AI Writing for Realtors: What Works, and What’s Quietly Flagging You as Spam
Gmail’s Gemini-era filters can spot AI-generated emails before a human sees them. Here’s what’s still landing in the inbox, and what’s quietly getting routed to spam.
May 19, 2026 - The Nudgey Team
In February of this year, the cold email tool TextPolish ran a quiet experiment that should scare every realtor using AI to draft follow-ups.
They sent 5,000 AI-generated cold emails through a properly warmed-up domain with all the right SPF, DKIM, and DMARC authentication in place. By every traditional measure, the campaign should have landed in the inbox.
Open rate: 0.2%.
The emails weren\u2019t filtered for technical reasons. They were filtered for syntactic ones. Gmail and Outlook had learned what AI-generated cold email looks like (the cadence, the phrasing, the structure) and they were quietly routing it to Promotions and Spam before a single human eye ever saw it.
For realtors, this is the story of 2026 in one experiment. AI writing tools are now table-stakes. The question isn\u2019t whether you should use them. The question is whether the AI you\u2019re using will get you flagged.
What changed in early 2026
The big shift happened in January, when Google integrated Gemini AI throughout Gmail. The feature most users see is the auto-summary at the top of long threads. The feature most senders don\u2019t see is the one that matters: Gmail now runs every incoming email through what deliverability researchers are calling a \u201crelevance test\u201d before deciding whether to surface it to the user at all.
According to deliverability platform Folderly\u2019s research, the old binary of \u201cinbox vs. spam\u201d is gone. There\u2019s now a gradient of visibility within the inbox itself. Their data found that up to 40% of emails reaching Gmail inboxes are being deprioritized by AI filtering. Technically delivered, effectively invisible.
Meanwhile, Gmail\u2019s spam infrastructure as a whole now blocks nearly 100 million spam emails every minute, with AI-enhanced filters catching more than 99.9% of spam, phishing, and malware before they hit inboxes. The filters run on Google\u2019s own AI models (RetVec, Gemini Nano, and TensorFlow) and they\u2019re trained on more spam than any human has ever read.
The relevant detail for realtors: over 51% of all spam is now AI-generated. Google\u2019s filters have seen vastly more AI-written email than human-written email at this point. They know exactly what it looks like.
The \u201cAI perplexity\u201d problem
Here\u2019s the technical concept that explains why generic AI writing fails so badly.
Modern spam filters use transformer models (the same underlying technology as ChatGPT and Gemini) to evaluate incoming text. One of the things these models can measure is perplexity: how surprising or unexpected the next word in a sentence is, given the words that came before it.
Human writing has high perplexity. We pick odd words. We start sentences with conjunctions. We use phrases that don\u2019t quite fit. We make small grammatical choices that no model would predict.
AI writing has low perplexity. It picks the most likely next word almost every time. That\u2019s literally how the model works.
A spam filter doesn\u2019t need to know an email was written by AI to flag it. It just needs to notice that the text reads like every other piece of low-perplexity content the filter has been trained on. And a huge share of that training data is now spam.
Practically, this means an email that opens with \u201cI hope this email finds you well\u201d and continues with \u201cI wanted to reach out regarding...\u201d doesn\u2019t fail because those phrases are inherently bad. It fails because every spam bot in the world is now using them, and the filter has learned to associate them with low engagement.
What \u201cgood\u201d AI writing for realtors actually looks like
The fix isn\u2019t to stop using AI. It\u2019s to use AI that produces output a human realtor would actually write, and then to edit it like a human would.
A few patterns that consistently survive filters and get replies:
Lowercase subject lines. \u201cquick question about the bayview listing\u201d outperforms \u201cQuick Question About the Bayview Listing.\u201d Sales bots almost always title-case. Humans texting from their phone don\u2019t.
Short, broken sentences. Two-word sentences. Sentence fragments. Starting with \u201cAnd\u201d or \u201cBut.\u201d Spam filters are trained on long, grammatically perfect sales copy. Real people write like they talk.
No greeting at all, or just a first name. \u201cHey Jordan\u201d beats \u201cDear Jordan.\u201d No greeting at all beats both.
A real-world reference. One specific detail the lead actually mentioned. The school district they asked about, the price range they specified, the neighborhood they walked through on Saturday. Generic personalization tokens like {firstname} don\u2019t count. The filter knows what those look like too.
A P.S. AI tools rarely use postscripts naturally. A handwritten-feeling P.S. (\u201cp.s. saw a new listing on Elm this morning, want me to send it over?\u201d) reads as human to both the recipient and the filter.
No images, no link tracking, no fancy HTML. Plain-text emails consistently outperform designed ones for one-to-one follow-up. They look like something a person actually typed. The moment you add a tracking pixel and an unsubscribe footer, you\u2019re sending a marketing email, and the filter treats it like one.
The deliverability checklist that still matters
The syntax stuff is new. The infrastructure stuff hasn\u2019t changed:
- SPF, DKIM, and DMARC authentication are required, not optional. Emails that fail DMARC in 2026 often never reach the spam folder at all. They\u2019re rejected at the mail server level.
- Domain warmup. A new sending domain needs 4\u20136 weeks of gradual volume ramping (5\u201310 emails/day to start) before you push real campaigns. Skipping this is the fastest way to torch a domain.
- Consistent send volume. Erratic patterns (500 emails Monday, nothing for three days, 1,000 on Friday) look like spam-bot behavior. Steady daily volume looks like a real business.
- Engagement quality. ESPs increasingly weight reply depth, read time, and conversation length when deciding inbox placement. One actual conversation is worth a hundred opens.
The threshold has gotten lower, too. By 2026, even a spam complaint rate as low as 0.3% can get a domain blacklisted. On a list of 500 leads, that\u2019s 1.5 complaints. One bad week and you\u2019re done.
The honest takeaway for realtors
AI-generated follow-up emails aren\u2019t dead. Lazy AI-generated follow-up emails are dead.
The realtor sending 200 generic \u201cJust checking in!\u201d messages a week from a generic ChatGPT template is now getting filtered into the void, regardless of how good their domain reputation is. The realtor using AI to draft a specific, short, human-sounding message that references something the lead actually said, and then sending it on a sensible schedule from a properly warmed domain, is still landing in the primary inbox at high rates. The key is having a system that actually works your database consistently rather than relying on manual effort.
The technology is doing more work than ever. It just has to do the right kind of work. The filters got smart. The writing has to get smarter to match. And the biggest ROI still comes from working your existing relationships \u2014 your past-client database is the most undervalued asset most agents own.
Five-minute response times still win the lead. But only if your message actually shows up in the inbox.