13. Formatting overkill
Bolding every key term, every bullet shaped as "Term: definition," numbered lists where prose would work.
Before: Scalability: The system handles large volumes. Reliability: The system rarely fails.
After: The system handles large volumes and rarely fails.
14. Compulsive summaries
"Overall," "In conclusion," "In summary" tacked onto short passages that don't need summarising.
Before: Overall, the project was a success.
After: Delete, or state the specific outcome.
15. Tailing importance clauses
Sentences that end with hazy significance: "...a testament to...", "...reflecting the continued relevance of...", "...emphasising the significance of...".
Before: The festival returned in 2024, reflecting the continued relevance of local music.
After: The festival returned in 2024.
16. Pretentious adverbial drift
Adverbs like "quietly," "silently," "slowly," "subtly" paired with active or abstract verbs to manufacture gravitas without earning it: quietly hurt, silently shape, slowly drift, quietly compound, silently kill, subtly erode. AI reaches for these when it can't get specific. Real writing either names what's gradual or doesn't need the adverb at all.
Fix: drop the adverb. If the sentence still works without it, it wasn't needed. If gradualness genuinely matters, name what's gradual ("compounds over six months" rather than "quietly compounds").
Before: These habits quietly compound over time.
After: These habits compound over six months. (Or: These habits compound.)
17. Abstract "way it Xs" constructions
Phrases that swap a concrete subject for a philosophical abstraction: the way the room actually feels, the way a brand lives on social, the way a listing breathes online. AI defaults to these when it doesn't have a concrete observation, so it abstracts toward the universal.
Fix: replace with what's concrete. Not "the way the room feels" — "what the naked eye sees" or "what someone walking through actually notices."
Before: Photography that captures the way the room actually feels.
After: Photography that matches what the naked eye sees walking through.
How to actually use these 17 rules
Knowing the rules is one thing. Getting them to fire every time you generate copy is another. Three ways to do it, in increasing order of how well they hold up under load.
Option A — Add the rules as a custom instruction in your AI tool
Most AI tools let you set a permanent instruction the model follows on every chat. ChatGPT calls these Custom Instructions. Gemini has Saved Info and Gems. Copilot has a similar custom-prompt setting.
Paste the 17 rules in there once, and the model checks against them on every reply with no need to re-prompt.
The catch: Two things to watch out for here:
Most tools have a character limit on these settings, and the full rule set is long. You may need to compress it down to a shorter version that captures the patterns without all the examples. The trade-off is fidelity. The examples are what teach the model what each pattern looks like in practice, so trimming them means the model has less to anchor against.
They cost you in usage/tokens. If you bake these into every instructions, you can burn through your token limits quickly.
Option B — Paste the rules into your prompt
The simplest version. When you ask the AI to write a listing description, vendor email, or social post, paste the full rule set at the top of your prompt and tell the model to apply it before returning the draft.
This works on any AI tool, with no setup. The downside is repetition: you have to paste the rules into every new chat where you want them applied.
Option C — Build it as a triggerable skill (Claude users)
Claude has skill functionality baked into the platform. A skill is a self-contained instruction file the model loads when triggered, either by a phrase you say (like "humanise this") or as a sub-step in a larger job.
Once the skill is installed, you can ask Claude to "write a vendor report on 14 Smith Street" or "draft a LinkedIn post about my latest sale" and tell it to run the humanising rules as part of the job. The draft comes back already humanised. No extra editing pass.
The full skill content is below. Copy it, read through the rules end to end, and then ask Claude to install it as a skill for you. Claude will handle the installation steps.
Read the rules before you ask Claude to install the skill. The skill is what runs on your copy. You should know what it's doing before you let it run.
A word of warning: Always look through third party skills before installing them. Devices like this can be used maliciously to compromise your personal information.
Use AI to draft, edit before you publish
Agents who edit their AI drafts properly will sound human. Agents who paste straight from ChatGPT into Trade Me will sound like everyone else who pasted straight from ChatGPT.
The 17 rules above are the editing pass. Whether you bake them in as a custom instruction, paste them into your prompt, or trigger them as a skill, the work is the same. Your copy stops sounding like AI, and your brand voice starts sounding more human.
If your AI tools are doing the writing on your behalf, make sure they're doing it like you would.
FAQ
Can vendors and buyers really tell when copy was written by AI?
Increasingly, yes. The phrases AI defaults to ("vibrant community," "rich tapestry of coastal living," "delve into," "a testament to") were uncommon in property writing five years ago. Now they show up across multiple listings every week. Anyone who reads property listings regularly (which includes most active vendors and buyers) recognises the pattern, even if they can't name what they're recognising. The signal lands as "this agent didn't write this."
Does humanising AI copy make it less efficient to use AI at all?
No. The point of using AI is to skip the blank-page problem and get a usable first draft fast. The 17 rules are an editing pass on that draft, not a replacement for using AI. Even with the editing pass, you're still writing in a fraction of the time it would take from scratch. You're just publishing copy that doesn't sound like every other agent who used the same tool.
Will Google penalise AI-written real estate listings?
Google's official position is that AI-generated content isn't penalised on the basis of how it was created. It's penalised when it's low-quality, unhelpful, or written for search engines instead of people. Most AI-generated property copy gets flagged because it's unedited, generic, and indistinguishable from a thousand other listings. The AI didn't cause that. The paste-without-editing did. Edited AI copy that reads naturally and gives buyers something specific to engage with does fine in Google's results.
What's a Claude skill and is it safe to install?
A Claude skill is a self-contained instruction file that tells Claude how to handle a specific kind of job. Once installed, Claude loads the instructions whenever the skill's trigger conditions are met. The skill in this article doesn't access your data, doesn't make API calls, and doesn't run code. It's plain text instructions that change how Claude rewrites text inside your chat. Safe to install. The reason we recommend reading the full instructions before installing is so you understand what the skill is doing to your writing, not because there's any technical risk.
