ChatGPT Ads Reach Everyone Except Paying Users

Ads run on Free, Go and logged-out sessions. Plus and Pro are ad-free. Four questions to answer before you fund the channel.

ChatGPT ads run on Free, Go and logged-out sessions. Plus, Pro, Business, Enterprise and Edu are ad-free.

That single line decides whether this channel can reach your buyer, and it’s the last thing agency decks mention.

Four questions to answer before you fund it, the setting you should switch off today, and the four OpenAI bots most teams block by mistake.

Last week OpenAI emailed me: ChatGPT Ads Manager is now open to advertisers in the Middle East, including Turkey, where I work. A few days earlier it had been the first item in a client meeting, and I listed my reservations then. Since then the industry tone has moved to “great new channel, get in early.”

The channel is real. But four things sit at the bottom of every pitch deck.

Grab a coffee.

The thesis

ChatGPT ads bring a third model to advertising. Google showed us what people search for. Meta showed us what they’re interested in. ChatGPT shows what they’re trying to get done right now. That’s a powerful signal.

But before you fund it, answer four questions: who sees the ad (answer: people who don’t pay for ChatGPT), how the ad gets chosen (context, not keywords), which data claim is true (neither of the two you’ll hear, and there’s a setting you should switch off), and what the technical gate is (four separate bots, and most teams block all of them at once).

Go in without these and your first month’s budget is tuition.

1. Who actually sees the ad?

The most important sentence of the week is the least repeated one. According to OpenAI’s own help centre:

That has two consequences.

Media planning. If you sell to businesses, your decision-maker probably has a Business seat or pays for Plus. Either way, they will never see your ad. The people who pay most for AI sit outside this inventory. The channel fits consumer offers and low-to-mid ticket purchases far better.

Ethics. When ad-free becomes a product feature, the ad load falls on people who can’t pay. That’s been the ad model since television, but advertisers should decide knowing it.

This is where the non-US view matters. The Free tier carries far more weight in emerging markets like Turkey, so the income skew of this inventory is sharper outside America. I haven’t seen one US-based analysis look at it from this angle.

The first question isn’t “what’s my budget.” It’s “which plan is my buyer on?” If you don’t know, start asking in your client calls.

2. How does the ad get chosen?

The same need reads very differently in Google and in ChatGPT:

No keywords, yet the need is completely clear. That’s the channel’s potential.

OpenAI picks ads using the conversation, your landing page and copy, and signals it calls context hints.

A context hint is not a keyword list. It will be the most common mistake early advertisers make. Take the second example:

Wrong: pet insurance, dog insurance, vet costs, cat insurance

Right: “Owners of ageing dogs or cats who are worried about unexpected vet costs and want to secure treatment expenses in advance.”

It should answer three questions: what do you offer, who is it useful for, and in what situation.

Critically, a context hint is not a targeting rule. You can’t guarantee appearing whenever a given sentence is typed. For teams used to exact match, that’s the hardest shift.

Success here comes less from writing ad copy and more from defining the problem. Ask your sales team first, not your agency: what sentence does a customer say before they come to us?

3. Which data claim is true?

You’ll hear two sentences. Both are wrong.

“Your ChatGPT conversations are being sold to advertisers.” No. Conversations aren’t shared with advertisers and data isn’t sold; advertisers see aggregated reporting. The ad system is also separate from the system that writes answers. You can’t pay to change ChatGPT’s answer or make it recommend your product. Ads sit below the answer, clearly labelled.

“The ads don’t use your data at all.” Also wrong. Selection is built on the context of the current conversation. With personalisation on, broader signals can be used too.

The truth in between: the ad system and the data pipeline are separate systems, but they live inside the same box.

Who’s inside the box

Barış Özcan, one of Turkey’s best-known tech communicators, just published a video that isn’t about ads but explains OpenAI’s approach well. It builds on 404 Media’s September report: some ChatGPT conversations are read and scored by human contractors.

Your username isn’t shown and personal information is filtered, but the company admits the filter can miss some of it. Some chats carry a memory summary on top: what you’ve used ChatGPT for, sometimes where you live. Nothing explains how conversations are selected.

His closing line puts it well: typing “keep this between us” at the end of a chat changes nothing. Switching off the setting does.

Three separate switches, three separate decisions

• Settings → Data controls → “Improve the model for everyone”. It’s on by default, on Free and on paid Plus and Pro alike. Turn it off and new conversations stop going into the training pool, but nothing changes retroactively. On business accounts and the API it already comes switched off.
• Temporary chat. Goes into neither memory nor training. If you’re writing something sensitive, use this.
• Ad personalisation. Users can see why an ad was shown and turn personalisation off entirely. This switch is independent of the other two.

On deletion: a deleted chat leaves the system within 30 days, but if it had already entered the training pool, it stays there. And this isn’t only OpenAI. Google does the same human review and says so in Gemini’s privacy notice. The difference is that the information sits in a help page, not on the screen you use every day.

The advertiser’s responsibility. OpenAI’s rule against ads near mental health and politics is a good rule. But it’s about where the ad goes, not what happens to the conversation. Your brand is stepping into an interface where people talk at their most vulnerable. Not everything that’s technically allowed is worth doing.

4. The technical gate: four bots, four decisions

Most teams treat this as one decision: “we blocked the AI bots.” It’s four.

The core distinction: to appear in LLM answers, keep the access bots open. If you don’t want your content training the model, block the training bot. They aren’t alternatives. robots.txt manages them independently:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: OAI-AdsBot
Allow: /

The common trap: a wildcard. A broad Disallow under User-agent: *, or Cloudflare’s “block AI bots” toggle, closes all four at once. As we saw last week, Cloudflare’s Bot Preference Sync only offers three categories (Search, Agent, Training) and no per-bot rules.

On OAI-AdsBot: it only visits pages submitted as ads, its data isn’t used for training, and there’s no published IP list, so you can’t verify a visit in your logs is genuine.

Blocking has limits. Blocking GPTBot today doesn’t recall data already collected, and your content can still reach a model indirectly through third-party sites. robots.txt isn’t a lock, it’s a declaration. Don’t promise a client “our data is protected.” Say “we’ve declared that we refuse crawling for training.” That’s the honest version.

From the field: what I told a client

Not industry data. A recorded conversation. Last week, the first item in a weekly client meeting was ChatGPT ads: “What do we know, should we test it?” My answers:

• “I don’t know exactly how it works.” I said it twice and I’m in the same place today. A channel opening isn’t the same as knowing how it behaves.
• “That’s not really us.” Inserting a brand into a conversation the user is having may not fit that brand’s story. Not every channel suits every brand.
• “Agencies are trying to be the pioneers, as with every new trend.” With no one holding six months of campaign history, there are no experts yet, only early testers.
• “It’s on free accounts now, and there’s no guarantee it won’t come to paid ones.” This is my real reservation. A media plan built on today’s plan split breaks the day OpenAI changes that split.

The most striking line from the client side wasn’t about ads: “I’m wary of even uploading a photo of myself or my family.” The real feeling at the table was trust. No dashboard measures it, but it will set this channel’s ceiling in markets like mine.

One meeting, no generalisation, no money spent yet, so no performance claims. These are the questions asked before money is spent.

The one thing to do this week

Don’t open an account. Fill in four boxes. It takes half an hour.

1. Which plan is my buyer on? Ask your last ten clients: paid or free ChatGPT? If it’s mostly paid, this channel doesn’t reach your decision-maker.
2. What’s my problem sentence? Describe the situation where your product becomes useful, in one paragraph, without keywords. If you can’t, you’re not ready for the channel.
3. Are the four bots listed separately in my robots.txt? If they’re blocked with a wildcard, fix it with three lines. Compare your Cloudflare policy against robots.txt; if they contradict, you don’t know which one wins.
4. Where’s my ethical line? Write down the conversation contexts where you don’t want your brand to appear: health, debt, family crisis, job loss. Today it’s a preference. Tomorrow it’s your brand safety policy.

And one personal, ten-minute job: in your own account, go to Settings → Data controls → “Improve the model for everyone” and turn it off. Tell your team and your family too. When they write something sensitive, they should use temporary chat.

If you test, use one problem sentence, one landing page and a trial budget. In a beta system, month one is learning.