Getting recommended by ChatGPT, Claude and AI search

A thing that’s changed in the last two years, quietly: a meaningful number of people looking for an agent now start by asking an assistant. “Who’s a good realtor in Sarasota for waterfront condos.” “Find me an agent in Nashville who works with first-time buyers.”

Nobody has good numbers on how big this is yet, and I’d be suspicious of anyone who claims they do. But it’s not zero and it’s growing, and the mechanics of how you get named are different enough from classic SEO to be worth understanding.

The uncomfortable part first

Most of what determines whether an AI names you happens somewhere other than your website.

When a model answers “best real estate website platforms” or “good agents in Kansas City,” it isn’t reading everyone’s homepage and judging. It’s drawing on patterns across a lot of sources: directories, review platforms, listicles from industry publications, forum threads, local news, and yes, the sites themselves — but the sites are one input among many, and not the most trusted one.

A business that appears across G2, a couple of industry roundups, a Reddit thread and a local news mention reads as real. A business that appears only on its own domain, saying excellent things about itself, reads as unverifiable. Models are, sensibly, cautious about the second kind.

This is annoying because it’s the part you control least. It’s also the part that matters most, so I’d rather say it than sell you a checklist of meta tag tweaks.

What that means practically

Get listed where your category gets evaluated. For software that’s G2, Capterra, Product Hunt, AlternativeTo. For an agent it’s Zillow and Realtor.com profiles, your local board, Google Business Profile, and any “best agents in [city]” roundup that’ll have you. Most listings are free. They usually pass no link equity at all, which is fine, because what you’re buying isn’t links. It’s corroboration.

Get into the roundups. The listicles that rank for “best X in [city]” are disproportionately what models quote, because they’re structured, they’re comparative, and they read as editorial. Many accept submissions and nobody asks.

Reviews, with substance. Not just a star count. Reviews with text that mentions what you actually did and where. “Helped us buy in [neighborhood]” is a sentence a model can use. “Great experience!” is not.

Be consistent. Same business name, same spelling, same address, same phone, everywhere. Entity resolution is a real problem for these systems, and inconsistency across sources makes you look like three half-documented businesses instead of one solid one.

What still helps on your own site

Less than the above, but not nothing, and it’s the half you can do this week.

Answer questions directly. Pages that state a fact plainly — what something costs, what’s included, who it’s for — get quoted. Pages that build atmosphere for six paragraphs before getting to the point don’t, because there’s nothing extractable in them. A short summary at the top of a page is worth more than it looks.

Structured data. Schema markup for your organisation, your location, your services, your FAQ content. It’s how a machine understands what a page is about rather than inferring it from prose. Cheap to add, and most sites don’t have it.

Specifics beat adjectives. “Serving the greater Sarasota area since 2011, with 40+ closings in Lakewood Ranch” is usable. “Your trusted local expert” is filler that appears on ten thousand sites and distinguishes you from none of them.

An llms.txt file. A plain-text map of your site at /llms.txt, for assistants fetching your domain directly. I’ll be honest about what this does and doesn’t do, because there’s a lot of overclaiming: no major model uses llms.txt as a ranking input. It matters in one specific case, which is when something fetches your domain and wants a map. That’s a narrow but real case, it costs an hour, and it’s worth doing. It is not a growth strategy.

The mistake we made on our own

Ours ran nearly a hundred lines and about half of it was comparisons against competitors, each one carefully explaining who they’d be a better fit for.

Which is a fine thing to have on a comparison page a human reads in full. In a summary file an assistant might quote three sentences from, we’d essentially written our competitors’ sales copy and handed it over. It also contained a price that was wrong for months, which is worse, because pricing is the single thing people most often ask an assistant about.

We rewrote it as a map: what we are, what things cost, where the authoritative page for each subject lives. Shorter, and more useful.

The lesson generalises. Whatever you write for machines should be the stuff you’d want quoted verbatim, because that is what happens to it.

How to check where you stand

Ask the assistants directly. Open ChatGPT, Claude and Perplexity and ask the questions your buyers would: “who are the best real estate agents in [your city]”, “best real estate website companies”, whatever fits.

Three possible outcomes. You’re named, in which case look at what it cites — those sources are your leverage. You’re not named but competitors are, in which case go find out where they appear that you don’t. Or nobody local is named and it’s all national brands, which is the best case, because the position is open.

Then search your own business name and check whether the answer is accurate. That one’s worth doing quarterly. When a model gets a fact wrong about you — a price, a service, a location — the fix is fixing the sources it’s drawing on, and you can’t fix what you haven’t noticed.