Generative engine optimization for real estate agents

First

Does any of this apply to real estate?

It applies, and almost nothing written about it says so. Search for how AI decides which real estate agent to recommend and the answers are about brands, or about tools for writing listing descriptions. The question a buyer actually asks an assistant, and the mechanism that decides whose name comes back, is largely undocumented.

That gap is not evidence the mechanism is absent. It is evidence nobody has counted. We counted: four assistants, one Florida county, weeks 31 to 36 of 2026, and the pattern is unambiguous. Which agent gets named is decided upstream, by which pages an assistant reads before it answers, and it is only loosely related to who is good at the job.

So the question is not whether this applies to you. It is whether anyone has measured it in your market yet.

The distinction

Is an agent just a small brand?

Most writing about generative engine optimization is about brands. The unit of analysis is a company, the question is share of voice across a category, and the outcome is measured in mentions. That work is real and it is not this.

An agent is not a category. A buyer does not ask which brokerage has the largest share of AI answers. They ask who to call about a waterfront condo on Las Olas Isles, and the assistant returns two or three human names. The unit is a person, the scope is a few square miles, and the question carries an intent, buying or selling or downsizing, that changes which sources the assistant consults before it answers.

That difference is not a matter of scale. It changes what you can measure and what you can act on. Brand-level GEO can tell a company it is under-represented. It cannot tell an agent that Gemini reads their website and ChatGPT never has, which is the finding that decides what to do next.

The definition

What does it actually mean to be optimised?

Generative engine optimization for a real estate agent is the practice of making a named individual findable and correctly described when a buyer or seller asks an AI assistant who to work with in a specific market.

It has two halves and they fail differently. The first is presence: whether an assistant names you at all. The second is accuracy: whether what it says about you when it does is true. Presence without accuracy is worse than absence, because a wrong specialty or a wrong neighbourhood attached to your name travels further than silence.

Both are decided upstream of you, by which pages an assistant reads before it answers. That is the mechanism, and it is measurable.

Worked example

What does the mechanism look like in real data?

Definitions are cheap. This is what the mechanism looks like in measured data. We traced every page four assistants opened across 35,925 citations in Broward County, then counted how often each one cited an agent’s own website. Tracing is the hard part: every figure below runs from the sentence an assistant produced back to the page it read to say it.

This is the whole argument for measurement over advice. The same site is a strong asset on one platform and nearly invisible on another, in the same market, in the same weeks. Any recommendation that does not say which platform it is about is guessing.

Read the full study, which traces each finding from the sentence an assistant produced back to the page it read to say it.

What follows from it

What follows from it?

Production does not predict presence. Among the 25 highest-producing agents in Broward County, closed volume varies by a factor of five while appearances in AI answers vary by a factor of three thousand. Six of those 25 were never named at all. Being good at the job is not the mechanism. Weeks 31–36 of 2026.

The platforms read different things. They are separate machines with separate appetites. An agent winning on one can be invisible on another for reasons unrelated to how well they work.

Your own site is a lever on some platforms and not others. Which is why the first question is not what to publish, but who is reading you now.

Limits

What can this page not tell you?

• It is one county. Every figure here is Broward County, Florida, weeks 31 to 36 of 2026. We have five further Florida markets seeded and unpublished, and we do not publish a market until it has corpus. Whether these patterns hold elsewhere is an open question and we will not assert that they do.
• It describes, it does not prescribe. The data shows which platforms read agents’ websites in this corpus. It does not establish that publishing more will cause an assistant to name you, and we have not run that experiment.
• 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.

The methodology states the four weights, the sample, and the four numbers we refuse to compute.