Write For The Conversation

I've bought ads on just about every screen people stare at. Search results, feeds, inboxes, that little box next to the weather forecast. Every one of them trained us to think in keywords and audiences.

ChatGPT ads don't care much about either one.

On October 21, OpenAI expanded ChatGPT ads to 26 more countries, which brings the total to 89. The new list includes Argentina, Chile, Colombia, Peru, Nigeria, South Africa, Serbia, Ukraine, Cambodia, Bangladesh and Pakistan. More than 100,000 advertisers have run campaigns since launch, and the ad business hit an annualized $1 billion run rate in August, up from $100 million in March.

That's real money moving, which means your clients (and your brother in law who "does marketing") are about to ask whether they should test this.

Maybe. But build it like a Google campaign and you'll light the budget on fire and blame the platform. The ad shows up in the middle of somebody's conversation, so you have to write for that conversation, and most of what follows is about how.

Start with the audience. It kills bad ideas early.

Ads only show on the Free and Go plans. Plus, Pro and Enterprise users never see them. So the paying power users you picture when someone says "ChatGPT user" are partly off the table.

Earlier rules kept ads away from users under 18 (or predicted to be minors) and away from sensitive topics like health, mental health and politics. Political advertising isn't allowed at all. Run ads for a chiropractor, a therapist or a candidate? Park this channel for now. You aren't the customer they're building for yet.

OpenAI also says conversations stay private from advertisers, data isn't sold, ads are labeled and kept separate, and advertising doesn't change ChatGPT's answers. They published a technical overview saying the ad system runs apart from answer generation. Translation for your client: paying for ads won't get you recommended in ChatGPT's organic answers. That's a separate fight. Anyone selling you otherwise is selling you something.

You buy through technology partners or self serve in Ads Manager, which is still in beta.

OpenAI's launch guide lays out three levels. If you've built in Meta, this will feel familiar for about ten seconds.

Campaign. Name, budget, start and end dates, objective (Views or Clicks) and country.

Ad group. Name, max bid and your context hints. The bid is a max CPM if you want reach or a max CPC if you want clicks. CPC bidding showed up alongside CPM on May 5.

Ad. Brand name, logo, title, copy, landing page URL and an image.

There's a bulk upload template, and every ad gets a policy review before it goes live.

One setup trap before you touch anything. Lilach Bullock's walkthrough points out that country, currency and time zone can't be changed after you create the account, and changing the brand or legal name later kicks off a review. Managing clients? Get these right on day one or plan on building a second account.

Now look at what's missing from that ad group level. There's no keyword list, no interest stacking and no lookalike option, just a box for hints. Most of the money in this channel gets made or wasted in that box.

Forget keywords for a minute. A hint describes a moment in a conversation where your offer would be the obvious next step.

The context hints guide puts it plainly: hints are plain language descriptions of conversations and situations, not exact match keywords, and they don't guarantee delivery. The system scores relevance across your hints, title, copy and landing page together. Then a relevance weighted second price auction ranks the ads, which means a sharp hint can beat a fat bid.

Every direct response person should sit up at that. It's one of the rare auctions where the better writer gets a discount.

My formula is who + situation + need + constraint.

Weak hint: "HVAC repair."

Better hint: "Homeowner whose AC quit during a heat wave and needs someone out today without paying emergency rates."

The first one is a category. The second is a person in a specific bind with a reason to click. Write that kind.

A few rules I follow, most of them lifted straight from the guide:

Five to fifteen hints per ad group. One idea per hint. Start with about five and give them room to work.

One buyer per ad group. Don't mix the homeowner with the property manager. They worry about different things and they'll need their own page anyway.

No locations or demographics in hints. Writing "show only to people in Chicago" won't restrict anything. Use the actual targeting controls for that.

Write the situation instead of the category. "Small business owner behind on six months of books before tax time" beats "bookkeeping services".

Use the buyer's words. If customers say "my books are a disaster," then your hint says "books are a disaster." Nobody has ever typed "financial reconciliation needs" into a chat at 11 p.m.

Tighten your wording, because there are no negative keywords. None. A loose hint puts you next to the wrong conversations and you can't exclude them. Sharper language up front is the only fix you get.

You already have the raw material. It's sitting in your call recordings and your inbox, and every prospect who ever said "okay, here's what's going on" handed you a context hint.

If you record calls with Fathom, export the transcripts from your last 50 sales calls. No transcripts? Paste in 50 customer emails or form submissions instead. Drop them into ChatGPT or Claude with this:

Read every hint out loud. If it sounds like a brochure, cut it. Keep anything that sounds like a stressed customer on a Tuesday.

The limits are tight. Good.

Title: 16 to 24 characters recommended, 50 max.

Copy: 32 to 48 characters recommended, 100 max.

Image: PNG or JPG, square, no bigger than 1200 x 1200.

That's barely a sentence, and it sits right next to an answer somebody just asked for. A billboard yelling over the conversation won't work here. What works is something that reads like the obvious next helpful step.

Say somebody asks ChatGPT how to tell if their AC compressor is dying. The answer walks them through the signs. Your ad underneath says "Same Day AC Repair" with "Flat diagnostic fee. Tech at your door today." That picks up right where the answer left off. "BEST HVAC IN TOWN! CALL NOW!" is a guy with a megaphone at a dinner party.

Write the title like a label for the fix. For the copy, find the one detail that removes the hesitation, whether that's price clarity, speed or a guarantee, and use only that one.

Models are bad at counting characters, so make it show its work.

Count them yourself anyway. I've watched models swear a 31 character line is 24.

And use the offer that's actually live on the landing page. If the page says "free estimate" and the ad says "flat diagnostic fee," you've broken relevance and annoyed a prospect in one move.

Relevance gets scored across hints, title, copy and landing page together. The page counts as much as anything you typed into Ads Manager.

Say your hint describes a homeowner with a dead AC in a heat wave, and the click lands on a homepage with a slider about your 40 years of family history. You lose that one. The first screen needs to pick up the exact situation the hint described.

Two technical checks. The page has to be reachable, and it can't block OpenAI's crawlers. Have your developer look at the robots file and any bot protection on the site. Plenty of security plugins block crawlers by default, and nobody notices until the ad gets rejected.

Send the lead straight into Go High Level so the follow up starts the minute the form comes in. A click from mid chat is warm for about as long as the chat lasts.

Picture a six person bookkeeping firm that works with small service businesses. They've run Google Search for years and they want to see if ChatGPT ads can add leads without cannibalizing what works.

Ad group one: the catch up client. Hints like "Contractor who hasn't touched their books in months and their accountant needs clean numbers before the tax deadline" and "Owner who just found out their bookkeeper quit and left a mess in QuickBooks." Title: "Catch Up Bookkeeping." Copy: "Months behind? We clean it up before tax time." Landing page: a catch up offer, a flat project quote, a short form.

Ad group two: the outgrown DIY owner. Hints like "Service business owner doing their own books at night and starting to make mistakes as revenue grows" and "Owner who spends weekends on invoices and wants that time back." Title: "Get Your Weekends Back." Copy: "Monthly bookkeeping for service businesses." Landing page: monthly plans, what's included, a call booking.

Ad group three: the lender prep client. Hints like "Business owner applying for a loan who needs financial statements a bank will accept." Title: "Lender Ready Financials." Copy: "Clean statements your bank will actually accept." Landing page: a page that says exactly that, with a timeline.

Each buyer gets a group and a page built for them. Nobody gets dumped on the homepage.

Same idea for a regional HVAC company. Emergency repair, replacement shoppers and maintenance plan buyers would each get their own group, their own hints and their own page.

The US pilot launched at a $60 CPM. An August table listed daily minimums starting at $25. Neither number makes this a cheap way to buy eyeballs, and for a lot of small budgets Meta or Google will still be the better spend until this channel proves itself in your own numbers. I'd only carve off test money you can afford to learn with.

A sane starter plan: one campaign, two or three ad groups with about five hints each, and a daily budget at or a bit above the minimum. Run it two full weeks before you decide anything.

Know how the reporting behaves before you panic. Impressions, clicks and CTR refresh about every 15 minutes, but spend can lag 7 to 8 hours, so the dashboard at 3 p.m. isn't telling you what you spent today. In the US you get pixel tracking and a Conversions API. Set up both before launch or you'll be judging the channel on clicks alone.

Then test like an adult:

OpenAI is also testing Sponsored Agents (a labeled chat with an advertiser's own AI agent) with some US advertisers. Interesting. I wouldn't budget around it yet.

Decide if you even qualify. If you sell anything in health, mental health or politics, skip this channel for now.

Export transcripts from your last 50 sales calls and run prompt one to get your first set of hints.

Pick one buyer and write one ad group with five hints. Just one.

Run prompt two, then count every character yourself.

Audit the landing page with prompt three and make sure your site doesn't block OpenAI's crawlers.

Set up the pixel and Conversions API before you spend a dollar, then leave it alone for a week.

Jordan

The AI Newsroom | Practical AI for people with a business to run.