There's an ad slot inside the answer now.

For twenty years, the game was Google rank. Get your listing, your name, your site as high on the page as possible.

Here's what changed. When a consumer asks ChatGPT "what should I know about buying a first home in Arizona," they don't get ten blue links. They get an answer. One answer. And underneath that answer, there can now be a paid ad.

OpenAI opened its Ads Manager, and the invitation landed in my inbox this week. Real estate is one of the categories you can select. So I built a full campaign for my brokerage — targeting, creative, budget, all the way to the publish button.

I haven't pressed it. I'll tell you exactly why, because the reasons are more useful than the campaign.

First, understand what this is, because it's genuinely different from every ad platform you've used. The ad appears below ChatGPT's response, clearly labeled and visually separated from the answer. OpenAI says ads run on separate systems and don't influence what ChatGPT tells you, and that advertisers never receive your chats, memories, or personal details — only aggregated performance data. Ads don't appear in Temporary Chats at all, and they aren't shown in contexts OpenAI considers sensitive.

The minimum spend is $25 a day. So $750 a month to find out.

This part will feel wrong to anyone who has run Facebook or Google ads.

There are exactly two targeting controls.

The first is geography — at the country level only. Not state. Not city. Not ZIP. My campaign says "United States." For a Phoenix broker, that means paying to show ads to people in Vermont.

The second is what OpenAI calls context hints, and it's the whole ballgame. You don't pick keywords or demographics. You write, in plain English, a description of the conversations where your ad should appear. OpenAI's own definition: hints that describe the conversations, topics, or keywords where your services may be relevant — explicitly not exact-match keywords, and with no guarantee of delivery in any particular conversation.

So instead of bidding on "phoenix realtor," you write something like: the user is asking whether they can afford to stop renting and buy in Arizona. Then a model decides, live, whether your ad fits the conversation happening in front of it.

That's it. A paragraph and a country. Which means localizing isn't a setting — it's a sentence you write and hope the matching engine honors.

You remember what happened to Facebook. Advertisers could exclude people from seeing housing ads by age, by gender, by ZIP code. HUD came after them, and Meta ended up having to build an entirely separate, stripped-down flow for housing ads with those controls removed.

So the first question I walked in with: does this platform have the same hole? The answer has three layers.

Layer one: the built-in demographics are gone. No age setting. No gender. No income. No ZIP. Not restricted for housing — simply nonexistent, for every advertiser in every category. You cannot discriminate through demographic targeting here because there is nothing to discriminate with. Structurally, that's cleaner than Meta ever was.

Layer two: the targeting field is now a sentence you write yourself. That's what context hints are, and that's where the risk went. The old fair housing problem was a checkbox somebody shouldn't have checked. The new one is a phrase somebody shouldn't have typed. If an agent writes a hint describing who they want — a family type, an age group, a neighborhood standing in for something else — that's the same violation wearing a different outfit, sitting in a free-text box with no warning label anywhere near it.

My rule, and I'd hold it hard: describe the question, never the person. "Someone asking whether they can afford a first home in Arizona" is a question. "Young families" is a person. One targets a conversation. The other targets a class.

Layer three: you can upload your own list. There's an Audiences tool that takes a CSV of emails or phone numbers to use as a filter on your campaigns. Read that carefully, because it means you are now deciding who does and doesn't see a housing ad. How you built that list becomes the entire question. "Everyone who's closed with me" is fine. A list assembled by neighborhood, by surname, by language, or by anything else standing in for a protected class is the Facebook problem with your hand on the pen instead of a targeting menu.

I could not find an answer to the most important version of that question: whether you can exclude an audience rather than only include one. Exclusion is precisely what HUD went after Facebook for — not who you reached, but who you deliberately kept from seeing the ad. The interface calls them "audience filters," which could mean either. If you run this, find out before you upload anything.

And here's what should focus your attention. OpenAI's FAQ says ads are not shown in sensitive or regulated contexts — it names personal health, mental health, and politics. Housing is not on that list. They knew to carve out regulated categories. They carved out three. Real estate wasn't one of them. There is no guardrail here — not a weak one, none. Every bit of fair housing compliance on this platform sits on what you type in that box and what you upload in that file.

One more thing that isn't a fair housing issue at all. Those emails and phone numbers in your database were given to you by people who wanted to buy or sell a house — not to be uploaded into an AI advertising system. OpenAI says advertisers receive only aggregated data and never see personal details, and the upload supports hashing. The decision is still yours to justify, and it's worth making on purpose rather than because there was a button.

Here's the part that genuinely surprised me, and it's sitting in the Law Book right now.

A.A.C. R4-28-502(J) says that the use of an electronic medium — and it names the Internet, Artificial Intelligence, and web site technology by name — that targets residents of this State with the offering of a property interest or real estate brokerage services constitutes the dissemination of advertising.

Arizona wrote artificial intelligence into its advertising rule. That means a ChatGPT ad aimed at Arizona consumers isn't some novel gray area. It is advertising, governed exactly like a yard sign or a postcard.

Which brings two requirements with it. Subsection (E) requires that all advertising identify, in a clear and prominent manner, the employing broker's legal name or the dba on the broker's license certificate. And subsection (J) adds that when you advertise online, all of those requirements must be satisfied on the advertisement itself, without the need to scroll.

Now look at the ad unit. It's a small card: advertiser name, a logo, a headline, one line of copy, an image. In my own preview, the advertiser name displayed truncated with an ellipsis — and what showed was my ad account name, not my brokerage's name.

So before anyone in Arizona funds one of these, that's the question to answer: does your employing broker's legal name appear, clearly and prominently, on the ad itself, without scrolling? If it doesn't, you have an advertising problem regardless of how the campaign performs. And remember subsection (G) — the designated broker is responsible for the advertising of all real estate activity. That means your broker, not the platform.

I don't have a clean answer yet for how to satisfy that inside a card this small. That's exactly why I haven't published mine.

I built the whole campaign and stopped at the publish button. Here's what I found that isn't in the marketing.

None of that makes the platform bad. It makes it early. And early is expensive when you're paying $750 a month to learn.

If you want to look at it yourself, it's at ads.openai.com. Building a campaign costs nothing — you only pay if you publish. Nobody pays me if you click that, and there's no referral code on it.

Have you tried advertising anywhere new this year — ChatGPT or anything else outside the usual portals? I want to know what you spent and whether anything came of it. Hit reply. I read every response, and I'll report back if I fund this one.