You can buy the slot, you cannot buy the sentence

OpenAI placed the sponsored unit below the answer, on a system advertisers are told they cannot touch. That single architectural choice makes AI paid media a downstream decision, and the only defensible way to spend is to first know where you already win organically, where competitors own the answer, and where high-value ground is still unclaimed.

What actually shipped

The timeline is short and it is worth being precise about it, because most of the (linkedin) commentary is not.

OpenAI published its approach to advertising on 16 January 2026 and began testing ads in ChatGPT on 9 February, with logged-in adults in the United States on the Free and Go tiers. Canada, Australia and New Zealand followed on 26 March; the UK, Mexico, Brazil, Japan and South Korea on 7 May, moving to full launch in those markets on 11 August. On 31 August, OpenAI announced self-service access across India, Europe, the Middle East and North Africa, bringing the platform past 40 countries, and disclosed that ChatGPT Ads had reached a $1 billion annualised revenue run rate in under 200 days.

The stack is no longer experimental either: CPC and outcome-optimised bidding, product feeds, platform and geographic targeting, custom audiences, a pixel and a Conversions API, tens of thousands of advertisers and more than 50 technology and measurement partners.

The reach, however, is capped by design. Ads appear only for Free and Go users; Plus, Pro, Business, Enterprise and Edu accounts see none. Under-18 users are excluded. Sensitive categories, including health, mental health and politics, are excluded. Users can dismiss an ad, ask why they are seeing it, turn off personalisation, clear their ads data, or pay their way out entirely.

So: a real channel, with a real media stack, and a structurally bounded audience. Which is exactly why the question is not whether to buy. It is where, and why.

The architecture that decides everything

Ads sit below the end of a response, clearly labelled as sponsored and visually separated from it. OpenAI is explicit about what runs where: ads “run on separate systems from our chat model, and advertisers have no ability to shape, rank, or alter ChatGPT’s responses.”

Take that literally for a moment, because it is the whole strategic story. The sentence above your ad is produced by two machines you cannot bid on: parametric memory, which is what the model absorbed about your category during training, and live retrieval, the pages it fetches, reads and cites at answer time. Your budget reaches neither. It reaches the block underneath.

Fig. 1. The paid unit is downstream of the recommendation. Budget moves the bottom block; only how the model represents you moves the top one.

In paid search, an ad sits above ten blue links the user still has to judge for themselves. In an answer engine, it sits below a verdict that has already been delivered, with reasons and often with sources. That is not a worse placement. It is a different job, and it fails when you treat it like the old one.

Why “we’ll just buy the placement” is the wrong first move

Three reasons, in ascending order of how much they cost you.

First, the decision is usually older than the prompt. 6sense finds buyers are roughly 70% through their journey and have 85% of their requirements set before they contact a vendor. Forrester finds 92% of buyers start with a vendor already in mind and 41% with a single preferred one. The AI answer is frequently confirmation, not selection, and your sponsored card arrives after the confirming sentence has been written.

Second, part of your category is unreachable by paid, permanently. Enterprise seats, the buyers you most want in B2B, are on ad-free tiers. Whole categories are excluded. For those audiences there is no slot to buy at any price. There is only how the model describes you.

Third, the numbers will flatter you. OpenAI cites an ecommerce advertiser at 3× return on ad spend over 28 days, and a technology partner reporting that more than 80% of ad-driven ChatGPT traffic came from new customers. Both are plausible, and both are measuring a click that a pre-qualifying answer handed over. This failure mode is not new. In Refine Labs’ hybrid-attribution work (620 conversions and $21.5M closed-won over twelve months), 53% of buyers self-reported a podcast as the reason they showed up, against 0% credited by the software; a measurement gap of roughly 90%. Same physics, new channel: paid gets credited for a decision the organic answer made.

You are not buying attention. You are buying a line underneath someone else’s verdict.

Three questions before the first bid

Every prompt that matters in your category can be placed on two axes: how present you are in the answer, and how present your competitors are. Four squares, four different jobs, and only two of them are a bid.

Fig. 2. Position prompts by measured presence in AI answers, not by search volume. The square, not the budget, decides the move.

1. Where are we already the answer? Start here, because it is the cheapest thing you own. Measure mention rate, position in the answer, whether the model can justify choosing you in one sentence, and which sources it cites when it does. In AYAN’s work with CALO, the brand became the default AI recommendation, at 83% recommendation-default on transactional intent, with AI-referred customers converting 4× higher. Bidding on that ground mostly buys a click you were already owed. The work there is defensive: keep the cited sources accurate and current, because defaults erode quietly.

2. Where do competitors lead? This is where paid does its one genuinely irreplaceable job: interrupting a default the answer just handed to someone else. Bid, but be honest about what the bid is: a bridge, not a fix. The fix is the evidence layer. A KDD ’24 study across roughly 10,000 queries found that adding quotations, statistics and citations to source content lifted visibility in AI answers by about 40%. If the model repeats a specific claim about a rival, find the page it is reading and supply the comparable.

3. Where is nobody positioned, and the value is high? The most under-priced ground on the map. These are high-intent prompts the model answers generically, with no brand strong enough to anchor. Cheapest inventory, no incumbent default to fight, and the fastest organic capture available in the category. Judge these on the decision the prompt precedes, not on volume. Volume is a search-era metric and it will point you at the wrong prompts.

And the fourth square, contested ground, where you and a rival both appear and the model hedges: the margin there is explainability. Whichever brand the model can justify in one clean sentence wins the hedge. With L’Occitane Middle East, ranking first on explainability in beauty AI queries came with an 8.6% conversion uplift from frontliners: the model’s ability to say why, converted into revenue.

Fig. 3. The same map, read as a budget decision. Each row splits what a bid can buy from what only representation can fix.

What this changes inside the company

The practical shift is small and it is organisational: ad budget becomes an output of the prompt map, not an input to it. You review the map before the plan, you allocate against squares rather than against last quarter’s channel split, and content, PR, product marketing and paid all read the same document.

The stakes are not speculative. Roughly 60% of searches now end without a click. Forrester’s 2025 buyers’ survey finds 94% of B2B buyers using generative AI somewhere in their buying process, and twice as many naming generative or conversational AI as more meaningful than any other information source. When that much of the decision happens inside an answer, the expensive gap in most marketing organisations is not budget. It is that nobody owns the sentence above the ad.

The playbook

1.  Map before you bid. Score your top commercial prompts on your presence, competitor presence and answer quality. No spend until the map exists. It is a two-week exercise and it survives the campaign.

2.  Bid where you are losing, build where you are absent, defend where you are already the answer. Three squares, three different budgets. Paid interrupts defaults; it does not create them.

3.  Measure the sentence, not just the click. Track mention rate, explainability and cited sources alongside ROAS. If the answer improves and the ad spend falls, that is the win, not a reporting problem.

ChatGPT ads are a good channel and they will get better. But the brands that will do well on them are the ones that already know what the model says about them when nobody is paying. Everyone else is buying distribution for a verdict they never read.

References

· OpenAI, “Our approach to advertising and expanding access,” 16 January 2026. openai.com/index/our-approach-to-advertising-and-expanding-access/

· OpenAI, “Testing ads in ChatGPT,” 9 February 2026 (updated through 11 August 2026). openai.com/index/testing-ads-in-chatgpt/

· OpenAI, “A milestone in expanding access to AI,” 31 August 2026. openai.com/index/expanding-access-to-ai-with-chatgpt-ads/

· OpenAI Help Center, “Ads in ChatGPT”. help.openai.com/en/articles/20001047-ads-in-chatgpt

· Forrester, Buyers’ Journey Survey, 2025 (94% of B2B buyers use generative AI in the buying process).

· Forrester, 2024 (92% of buyers start with a vendor in mind; 41% with a single preferred vendor).

· 6sense, B2B Buyer Experience Report, 2024 (buyers ~70% through the journey, 85% of requirements set before contacting sales).

· Aggarwal et al., “GEO: Generative Engine Optimization,” KDD ’24 (~40% visibility lift from quotations, statistics and citations across ~10,000 queries).

· Refine Labs, hybrid attribution analysis (620 conversions, $21.5M closed-won, 12 months).