What AI reads
before it
names you.

Ask ChatGPT, Perplexity and Gemini the same question about who to hire in Fort Lauderdale and you get three different lists. Anyone can see that in an afternoon.

What nobody could see was why. It isn't that one is smarter. Each one goes looking somewhere different, and where it looks decides who it finds.

These aren't three descriptions of one behaviour. They're separate machines with separate appetites, and an agent winning on one can be invisible on another for reasons that have nothing to do with how good they are at the job.

We got the third one wrong the first time. Our initial reading was that Perplexity names you because you are on somebody else's list. It held for five agents and broke on two, and we found the two in our own data before publishing. The corrected version is below, and it's the most useful part of this.

Every source,
for six weeks.

Each week we ask the assistants the questions buyers and sellers actually type. We record the answers. What we'd never done until now is record where those answers came from.

Every assistant hands back its sources. Perplexity lists them. Gemini attaches them to the sentences they support. ChatGPT marks them inline. Google returns them as results. They'd been sitting in our stored responses for six weeks, unread.

We read them. 35,925 citations, across 65 readings. Then we sorted every cited page by what kind of thing it was: a listing portal, a ranking article, an agent's own website, social media, trade press.

Kind of page	ChatGPT	Gemini	Perplexity	Google
Agent and team websites	41.4%	74.2%	53.0%	40.6%
Listing portals	40.8%	16.8%	24.3%	23.5%
Ranking lists	14.3%	4.9%	17.9%	6.7%
Social and forums	0.2%	3.2%	4.8%	21.3%
Trade press	3.3%	0.8%	0.1%	7.9%

Share of all cited pages by type, per platform. Broward County, June through August 2026.

Read that table slowly and three personalities separate. Gemini spends three quarters of its attention on agents' own websites. ChatGPT is the portal reader, and it barely touches social at all: five citations out of three thousand. Google is the only one where what people say about you in public counts for much, at a hundred times ChatGPT's rate.

Door oneGemini

It reads your website.
Actually reads it.

If you want one clean example of a machine doing exactly what it looks like it's doing, here it is.

Agent's own domain	ChatGPT	Gemini	Perplexity	Google
juliejonesluxury.com	1	631	89	28
gillesraisfinehomes.com	0	372	337	408
lauriereader.com	19	216	191	92
whitneydutton.com	0	74	140	41

Times each agent's own website was cited, by platform, across all readings.

Julie Jones is the cleanest case in the whole corpus. Her site was cited 631 times by Gemini and once by ChatGPT. She is named by Gemini four times more often than an average agent, and the reason is in the sources.

On the sentences that name her, 91% of what Gemini cited was her own website. For Gilles Rais it is 86%. This is not a correlation across a market, it is what the machine was reading at the moment it said the name.

So if Gemini is where you're thin, the fix isn't mysterious. It's your own website, and whether there's anything on it worth quoting.

The forkPerplexity

Three ways in, and agents
sort cleanly
across them.

We nearly published something simpler. Five of Perplexity's most-named agents arrive through a ranking list, so "Perplexity reads other people's lists" looked like the answer. Two agents in our own data would have been counterexamples.

The honest version needed the sentences Perplexity actually wrote. So for each agent we took only the sentences naming them, and asked what got cited on that sentence.

What Perplexity cited on the sentences naming each agent. Broward County, six weeks.

Read down the table and three routes separate. Five agents arrive through a ranking list. One arrives through her own website. Two arrive through a portal profile.

Two of them have no website we could find, and they arrive by completely different routes. One goes through a ranking list at 95%. The other goes through a portal profile at 79%. Not having a site doesn't push you onto somebody's list, and if we'd published the simpler sentence the second one would have been sitting in our own data disproving it.

Perplexity · 31 August 2026 · "top real estate agents in fort lauderdale"

"Their list includes agents such as Chris Toomey, Ron Haibi, Sandra Rathe, Larry Mastropieri, Liz Caldwell, Brian Boles, and others."

What it read to say that

One directory’s ranked list for this city

One page. Five of the eight agents Perplexity favours most in this market, named in one sentence, from one URL. It accounts for 92% of every citation we recorded from that source.

We know which page. We’re not naming it here, because for each market there is a different one, and finding it is the thing we do.

Not a domain. Not a genre of page. One ranked list for one city, which Perplexity reads and recites back almost verbatim.

Which is why the useful question isn't "how do I get on lists." It's which page, in your market, this assistant happens to read. That answer is different in Fort Lauderdale than in Naples, it changes, and you can only know it by looking every week.

Door twoChatGPT

It names people
it has never read.

This one we didn't expect, and it's the one worth sitting with.

ChatGPT · 31 August 2026 · "best luxury real estate agent in fort lauderdale"

"For luxury waterfront in Fort Lauderdale, Josh Dotoli and the Dotoli Group are among the most consistently referenced names."

What it read to say that

Nothing on his website. Not once in six weeks.

ChatGPT names him more than four times as often as an average agent in this market, and has never cited joshdotoligroup.com. Perplexity, which barely names him, has cited it 262 times.

He is not alone. John D'Angelo has one of the largest agent websites in the entire corpus, cited 3,002 times across the platforms. ChatGPT accounts for 0.6% of that.

Across ChatGPT's most over-represented agents, 19 own-site citations out of 3,385. Six-tenths of one percent.

So ChatGPT knows these people from somewhere that isn't their website. It reads listing portals, it recalls what it absorbed in training, and it produces a name without having looked the person up. Different machine, different door. Improving your site won't open it.

One problem, three fixes,
and they don't overlap.

Every agent we've spoken to assumes AI visibility is one thing. Get more visible, appear more often, everywhere at once.

It's three things. Invisible on Gemini is a website problem. Invisible on Perplexity is a not-on-the-lists problem. Invisible on ChatGPT is a problem that fixing either of those won't touch.

Work on the wrong one and you can spend six months improving something the platform you care about never looks at.

What this doesn't show,
and what we got wrong.

We wrote a version of this with three doors and had to withdraw one. The Perplexity claim rested on five agents who fit and two who didn't, and the two were sitting in our own data the whole time. What fixed it was measuring the sentences rather than the market.

It's one county. Broward, six weeks. Miami-Dade and Palm Beach are running now and we'll publish what they say, including if they disagree with this.

A citation next to a name isn't proof it's about that name. We record where each source sits in the answer, so the sentence-level version is coming. Today we can say a page was cited in a reading, not that it was cited about one particular person.

A missing website in our data is a floor, not a fact. Our matching is conservative and misses some domains. When we say a site was never cited, we mean we couldn't find it, which isn't quite the same thing.

And this is observation, not experiment. We watched what the assistants read. We didn't change anything and then measure what happened.

That study is running now. One agent in this dataset got himself onto the page Perplexity reads in his market, on 2 September, deliberately, to test the finding. His position beforehand is on record, a nineteen-agent control was captured the same way, and the first honest reading is two weeks out. We'll publish it whether it moves or not.

What we changed, and why.

• 4 September 2026 · corpus figures corrected from the whole table to the evidence. This study’s corpus figures counted every row in our database, of which 68.2 per cent came from test and stress accounts. Corrected: 35,925 citations (was 113,097), 65 readings (was 1,018, a count of per-platform result rows), six weeks (was thirteen), 369 agents discovered (was 1,079), 171 corroborated (was 391). The thirteen weeks reproduces from no count; it is the span from our first test reading to our first real one, the window before evidence existed, and we published the length of our setup as the depth of our data. Corroborated now means seen in three or more separate real readings, read from the per-reading record rather than a stored counter that counted stress readings. Every figure on this page is now served live from the same source the product reads.