On 31 August 2026 the self-service Ads Manager opened to advertisers in Greece. Running ads inside ChatGPT was possible before this through other networks; what changed is that campaigns can now be built and managed directly. The platform is still in Beta and its capabilities are widening steadily, so treat everything below as accurate today rather than settled.
How the Ads Work
Ads appear below the answer, as clearly labelled paid placements. They do not alter the content of the response - the model answers, and the ad sits separately underneath.
Selection is driven primarily by the context and intent of the current conversation, combined with your ad copy, your landing page and the context hints you define. That list is worth reading twice: the landing page is an input to whether your ad is selected at all, not just where the click arrives. A thin page that limps along on other channels is a delivery problem here.
Context Hints, Not Keywords
Context hints describe the needs or situations where a product might be useful. They are not exact-match keywords, and they do not guarantee delivery in any given conversation.
This is the single biggest adjustment for anyone arriving from Search. Hints written as a keyword list underperform hints written as the situation a customer is actually in. "waterproof hiking boots" is a keyword. "someone planning a multi-day hike in wet conditions and asking what gear they need" is a context hint. The system is matching conversations, so describe the conversation.
Who Actually Sees Them
Ads can serve to adult users on the Free and Go plans. They do not appear for:
- Plus, Pro, Business, Enterprise or Edu accounts
- Users under 18
- Temporary Chats
Plan your audience expectations around that. The reachable pool skews toward general consumer use rather than the heaviest professional users, and a meaningful slice of high-intent B2B research happens on paid tiers you cannot reach.
Objectives and Bidding
Three objectives are available, and bids are set at Ad Group level:
- Reach / Views - CPM. Optimises for impressions, charged per thousand.
- Clicks - CPC. Optimises for website traffic, charged per valid click.
- Conversions - oCPC. Optimises toward a specific conversion event, still charged per valid click.
The auction weighs your bid alongside relevance to the conversation and the expected performance of the ad. In other words, bidding your way past a poorly matched ad does not work well - the same lesson every relevance-weighted auction eventually teaches.
Targeting: What Exists and What Does Not
Available today:
- Geographic targeting
- Web, iOS and Android targeting
- Conversational targeting through context hints
- Custom audiences built from email addresses or phone numbers
- Audience inclusions, exclusions and bid adjustments
Two constraints matter for planning. Custom audiences require at least 25,000 matched users, which puts them out of reach for most Greek advertisers at the start. And traditional interest, demographic and lookalike targeting has not been announced.
So the levers that carry campaigns on other platforms mostly are not here yet. What remains is context, geography, the quality of your product data and the offer itself. That is a different game rather than a worse one, and it rewards a specific kind of advertiser: one whose feed is clean and whose landing pages answer the question the user just asked.
Ad Formats
The base format is:
- Brand name and logo
- Headline
- Description
- Square image
- Landing-page URL
For retail there are also product-feed campaigns, which surface product image, title, price, offer and rating. If you sell products, this is the format that matters.
The Product Feed
OpenAI made a sensible decision here: the schema is Google-compatible. A healthy Merchant Center feed transfers with little change, since the documentation asks for the fields marked Required in the stable product feed specification.
There is one Ads-specific addition, and it is the detail that quietly breaks feeds on the first attempt: every product you want processed needs is_ads_eligible set to true, using that canonical field name rather than the legacy alias. Miss it and the feed uploads cleanly while nothing becomes eligible to serve - a silent failure, not an error message.
Two mechanics to know before scoping the work:
- The feed connection is created in the Feeds area of Ads Manager, not through the API, and the catalogue goes up over SFTP. The public Advertiser API has no endpoint for creating a connection or uploading a catalogue, so that step cannot be scripted.
- After the first upload, the Delta Feeds API handles incremental updates - availability, prices, titles - without re-sending everything. For a catalogue of any size this is the difference between a workable daily refresh and a nightly ordeal.
Measurement, Precisely
This is where accounts quietly lose money, so it is worth being exact.
The pixel belongs near the top of the <head> on every page where a conversion can happen - the documentation is explicit about placing it early so fast conversions are not lost. It exposes one function, oaiq(), with three commands you will use: oaiq("init"), oaiq("measure") and oaiq("consent").
When someone clicks an ad, a privacy-preserving identifier called oppref arrives on the landing page URL. The SDK captures it and stores it in a first-party cookie named __oppref so it survives across pageviews.
Standard events cover the usual funnel: page_viewed, contents_viewed, items_added, checkout_started, order_created, lead_created, registration_completed, appointment_scheduled, subscription_created and trial_started. Lead generation and subscription businesses are covered, not only retail.
An attribution_window_days value is required when configuring event settings; the documentation recommends 30. Alongside the pixel there is an image tag for pages where JavaScript is not an option, and a conversions API for server-to-server events. Plan for the conversions API from the start rather than as a later upgrade - browser-side measurement degrades for all the usual reasons, and that third pixel command, oaiq("consent"), is a reminder that consent state belongs in the first build. If you have already done server-side tracking and consent work elsewhere, the same discipline applies.
What Data You Actually Get
Worth being clear about, because clients ask and the answer is reassuring in one direction and limiting in the other.
OpenAI does not share conversations, chat history, memories or users' personal details with advertisers. What you receive is aggregate campaign performance data. Users can turn off ad personalisation, delete the data used for advertising, and hide or report an ad.
The limitation follows directly: you will not be doing conversation-level analysis, and you should not promise anyone that you can.
Policies and Brand Safety
Ads do not serve in sensitive or brand-unsafe conversations - health, mental health, politics, violence, self-harm, illegal activity. That is a meaningful protection and also a reach constraint worth setting expectations about.
Broadly, the permitted categories include consumer products, local services, travel, entertainment, digital products and education. Financial services, healthcare and legal services face additional restrictions and approvals. Political ads, gambling, tobacco, alcohol, weapons and adult content are prohibited or tightly restricted.
If you operate in a restricted category, budget time for approval rather than assuming a same-day launch.
What Nobody Can Honestly Tell You Yet
OpenAI suggests starting CPC bids in the region of three to five dollars. Be clear about what that is: a suggested starting point, not an average CPC and not a benchmark. OpenAI has said plainly that performance benchmarks across advertisers, industries and campaign types do not exist yet.
Anyone quoting you an expected cost per acquisition on this channel today is estimating, not reporting. That is not a reason to stay out - it is a reason to enter with a budget you can afford to learn with, and measurement clean enough that you produce your own benchmark instead of waiting for someone else's.
What I Have Already Built
I have been working in the platform since it opened in Greece: campaigns running, a product feed connected and processing, the pixel deployed, and conversion events mapped to what the business actually cares about rather than to whatever was easiest to fire.
Four things worth passing on ahead of any strategy conversation:
- The feed is where the time goes. A healthy Merchant Center feed migrates quickly. An unhealthy one becomes visible here, which is uncomfortable but useful.
- Check is_ads_eligible first when products are not serving. It fails silently, so it is rarely the first thing people look at.
- Write context hints as situations, not keyword lists. This is the change that most affects delivery quality.
- Get the conversions API in early. Retrofitting server-side measurement after a month of decisions made on browser-only data means re-reading that month instead of building on it.
Who Should Move Now
Not everyone. A Beta channel deserves an honest look at whether your unit economics can absorb an unfamiliar cost per acquisition while it settles, and whether your category is even permitted without a lengthy approval.
The case for moving early is that competition inside new inventory is thin, and accounts with clean measurement now will be the ones able to read the channel as it matures. The case for waiting is that you will be forming your own benchmarks rather than inheriting any. Both are defensible. What is not defensible is arriving in a year with a broken feed and no measurement, which is what usually happens.
If you already run Performance Max or Meta campaigns at scale, your feed and tracking discipline transfer directly. If your measurement is shaky, fix that first - a new channel will not be kind to reporting you cannot trust.
Working On This Together
If you want ChatGPT Ads set up properly - feed connected and eligible, pixel and conversions API in place, events mapped to real business outcomes, context hints written for how the matching actually works, and campaigns structured around the targeting that exists rather than the targeting you are used to - this is work I am doing now and taking on projects for.
The most useful way to start is short: tell me what you sell and how you currently measure it. One conversation is usually enough to establish whether this channel deserves your budget yet, which is a more honest answer than a proposal.
Sources and caveats
Platform mechanics come from OpenAI's developer documentation for Ads - product feeds, measurement pixel, conversion setup, campaign targeting and the Advertiser API - from OpenAI's announcement of the European expansion, and from my own work inside the Ads Manager since it opened in Greece. The platform is in Beta and changing: verify against the current documentation before implementing.