A seller in your ZIP code asked ChatGPT who to call.

A man built a sentence by pulling words out of a hat

Type your market into ChatGPT and ask who the best agents are. Then open a new chat and ask it exactly the same way. The second answer is not a longer version of the first. It is a different answer, sometimes with a different name in it, and the machine is working properly when that happens.

The reason goes back to 1948 and an engineer at Bell Labs named Claude Shannon. Shannon wanted to know how much of English is just habit, so he drew one word at a time from a hat. A fresh hat after every word, each one holding only the words that follow the last one in print.

Nobody wrote what came out. Probability did, one word at a time, and it almost reads. That bet is seventy-eight years old and it sits under every answer an AI assistant writes today. What changed is the size of the hat.

So what is it? A prediction machine. It builds the paragraph while the seller waits. At every step it holds a list of what could come next, with odds on each one, and it does not always take the favorite. That room in the pick is why the same question came back two ways.

Two words are worth learning, because the business problem sits in the difference. Deterministic means the same question in, the same answer out. The MLS is deterministic. The engines are non-deterministic, so there is no such thing as checking once.

Sources. Claude Shannon, A Mathematical Theory of Communication, Bell System Technical Journal, 1948. The sentence is his second-order word approximation, printed in the paper. The hat is our image for his documented method: he used tables of random numbers and books opened at random. That these systems work out the next piece rather than looking up a stored record is how their makers describe them, and the same question comes back two ways because there is a deliberate roll of the dice in the picking. The MLS and the past client are ours, September 2026. Inferred.

You ask it whether your bio is any good. It says yes.

Write short to ChatGPT and it goes short. Write like a lawyer and it tightens right back up. By about the third message it is writing the way you write, and people read that as warmth.

The agreeing has a source. After a model is built it gets finished by people. They read two answers side by side and pick the better one, millions of times over, and those picks set its manner. People pick the answer that agrees with them. Reward that often enough and agreeing becomes the house style.

On a listing presentation, that costs you something. You paste your bio in and ask if it works. You paste your pricing rationale in and ask if it holds. It says yes to both, and you walk into the appointment carrying a compliment instead of a read. Then the seller asks the one question nobody rehearsed.

One sentence changes the answer. Tell it to answer as the seller who is about to list with somebody else, and to say why. That takes you out of the field it narrows on, which is the only way to get an objection instead of a compliment. Ask what is wrong with the bio, in the same message as the bio.

There is nothing to fear in a monster that turns out to be a mirror. The agreeing came from a few million people picking the answer they liked better, and it was handed back to us as a personality.

Sources. Reported: OpenAI, Sycophancy in GPT-4o, 29 April 2025, on the update it rolled back and on overweighting short-term user feedback. Sharma and others, Towards Understanding Sycophancy in Language Models, Anthropic 2023, at ICLR 2024: an answer matching a person’s stated view is more likely to be preferred, and both people and the rating models prefer convincingly written sycophantic answers over correct ones a non-negligible share of the time. That these models adopt a user’s vocabulary, sentence length and framing across a conversation is observable in any current major model. Finishing a model on human preference ratings is standard practice at every major lab. The one-sentence fix is ours. Inferred.

Forty-three percent got your name from a friend. Then they checked.

NAR’s 2025 profile has the number that has always run this business. 43 percent of buyers were sent to their agent by a friend, neighbor or relative, and 88 percent used an agent at all. None of that has moved. What moved is what happens in the four minutes after the referral.

Your name gets said at a dinner. She types it in on the way home. What comes back is a paragraph assembled out of whatever is written about you, and it decides whether she is still holding your name when she gets to the driveway.

That is the same job the first page of Google did for twenty years. The difference is that the answer now arrives already decided, in one paragraph, with two or three names in it, and nobody can see which pages it was built from.

Two thirds of first-time buyers hire the first agent they speak to. If the paragraph changes who that first agent is, it changes everything downstream of it, and it does that before anybody picks up a phone.

Sources. Reported: National Association of Realtors, 2025 Profile of Home Buyers and Sellers, published 3 November 2025: 88 percent of buyers used an agent or broker; 43 percent found their agent through a referral from a friend, neighbor or relative; 67 percent of first-time buyers and 76 percent of repeat buyers hired the first agent they spoke to. Bank of America Homebuyer Insights Report, fieldwork by Sparks Research 13 April to 10 May 2026 among 2,000 respondents, released 23 June 2026: one fifth of prospective buyers and current homeowners used AI tools or chatbots for homebuying research in the past year, rising to 28 percent of millennials and 32 percent of Gen Z. Peec AI, Top domains cited by AI search, 31 March 2026, an analysis of 30 million cited sources, published without percentages. Where the engines sit relative to the referral is our reading, September 2026. Inferred.

By Thursday it has forgotten the positioning you set on Monday

People ask whether anything is lost when ChatGPT rewrites a piece of writing. Two things are, and they are different from each other. It loses the conversation, and it loses the specifics. Nothing about this is lossless.

Every chat lives inside a window. New words go in the front, old words fall off the back, and apart from a few facts some products save, nothing that falls off is kept. So hour three argues with something you settled in hour one, and it does not flag the gap. It fills the hole with the likeliest thing, in the voice it used when it knew.

Before a long chat fills up, ask it for a handoff, half a page of what the two of you settled. What it leaves out is what you put back. Open a new chat, paste the handoff in first, and use the same opening words every time so you can tell it took. You are not saving a conversation. You are handing over a brief.

The specifics are the one that costs you something, and on a listing they are the whole asset. Ask it to write the description for a house and it hands back the likeliest version, which is the one every house gets. Charming. Light-filled. Nestled. The jacaranda that drops on the drive in May, the odd little step down into the den: those are the least likely pieces in the whole thing, and they go first.

That is true even when the sentences come back better. A rewrite is not a cleaner copy of what you gave it. It is the average of everything written in that shape. Hand it your listing and you get back a better-written listing with less of the house in it. That trade is sometimes worth making. It is never free, and nobody quotes you the price at the time.

Sources. Context windows are published for every major model, and vendor documentation for several 2026 models states a ceiling at or above a million pieces, with reliable recall falling off below the stated ceiling. Content that leaves the window is not retained. Some products add a memory feature on top, which saves a handful of facts and not the conversation. That these models continue a thread rather than flag a gap is observable in any current major model. That a rewrite regresses toward the likeliest phrasing follows from the mechanism in the hat and is our reading. The handoff list is ours, September 2026, a habit rather than a measured result. Inferred.

Go and read what it says about you

Nobody knows exactly what one of these will do next. There are four answer engines people actually use, none of them deterministic, and no clean way to watch any of them in analytics. So there is one honest method left: ask, write down what came back with the date on it, change one thing, and ask again.

In the first ninety seconds you learn whether you are in the paragraph at all, whose names are in it instead, what it thinks you sell, which is often not what you sell, and whether it has invented a number about you, which what it makes up says it will do in the same voice it uses for a real one. That is a baseline, and a baseline is worth more than an opinion.

Sources. That these systems are non-deterministic and return different answers to the same question is in the hat and is how their makers describe them. That they are not observable in standard web analytics follows from the Pew figures in the invisible click, where the reading happens without a click. The three steps are ours, September 2026, a working habit rather than a measured result. The free check asks ChatGPT, Claude and Perplexity live and dates what they return. Nothing on this page is a claim that any specific action will change what an engine says. Inferred.

The paragraph a seller read instead of calling you

Go back to the seller in your ZIP code who asked ChatGPT who to call. Read that paragraph the way she read it. It is clean, it is fair, it names two or three people, and there is nobody in it. Both of those are true at once, and telling them apart is the whole job now.

The left column is the part of this business that is about to be worth less. The listing description, the market update email, the neighborhood blurb, the first draft of everything. That is the work you did on autopilot at nine on a Sunday, and losing it is not a small thing, because it was a lot of the hours.

The right column is what stays, and it is the part you would have said was the job if anybody had asked. It is also the part nobody has ever been able to write down, which is exactly why no engine can read it back. Write down one thing you handle that nobody else in your market writes about. It takes an afternoon.

Sources. The two columns describe what these systems are observed to do well and badly on questions about local professionals, and the reading of which column gets cheaper is ours rather than a research finding. Nothing here is a claim about commissions, employment or any individual firm. Inferred.