The Goal Isn't to Be Linked. It's to Be the Answer.

AI search has two currencies: citations and mentions. They look similar, they get conflated in every GEO conversation — and most brands are optimizing for the wrong one.

On February 9 this year, someone in the United States asked ChatGPT a question, read the answer, and noticed a small card underneath it labeled Sponsored. Nobody had seen one before. Two hundred days later, OpenAI said the ad business behind that card was already running at a billion dollars a year. Google's search ads took about four years to get there.

That speed is what everyone is talking about. We think the more useful question is simpler: what does that card actually buy you?

Look closely and the answer is a little surprising. The brand paying for the ad shows up in the answer above it less than 4% of the time. The paid layer and the organic layer of ChatGPT barely touch. Everything else in this post follows from that.

What the ad looks like

OpenAI's example of an ad in ChatGPT: an answer with potluck tips, then a card labeled Sponsored from a grocery brand showing a meal kit with price and cook time.

OpenAI's own example. The answer finishes, then a labeled card from an advertiser. Source: OpenAI.

That is the whole format, at least for now. The answer ends. Beneath it sits a card with the advertiser's name, a Sponsored label, and one or two product tiles: headline, image, price, link. Under that, a line reminding you that ads don't influence the answer and that your chats stay private. An interactive version, where you can ask ChatGPT questions about the ad without leaving the thread, is being tested with a handful of enterprise advertisers.

Who sees it. Logged-in adults on the Free and Go plans. Plus, Pro, Business and Enterprise never do. That still covers nearly all of ChatGPT's roughly 900 million weekly users, which is the point.

How it gets picked. Not by keyword. Advertisers choose intent categories, geographies and, since summer, custom audiences. ChatGPT matches those against the topic of the conversation and, if you allow it, your past chats, and it picks the ad after the answer is written. Advertisers get aggregated views and clicks, never the conversation. No ads appear on health, politics or other sensitive topics, and none for users under 18.

Two phone screens from OpenAI showing the Ads controls page, with toggles for personalizing ads and using past chats, and an Ads history list.

Users can review, delete and switch off ad personalization. Worth knowing before you assume you can target past behavior. Source: OpenAI.

What it costs, and how quickly that changed

The pilot opened for big budgets only. It did not stay that way for long.

DateWhat changed
Feb 9, 2026First US ads; $60 CPM guidance; commitments of $200,000 and up
Apr 13Minimum drops to $50,000; negotiated CPMs reported at $25 to $45
May 5Self-serve Ads Manager opens in the US; no minimum at all
AugUK, Mexico, Brazil, Japan, South Korea, then 31 European countries and India
Aug 31$1 billion run rate; self-serve opens across Europe, India, the Middle East and North Africa; 52 advertiser countries
Sep 10Ads for image and audio generators that compete with OpenAI's own are cut off; Adobe loses its Firefly campaigns without notice

Today a US business can start at $25 a day. Click bids are guided at $3 to $5, and advertisers running their own tests report paying between $1.72 and $2.74. The Ads Manager now has reach, click and conversion objectives, an on-site Pixel and a server-side Conversions API. By August, most of the money was flowing through cost-per-click.

A few doors stay closed: adult content, gambling, alcohol and tobacco, crypto, politics and legal services are prohibited, and finance and healthcare need approval. Turkey, as of the end of August, is not on the available list or the "coming soon" list. We come back to that.

Is it working?

Depends who you ask.

OpenAI's numbers are loud. A billion-dollar run rate, a plan for $2.5 billion this year and $11 billion next, and a story told to investors that ends at $100 billion in 2030. The inventory backing that is real: SE Ranking ran 50,006 commercial prompts in August and found a paid card under one in four of them, from more than a thousand advertisers. The early heavy spenders are the ones you would guess, SaaS companies and comparison sites, HubSpot and Semrush next to Top10 and BestMoney.

The advertisers' side is quieter. Reported click-through rates sit between 0.5% and 1.8%. One Adthena client measured 0.91% where Google search does 6.4% in the same category. One enterprise had managed to spend 3% of a $250,000 budget after several weeks, because ChatGPT couldn't find enough matching conversations. The best public case we found, a 15-day test by Opascope, returned 1.49 times its spend. Encouraging, small, and not yet a pattern.

Then there is trust. In an Ipsos survey this February, 63% of US adults said ads in AI search would make them trust the results less. OpenAI says its own trust metrics haven't moved. Perplexity looked at the same question and walked away from ads in February with one line: "We are in the accuracy business."

Our read: the audience is enormous, the inventory is real, and the craft of buying it is about six months old. Treat every benchmark you see, including the ones in this post, as a first draft.

The number that should organize your budget

Here it is. SE Ranking: the advertiser behind a ChatGPT ad is also cited in the answer above it in 3.63% of placements. Seer Interactive, different sample, different method: the advertiser is mentioned in the answer 7.3% of the time and cited as a source 3.3% of the time.

This isn't a bug OpenAI will patch. It is how the thing is built.

Diagram: the answer is written first from the model's sources; the ad is matched afterwards to the conversation topic. The two steps share no decision.

The ad is placed after the answer is written. You can pay for the box beneath the verdict. You cannot pay for the verdict.

Three things follow, and each is a line in your budget.

Buying the slot doesn't buy the recommendation. A brand missing from the answer can pay to sit under it. The reader sees "most people land on A, B or C," then a card for D. That's not nothing; it's how comparison sites are already spending. But it is a different product from being C.

Being recommended doesn't protect the slot. Seer's sharpest observation was that most brands haven't claimed their own branded prompts. When someone asks ChatGPT about you by name and the answer is glowing, a competitor can still buy the card underneath. Conquesting, the oldest trick in paid search, is back, and the inventory is cheap because so few brands have noticed.

One company controls both layers. On September 10, OpenAI stopped selling ads for image and audio tools, days after launching its own. Adobe's search lead described ChatGPT as "narrowing both the paid and organic discovery pathways" in categories it now competes in. If your category is one OpenAI might enter, your paid plan and your organic plan share a single point of failure.

Where the other engines stand

EnginePaid layer, September 2026
ChatGPTAds on Free and Go in 52 countries; self-serve; CPM, CPC and conversion bidding
Google AI Overviews and AI ModeAds in roughly a quarter of AI results, up from about 5% in early 2025; offers and checkout inside AI Mode
Microsoft CopilotExisting Microsoft Advertising campaigns can appear inside Copilot answers
PerplexityDropped ads in February 2026; subscriptions and shopping instead
Gemini app, ClaudeNo ads

Google folds the ad into the answer surface; ChatGPT keeps it outside and says so loudly. For a brand, both roads lead to the same place: the ad and the recommendation are separate assets, made by separate systems, measured separately.

A note on Turkey

In Turkey, ChatGPT carries 94.49% of AI referral traffic, the highest single-engine share anywhere. ChatGPT Go is available there. Ads are not, and Turkey isn't on the announced roadmap.

That combination is unusual. When the switch flips, almost the entire Turkish AI-search audience becomes ad-eligible on the same day, and the first advertisers in the auction will be the ones who already know which prompts they are missing from. Until then, Turkish brands have exactly one lever inside ChatGPT: the organic answer. It is worth pulling while it is the only one that works.

What to do

That first step is the one most teams skip, because until now nobody sold them a reason to look. The Lens measures the organic layer prompt by prompt, engine by engine, against named competitors: which answers include you, which don't, and who took your place. That's the map that tells a media team where an ad is defensive, where it is wasted, and where it is the only option. The PR Factory builds the source material that moves the answer itself, the part no auction sells. And The Shield watches sentiment in the answer layer so you hear about a bad verdict before you pay for the box beneath it.

ChatGPT now sells the slot. The answer is still earned. Measure the second before you pay for the first.

Sources