Reference
AI SEO for real estate agents
Search optimisation assumed a ranked list of links. Assistants do not return one. They return a short answer naming two or three people, assembled from pages they chose to read. AI SEO is the practice of influencing whether your name is in that answer, and this page is about doing it as a named agent rather than as a brand.
Corpus as of W36 2026 (weeks 31–36) for the traced findings, with the live corpus dated separately where it appears. Every figure states its window. The method is published at methodology, including what it cannot see.
First
Does AI visibility apply to real estate?
Search for it and you will find almost nothing that says so. The pages that rank for AI SEO are written by SEO platforms about brands, or by small sites with no measurement behind them. Not one company that measures AI visibility appears in them. The category has not arrived in this market yet.
That absence is not evidence the mechanism is missing. It is evidence nobody has counted it here. We counted: four assistants, one Florida county, weeks 31 to 36 of 2026. Which agent an assistant names is decided upstream, by the pages it reads before answering, and it is only loosely related to who is good at the job. The parent concept is generative engine optimization for real estate agents; this page is the practical half.
What makes it different from SEO
Three things, and each changes what you can act on.
There is no position to rank in. An answer names people. You are in it or you are not, and a near-miss is invisible. Rank tracking has no equivalent here.
The assistants disagree with each other. They are separate machines with separate reading habits. One answer is not a sample of the others, so a single check tells you almost nothing.
The unit is a person, not a domain. A buyer asks who to call, and the answer is a human name attached to a neighbourhood. Domain-level thinking does not reach it.
What carries over
Being read still matters, and being read is still mostly a function of pages existing, being crawlable and saying something specific. What changes is which pages, and that is measurable rather than assumable.
Worked example one
Where does each assistant look?
We traced every page four assistants opened across 35,925 citations in Broward County, and sorted what they read by source type. Tracing is the hard part: each figure runs from the sentence an assistant produced back to the page it read to say it.
Share of each platform’s citations by source type. Broward County, weeks 31–36 of 2026.
| Platform | Agent’s own site | Portals and directories | Social and community |
|---|---|---|---|
| Gemini | 74.2% | 21.4% | 4.4% |
| Perplexity | 41.4% | 53.0% | 5.6% |
| ChatGPT | 27.1% | 72.7% | 0.2% |
| 38.9% | 41.5% | 19.6% |
Gemini spends three quarters of its attention on agents’ own websites. ChatGPT is a portal reader and barely touches social at all, at two citations in a thousand. Google is the only one where what other people say about you in public counts for much, at roughly a hundred times ChatGPT’s rate.
What this changes about the advice
Publishing on your own site is a strong lever on Gemini and a weak one on ChatGPT. Getting your portal profile right is close to the whole game on ChatGPT and secondary on Gemini. Neither of those is a general recommendation, and any advice that does not name a platform is averaging across machines that do not behave alike. The full study traces each row to the pages behind it.
Worked example two
How much does each platform actually cite?
Where a platform looks is one question. How much it cites at all is another, and the two together decide where effort pays.
Citations by platform in the live corpus, as of 7 September 2026. This figure moves weekly; the traced findings above do not.
| Platform | Citations | Share |
|---|---|---|
| Perplexity | 22,687 | 49.4% |
| Gemini | 12,150 | 26.4% |
| 7,035 | 15.3% | |
| ChatGPT | 4,067 | 8.9% |
Perplexity produces about half of all citations and Gemini a quarter. ChatGPT cites least, which does not mean it names agents least: it names them while showing less of its reading. A platform that cites little is not a platform that matters little, and the two are easy to confuse.
Reading the two tables together
Gemini reads your site and cites heavily, so a site is a lever there twice over. ChatGPT reads portals and cites lightly, so a portal profile matters and will be hard to verify from citations alone. That combination is not visible from either table by itself, which is the argument for measuring both rather than picking a metric.
Practice
What can an agent actually do?
Three things the data supports, and one it does not.
Know which platforms read you now
Before any change, the useful fact is which of the four currently name you and which never have. That is a measurement, not a guess, and it is the same reading we run in an audit.
Make the pages that get read say something specific
Every platform reads either your site or your portal profile. Both are pages you control, and both are frequently generic. Specificity here means a named neighbourhood, a named transaction type, a named price band, because those are the terms buyers put in their questions.
Check what is said about you, not only whether
Presence without accuracy is worse than absence. A wrong specialty attached to your name travels further than silence, and it appears in the same reading that tells you whether you were named at all.
What the data does not support
That publishing more will cause an assistant to name you. We can show which platforms read agents' websites in this corpus. We have not run the experiment that would establish cause, and we will not assert it until we have. Study 01 shows that production does not predict presence; it does not show what does.
Limits
What can this page not tell you?
- It is one county. Broward County, Florida. Five further Florida markets are seeded and unpublished, and we do not publish a market until it has corpus. Whether these shares hold elsewhere is open, and we will not assert that they do.
- Citation shares are not naming shares. A platform that cites less may still name agents often. The two tables above measure different things and neither is a proxy for the other.
- Discovered and corroborated are floors. The per-reading record disagrees with the running counter on 131 of 1,079 discovered names, and 252 entries point at readings that can no longer be found. A reading the record does not carry cannot be counted.
- Agents measured as part of a team are read on three platforms rather than four, so their figures sit against a smaller denominator than the headline implies.
- No causal claim is made anywhere on this page. Every figure describes what was observed in a window. The methodology states the four numbers we refuse to compute and why.
Published research so far is listed here.