How ChatGPT ads work, and what they cost

A ChatGPT ad is one card under the answer. It shows to adults on the Free and Go plans and it is matched to the conversation, not a keyword. The floor is $25 a day. OpenAI suggests a $3 to $5 bid. Insurance advertisers are approved one at a time, in the US only.

Most of what you will read about ChatGPT ads was written for a platform that no longer exists. Free users only, no minimum, no conversion bidding. All of that changed by summer.

Get the mechanics wrong, and you set a $100 budget you did not mean to set, pick an objective that cannot learn from a quote, and let AI rewrite copy your compliance officer signed. Here’s how it actually works today, checked in Ads Manager and against OpenAI’s own documentation.

Who sees ChatGPT ads, and where?

Adults 18 and over on the Free and Go plans. Signed in or not.

OpenAI’s ads team said both in its October webinar, and the second part matters for insurance: a signed-out user on a phone is a real renter or driver, and they count. Plus, Pro, Business, Enterprise and Edu users see nothing. OpenAI says ads run in 89 countries. For insurance, the policy limits you to the US.

Signed-out users get ads suitable for every audience. Mature categories only reach signed-in adults. Insurance is fine either way.

The ad is one card under the answer, introduced by a line the platform writes (“This might also interest you”). The card: an image panel, the advertiser’s name, a grey Ad label, a headline, a description. No web address. OpenAI’s help center says one ad unit per response, and that unit may carry more than one image. OpenAI is also testing placements with more than one advertiser, priced by a second-price auction, which Vital Design reports from inside the platform. Expect competition on good prompts to rise.

Ads never show in conversations OpenAI classes as sensitive: health, mental health, politics, personal data, anything with a minor. The Brookings audit of 3,600 ads found no case where that guardrail failed.

How does matching work when there are no keywords?

You write context hints. Short descriptions of the offer, who it helps, and the situation they are in. OpenAI’s support team told Symphonic Digital that hints “guide matching but aren’t exact-match keywords and don’t guarantee delivery.”

ChatGPT then decides whether your ad fits. OpenAI’s help center lists the signals: the context and intent of the current conversation, your landing page, your title and copy, your context hints and targeting, and, when the person has ad personalisation on, signals from their wider ChatGPT use, including past chats, memory and how they treated earlier ads. When several ads are eligible, ranking weighs relevance and bid.

That last part is the piece most guides miss. In the webinar, OpenAI’s example was a dealership ad for the exact minivan trim someone researched days earlier, shown while they planned a weekend trip. For an insurer: the person who asked whether their policy covers a stolen bike last Tuesday can see your ad on Saturday, in a conversation about Asheville.

The match can also be loose. Brookings’ researchers got a DoorDash ad on a pad Thai recipe, and ads on “give me tips for work-life balance.” In our insurance sample, 312 of 1,279 ad-triggering prompts had no insurance word in them. Liberty Mutual’s “Switch and Save” on a phone-deals prompt. GEICO on a question about Valvoline oil. Lemonade’s pet ads on dog-vet questions.

Loose is not bad. Loose is where the life moments live.

What does it cost?

$25 a day, minimum. We typed $10 into our own Ads Manager and got “Enter a daily budget of at least 25.00.” The form pre-fills $100. Change it.

Spend can reach twice your daily budget on a given day and seven times it across seven days. A $25 budget can bill $50 on a Tuesday. Tell finance before the first invoice.

OpenAI suggests a maximum bid of $3 to $5 per click to start. What people actually paid:

Billing is postpaid. Your card is charged when unpaid spend reaches a threshold that starts low and rises, or at month end. The currency and country you pick at setup cannot be changed. New advertisers are often offered free ad money once they start spending. We saw spend $500, get $500 back. The amount varies by account, so check yours.

Which objective should an insurer pick?

There are four. Reach pays per thousand impressions. Clicks pays per click. Conversions comes in two flavours: billed on clicks, which learns only from people who clicked, and billed on impressions, which also learns from people who saw the ad, did not click, and converted within a day.

Why the last row matters: OpenAI says 52% of eligible one-day view-through conversions happen within an hour of the impression. A quote request is the kind of thing people finish on a laptop after reading on a phone. Give the system credit for it, or it cannot learn from it.

For bidding, you get Maximize results (automatic) or a manual maximum. OpenAI recommends automatic. Here’s what I do on a brand-new account: a manual cap for the first three days, so a bad hint cannot spend $50 on nothing. Then automatic once the first conversions land.

What targeting exists?

• Location. Country, and in the US, state, DMA (a TV market) and ZIP code. Exclusions exist. For an insurer this is the setting that does the compliance work. Target the states you are licensed in. Exclude the rest.
• Platform. iOS app, Android app, iOS web, Android web, desktop web. Leave all five on for a test.
• Custom audiences. Your own list of emails or phone numbers, hashed. Inclusion needs 25,000 matched people, and OpenAI recommends 100,000. Exclusion has no minimum, so a policyholder list can keep your acquisition ad away from existing customers from day one. Not available in the EEA.
• Context hints. The main lever. Free text per ad group. Chapter 4 covers them.
• Negative Phrases. Announced October 5 for qualifying advertisers with brand-safety needs. Not a performance tool. Not something a new account should plan around.

No interest targeting. No behaviour targeting of the Meta kind. And you never see the conversation. Search Engine Journal’s line is exact: “Advertisers don’t receive users’ conversations or chat histories.” You see impressions, clicks and conversions by ad, ad group, device and country. That’s the view.

What rules apply to insurance?

OpenAI’s ad policy puts financial services, insurance included, in the US-only, approved-advertiser category. Proof of licensure may be required. Health insurance is restricted to approved US advertisers. Credit repair, debt settlement and legal services are prohibited outright. Guides quoting OpenAI’s help text put business verification at up to seven business days. Ad review runs a day or two.

Two settings deserve a compliance officer’s eye.

The first is creative enhancement. It is on by default, and it lets AI rewrite your headline and description to fit the conversation. OpenAI’s own wording: rewritten copy “may be shown without your individual review or approval.” For state-filed language, turn it off.

The second is location. The Brookings audit found that accounts in lower-income ZIP codes were significantly more likely to receive ads, with no evidence that race played a part. That is the platform’s delivery pattern, not yours. But if you hand-pick ZIP codes, you own the choice. Target by licensed state and let the rest be.

FAQs

Can I see which prompts triggered my ad?

No. You get performance by ad, ad group, device and country. The libraries in Chapter 3 are the only public window into prompts, and they are built with researcher-written questions.

Is there a carousel or video format?

Not for insurance. OpenAI says one ad unit per response that may carry one or several images, and it is testing a multi-image ad during image generation. Product-feed carousels exist for retailers with a catalogue. Text, image and the landing page are what you have, and OpenAI reviews all three together.

Can a person reply to my ad?

Yes. OpenAI’s help center says a user can message an advertiser through an ad, and the advertiser sees only those messages. The user can also hide the ad, report it, read “About this ad,” or ask ChatGPT about it. Write the card as something you would be happy to answer questions about.

Can my agency run this from its own account?

OpenAI wants agencies added as users on the advertiser’s own account, not creating accounts for clients. In practice: invite as Member, then upgrade to Admin. The direct Admin invite failed in one agency’s test.

Do EEA rules change anything?

Personalisation is off in the EEA and Switzerland, so custom audiences are unavailable there. Insurance is US only under the policy, so this only matters if you also sell elsewhere.