ChatGPT Advertising: How Ads on ChatGPT Actually Work

ChatGPT advertising went from a rumour everyone argued about to a self-serve platform you can log into and spend money on, in roughly four months. That is fast. Google took years to get from "we might put ads on search" to an ads manager, and Meta took longer still.

So the useful question is no longer whether OpenAI will sell ads. It does. The question is what the platform actually looks like from inside the account, what it costs this month rather than at launch, and whether your business should be spending anything on it yet.

I'll be straight with you about that last part: for a lot of businesses reading this, the answer is no. Not yet. I'll explain where that line sits.

What ChatGPT Advertising Actually Is

ChatGPT advertising is a paid placement system that shows sponsored cards to ChatGPT users based on what they are currently talking about. The ad sits below the assistant's answer, labelled "Sponsored", in its own visually separated block. The model answers the question first. Then the ad appears.

That sequencing matters more than it sounds. On Google, the ad sits above the thing you came for, and you scroll past it. Here the answer comes first and the ad arrives after the user has already got what they wanted. Whether that makes the ad more useful or easier to ignore is genuinely unresolved, and anyone telling you they know is guessing.

The rough timeline: OpenAI began testing with logged-in US adults on 16 January 2026. The pilot went official on 9 February. Self-serve buying through Ads Manager opened on 5 May, and that is the date that actually changed things — before it, you needed a large committed budget and a conversation with a salesperson. After it, you needed a credit card.

Who Sees the Ads — and Who Never Will

Ads show to logged-in adults on two tiers: the free plan, and ChatGPT Go at around $8 a month. Everyone above that — Plus, Pro, Business, Enterprise, Education — sees nothing. Logged-out users are excluded. Under-18 accounts are excluded.

Think about what that means for your targeting. The people seeing your ad are, by definition, the people who declined to pay $20 a month. That is not a bad audience. It is a very large audience and it skews toward consumers, students, and price-sensitive small businesses. But it is not where the senior enterprise buyer with a corporate card is sitting, because that person is on a Business seat and will never see a single ad you run.

If you sell to enterprise, read that paragraph again before you allocate budget.

The Ad Formats, Plainly

Three things exist right now, and the naming across the industry is inconsistent enough that it's worth being concrete about what each one is.

Sponsored recommendation cards. The main format. A product or service card that appears below a conversational response, when the conversation topic matches what you've told the system you sell. This is the one most advertisers will actually run.

Sponsored results in search. When a user's query triggers ChatGPT's search behaviour, sponsored results can appear among the returned links. Closest thing to a traditional search ad in the whole system.

Native content cards and companion display. Visual units that render after a model response without interrupting it. Product feed campaigns for retail plug into this — if you already maintain a shopping feed for Google, the lift here is small.

What does not exist: ads inside the model's answer. OpenAI has been firm that sponsored content does not get woven into generated text, and so far that has held. I'd treat it as a policy that could move, not a law of physics.

How Targeting Works Without Following You Around the Web

This is the part that is genuinely different, and it's the part most guides get wrong by describing it in Facebook terms.

You do not build an interest-graph audience. You give the system broad context hints — topics, scenarios, product categories, the situations in which someone would want what you sell. The matching engine pairs those hints against the live intent of the conversation happening right now. Three signals feed it: the topic of the current chat, the user's past chats and memory, and how they've interacted with ads before.

What is not in the mix: third-party cross-web tracking. No pixel-following you from site to site, because that infrastructure doesn't exist here.

And advertisers receive no user data back. Not chats, not names, not emails, not IP addresses, not precise location. You get aggregate performance. You do not get to see who you reached.

What ChatGPT Ads Cost in August 2026

Almost every guide you'll find quotes $60 CPM. That number is stale, and repeating it will cause you to make a bad decision.

$60 was the default maximum CPM bid at launch. It was never a market-clearing price — it was a ceiling. Pilot advertisers have reported actual delivered CPMs falling toward $25 as inventory expanded and more markets came online. On the click side, the recommended entry bid sits at $3 to $5, and one published test by Choice OMG in June measured averages of $4.42 to $4.88.

Let's do the arithmetic nobody else does. Take a modest test budget — say $2,000. At a $25 CPM that buys 80,000 impressions. At the 1.30% CTR that's been reported, you get roughly 1,040 clicks, for an effective cost per click of about $1.92. At $4.50 CPC on a click-billed campaign, the same $2,000 buys around 444 clicks.

Which means, on those numbers, CPM buying is currently cheaper per click than CPC buying. That is unusual and it will not last — it's the signature of a young auction where impression supply outruns advertiser demand. If you're testing, test CPM first.

One caveat I want to state clearly: OpenAI itself says it does not yet publish performance benchmarks. Every figure above comes from third-party advertiser reporting, and the sample sizes are small. Treat them as directional, not as a media plan.

Bidding: CPM, CPC, and the New oCPC Objective

Three objectives. CPM and CPC do what you'd expect. oCPC arrived in summer 2026 and is the interesting one — it bills per valid click while optimising toward conversions, so you pay click prices for conversion-shaped delivery.

There are constraints worth knowing before you build a campaign around it. Custom conversion events are not supported. You must nominate exactly one standard conversion event per campaign. And OpenAI states there is no recommended bid amount for oCPC at this time, which is a polite way of saying nobody has enough data yet.

There's also a "Maximize results" automated bidding mode, and targeting can be split across iOS app, Android app, and web.

My read: run CPM to learn what your creative does, then move to oCPC once you have enough conversion volume for the optimiser to have something to chew on. Starting on oCPC with a thin conversion signal wastes budget teaching a model nothing. If you have someone managing paid campaigns already, this is the sequencing conversation to have with them before launch, not after.

Measurement: The Pixel, the Conversions API, and What Still Breaks

This is the section every other guide skips, and it's the one that will determine whether your campaign is measurable or a black hole.

The pixel

OpenAI ships a JavaScript pixel via the OAIQ SDK. It sets a first-party cookie named __oppref with a 30-day lifetime, and includes automatic advanced matching — SHA-256 hashing done in the browser, on by default for new web pixels. Version 0.1.3 is current at time of writing.

Note the phrase "OpenAI generates pixels based on advertiser needs." You do not hand-write the tag. That is more restrictive than the Meta pixel and it means some of the custom event gymnastics you're used to simply aren't available.

The Conversions API

Server-to-server event sending, with deduplication against pixel events via shared event IDs. It accepts events from web, mobile app, offline, physical store, phone call, and email sources — which is a genuinely broad list, broader than I expected.

What's missing: there are no native CRM or CDP connectors published yet. If you want Salesforce or HubSpot conversions flowing in, someone on your side is writing that integration. Budget for it.

The 30-day window problem

Thirty days is short if you sell anything with a considered purchase cycle. Travel, B2B software, education, real estate, high-ticket services — all routinely run 45 to 90 days from first touch to close. Those conversions will attribute somewhere else, and ChatGPT will look worse than it performed.

Plan for that by holding a clean holdout region or a pre/post spend comparison. Do not judge this channel on last-click alone. You'll kill something that was working.

Custom Audiences and Their Very Real Minimums

You can upload email or phone lists, raw or SHA-256 hashed, up to 5 million identifiers per upload. Use them for inclusion, exclusion, or as bid multipliers at ad group level.

Now the number that stops most small advertisers cold: you need a minimum of 25,000 matched users before an audience is usable, and OpenAI recommends 100,000.

Matched, not uploaded. If you have a 40,000-person email list and half of those people don't have ChatGPT accounts under that address, you don't clear the floor. For most local businesses and most early-stage companies, custom audiences on this platform are simply not available yet, and pretending otherwise wastes a week.

ChatGPT Advertising vs Google Ads

The honest comparison, without the breathless framing.

The one row that should make you pay attention is competition. In most verticals, almost nobody is bidding here yet. That is the entire argument for testing now — not superior technology, not better intent, just an auction with fewer people in it.

The row that should make you cautious is the absence of an ad library. On Meta you can look at what competitors are running. Here you cannot, and I don't understand why OpenAI hasn't shipped one — every mature ad platform ends up with transparency tooling, usually because a regulator insists. Expect it eventually. Until then you're flying without that instrument.

Where It's Live Right Now

As of late August 2026, self-serve is available in nine markets: the United States, United Kingdom, Canada, Australia, New Zealand, Japan, Korea, Brazil, and Mexico.

Thirty-one European markets are opening through partners from 24 August, with self-serve following later. Europe moving through partners first rather than direct is almost certainly a regulatory sequencing decision — DSA and GDPR obligations around ad transparency and profiling are heavier there, and partner intermediaries absorb some of that.

India Isn't Live Yet — What This Window Is Actually For

India is not on the self-serve list. You cannot buy ChatGPT ads for Indian audiences today.

What is happening: OpenAI has begun expanding the advertising test to India and has emailed eligible users about the change. Ads will appear for selected adult accounts on the Free and Go plans, same as everywhere else. Reporting points to a Q3 2026 rollout, though OpenAI has not committed publicly to a completion date, and early CPM estimates in the region of ₹5,000 per 1,000 impressions have circulated — treat that as an estimate from the trade press, not a rate card.

The groundwork has been visible for a while. ChatGPT Go launched in India in August 2025. A major OpenAI brand campaign ran here in December 2025. You do not spend on brand marketing in a country you aren't about to monetise.

So what do you actually do in the gap? Three things, in order of how much they'll matter later.

Install the pixel now. Not when ads launch. Now. Attribution improves with historical data, and a pixel that has been collecting for four months before your first campaign is worth considerably more than one installed the morning you launch.

Fix your product feed. Retail campaigns run off feeds. If your store feed is the neglected Google Merchant Center export nobody has audited since 2023, that's a week of work you can do now instead of during launch week when everything is urgent.

Write your context hints. This is a copywriting problem, not a media buying one, and it's the skill nobody has yet. Sit with your sales team and write down the actual situations in which somebody needs what you sell — the scenario, not the keyword. "My AC stopped cooling and the compressor is making a noise" rather than "AC repair Jaipur". Conversational intent is described in sentences, and the advertisers who write those sentences well will beat the ones who paste in a keyword list.

Who Should Run These Ads, and Who Should Wait

Run them if you sell direct to consumers, your conversion happens in under 30 days, you're in one of the nine live markets, and you can afford to spend $2,000 to $5,000 learning something that may not work. The thin auction is a real edge and it closes as more advertisers arrive.

Wait if any of these apply to you:

• You sell to enterprise. Your buyers are on Business seats. They will not see your ad.
• Your sales cycle exceeds 30 days. The attribution window will make good performance look like bad performance.
• Your list is under 25,000 matched users. Half the targeting toolkit is locked to you.
• You're in India or another market that hasn't launched. There's nothing to buy.
• Your Google and Meta accounts aren't already profitable. This is the one people don't want to hear. A new channel with no benchmarks will not fix an offer that isn't converting on mature channels. It will just cost more to discover the same thing.

That last point is the one I'd push hardest. The excitement around a new ad platform makes people forget that channel selection is rarely why campaigns fail.

Setting Up Your First Campaign

Assuming you're in a live market and you've decided to test, here's the order of operations that wastes the least money.

One. Install the pixel and confirm __oppref is being set on your domain. Verify it fires on your actual conversion page, not just the homepage. This takes twenty minutes and people skip it constantly.

Two. Stand up the Conversions API alongside the pixel, with shared event IDs for deduplication. Yes, both. Browser-side tracking degrades and server-side is your floor.

Three. Pick one standard conversion event. Not three. The optimiser wants a single clear signal, and oCPC requires exactly one anyway.

Four. Launch on CPM, not oCPC. You're buying information about creative performance at this stage, and CPM is currently the cheaper way to buy clicks regardless.

Five. Write three to five genuinely different context hint sets. Not variations on a theme — different framings of the problem you solve. This is where the learning is.

Six. Run for at least two weeks before judging anything. With a 1.3% CTR and small budgets you will not have significance before that, and every early read will be noise.

What ChatGPT Ads Mean for Your SEO

Short version: less than the panic suggests, but not nothing.

Sponsored cards sit below the answer, in a separate block. They don't replace the model's response, and the model's response is still assembled from sources it trusts. Your organic visibility inside AI answers — whether ChatGPT cites you, recommends you, names you — remains a content and authority problem, not a budget problem. You cannot pay your way into the answer itself.

What changes is the space below the answer. That real estate used to be empty, and now it's commercial. If a user reads the answer and then sees a competitor's sponsored card, you've lost the click even though you won the citation.

Which argues for doing both, and for not treating them as separate departments. The businesses that do well here will be the ones whose organic content gets them named in the answer and whose sponsored card sits underneath it. Same brand, twice, in one screen. That's the play.

Frequently Asked Questions

Budgeting a digital project rather than a campaign? The same question-by-question breakdown for build costs is in our FAQ on ecommerce website cost in India.

Where This Leaves You

ChatGPT advertising is a real platform with real tooling and a genuinely uncrowded auction. It is also six months old, missing an ad library, capped at a 30-day attribution window, and unavailable in a large part of the world including India.

Both of those things are true at once, and the correct response to both being true is a small, well-instrumented test rather than either a land grab or a wait-and-see. Install the measurement layer regardless of where you are — that costs a morning and pays off whenever your market opens. Then spend the smallest amount that can teach you something real.

And if your existing channels aren't working, fix those first. No new auction has ever rescued a bad offer.