Inman published an interview on September 10, 2026 with Sam Mehrbod, CEO of the real estate marketing platform Roomvu, about what actually drives AI search visibility for agents. Most of what he describes matches what we see when we open up agent websites. One part of it is where we take a different line, and that difference is the reason Saige exists.
The study sitting underneath the conversation
The interview follows a piece of research worth reading on its own. Local Falcon, a company that tracks local search rankings, published The Realtor AI Visibility Index on August 18, 2026. It ran 37,500 searches across the 100 largest US cities on five systems: Google AI Overviews, Google AI Mode, Gemini, ChatGPT and Grok. Three phrases were tested, "best realtor near me", "best real estate agent near me" and "top rated real estate agent near me", each run from 25 points on a grid covering a two mile radius from the city centre.
The number that travelled is the one Inman reported on August 20: 91.5% of ranked, website-publishing agents were never named once. Two more findings from the same study matter as much. Only 21.2% of the highest-volume individual agents in those cities were cited by any platform, and 82.6% of the agents who were named at all were named by a single platform, with almost no overlap between systems.
Read the methodology before you read the headline, though. Three "near me" phrases are one narrow family of questions. A buyer asking "who should I talk to about selling a townhouse in Kitsilano before the school year starts" is asking something different in shape, and the study did not test it. The 91.5% figure is a real result about broad discovery phrases in big US cities.
Where the interview matches what we see
Three of Mehrbod's points line up with what shows up in our own work on agent sites.
The first is that long-tail terms are the winnable ones. His example is that every agent in San Francisco wants "top agent in the Bay Area", which is close to impossible, while "top agent in Berkeley who speaks Spanish" is achievable. The mechanism behind that is simple. Ask a broad question and thousands of pages could plausibly answer it, so the engine picks from a crowded pool. When someone names a neighbourhood, a language and a property type in the same question, the pool shrinks to the handful of pages that actually contain those words together.
The second is that hyperlocal content beats generic content. "First-time homebuyer tips" has been written ten thousand times. What a stadium renovation does to housing values four blocks away has been written once, if at all. Maybe never. We have made the same argument in our post on what to put on a neighbourhood page: the page earns a citation by containing facts nobody else bothered to write down.
The third is that this is still trial and error. No AI company publishes a ranking algorithm the way Google publishes parts of its guidance, so anyone claiming a repeatable formula is extrapolating from a handful of tests. Mehrbod says the same thing plainly in the interview, which is more honest than most vendor commentary on this topic.
The claim we read differently
The line in the interview that we would push back on is this one: the front door to an agent's business used to be a website, and now Google Business Profile, YouTube, Instagram, Facebook and email share that role.
The first half is accurate. Discovery genuinely does start in more places than it used to, and an agent with no Google Business Profile and no video is absent from channels that matter. Our own post on Google Business Profile for agents says as much.
The second half is where the Local Falcon data argues against the conclusion. Eighteen of the 25 most-cited agent websites in that study carry some version of a "best real estate agents in [city]" page. The agents who got named by AI systems largely got named through a page published on a domain they own.
The website's job changed. It used to be the place a buyer browsed after they already knew your name. It is now the place an engine quotes from, and the destination every other channel points back at. A Google Business Profile post links somewhere. An AI answer names a source. A referral looks you up. All three land on a page, and the quality of that page decides what happens next. That is why we treat the site as the foundation rather than one channel among several, and why an agent relying on a brokerage profile page instead of their own site is handing that decision to someone else.
Watermarking changes less than the headlines suggest
Inman opened the interview by asking whether AI text watermarking works for or against agents who use AI to write. It is worth knowing what the watermark actually does before deciding.
Anthropic's own explanation, How Claude's text watermarking works, published August 14, 2026 and updated September 1, says the watermark works by using a key and the preceding words to settle which word the model picks next. The pattern is undetectable to a reader and detectable to anyone holding the key. It adds no tokens, costs nothing in speed, and carries no identifying information about users or organizations. Anthropic also notes that detection is weaker on short passages and on factual content, where the model has fewer word choices available to encode anything.
So the watermark answers one question: did this text come out of this model. Reading it requires the key, which Anthropic holds, and the answer it gives says nothing about whether the page is useful.
Mehrbod's answer lands in the same place from a different direction, and we agree with it. A neighbourhood page that names the streets, the price band and the school catchment can be quoted by an engine whether a person or a model typed the first draft. A page that says the community offers something for everyone cannot be quoted by anything, because there is no fact inside it to lift.
What an AI system can actually verify about you
Asked what an engine can confirm about whether an agent is legitimate, Mehrbod named three things: testimonials written by other people, sold listings with a written story behind the sale, and hyperlocal FAQs. That list matches our experience, and the reason it works is worth spelling out.
Each of those is a fact that comes from a source other than your own marketing copy. A review sits on Google under someone else's name. A sold listing appears in board records and on portals. A neighbourhood FAQ can be checked against reality by anyone who lives there. Claiming to be the top agent in your area is a sentence you wrote about yourself, and a machine has no way to confirm it.
The fourth item we would add is consistency. Your name, brokerage, phone number and service area should match exactly across your website, your Google Business Profile and every directory you appear in, down to whether you write "Street" or "St." AI systems treat agreement between independent sources as a confidence signal. Disagreement gives them a reason to state nothing.
Why we built Saige
The pattern that led to this product kept repeating. An agent does everything the advice says. They post weekly, keep the Google Business Profile current, film video, send the newsletter. Traffic arrives. The phone stays quiet, and no AI tool ever names them.
When we opened the website underneath, the reason was close to identical each time. There was no page that answered one question in a form anything could quote. The bio said the agent was dedicated to exceeding expectations. The areas page was a list of neighbourhood names linked to IDX search results, the feed that puts MLS listings on an agent's site, with no sentence written about any of the neighbourhoods. The FAQ had six questions with one-line answers. Nothing on the site stated, in plain text a machine could lift, who this person is, where they work, what they sell and what a visitor should do next.
Publishing more content on top of that structure adds volume without adding anything a machine can quote. That is why an agent can follow the content advice for a year and end up close to where they started.
So Saige is one product: an AI-driven rebuild of the website you already have. It starts from your existing site and keeps what works. It restructures the pages so each one answers a specific question in its first two sentences. It adds structured data, which is a layer of code invisible to visitors that states who you are, the area you serve and how to reach you in a format machines read directly instead of guessing at from prose. And it puts one clear next step on every page, because a citation that lands on a page with no way to contact you is a wasted citation.
The honest reason we lead with the website rather than a posting schedule is that the website is the only piece of this an agent fully owns. Platforms change their rules. The page you publish stays yours. You can see how that looks on your own site with a free rebuild preview, which generates the rebuilt version and shows it to you before anything on your live site changes, or read more about how we work.
Three limits on all of this
This topic attracts confident promises, so here is where ours stop.
No AI company publishes its citation rules. Anyone guaranteeing you a spot in a ChatGPT answer is guessing, and Local Falcon's finding that 82.6% of named agents appeared on one platform only shows how unstable these results are between systems. Measure across several tools, on several days, before drawing conclusions. Our post on free AI visibility scoring tools covers what those checks can and cannot tell you.
A rebuild also fixes only what is on your site. It cannot create reviews you have not earned, sales you have not made, or a Google Business Profile you have never filled in. If the citation gap in your market is a reputation gap, a better website will surface it faster rather than close it.
Changes also take time to appear. Search engines and AI systems have to recrawl and re-evaluate a site before anything shifts, and how long that takes varies by domain. We cover the mechanics in the FAQ, and you can always ask us directly.
The takeaway
The interview and the study point at the same conclusion from opposite ends. Broad discovery phrases are close to unwinnable for an individual agent, specific local questions are winnable, and the agents who do get cited are getting cited through pages they published themselves.
That is the case for treating the website as the thing you fix first. Every other channel is a way of pointing at it, and none of them can be quoted the way a well-built page can.
