ChatGPT Ads Reach 89 Countries, but Auction Wins Are Not Impressions

AI-generated editorial illustration.

On October 21, OpenAI gave advertisers two numbers to take into a budget meeting. ChatGPT Ads had become available in 89 countries, after adding 26 markets that day. More than 100,000 advertisers had run campaigns since launch. A separate engineering account said the advertising business had reached a $1 billion annualized revenue run rate less than 200 days after it began.

Then the engineers described an ad that wins its auction and never appears. ChatGPT may finish its answer before the ad system returns a placement. The result can be selected on the server and still arrive too late for the person reading the conversation. OpenAI calls that difference the gap between delivery and a user-facing impression.

An auction report and an invoice can describe different events. A media buyer can pay for impressions or clicks, connect a conversion event and compare the channel with search. Before moving money, the buyer needs to find out which point each report measures, what was excluded before it, and whether a sale would have happened without the ad.

The 89-country launch makes that decision less hypothetical. OpenAI is taking a new buying surface to more local businesses and global brands, including 14 additional Latin American markets and countries in Europe, the Middle East, Africa and Asia. Its 1.2 billion weekly ChatGPT users represent the whole product, according to the company. They are not the number of people who can see a paid placement, an impression count or an addressable market estimate for a particular advertiser.

Eighty-nine markets, two buying models

The October 21 country announcement puts ChatGPT Ads in 89 markets. OpenAI says advertising appears only on Free and Go plans. Plus, Pro and Enterprise remain ad-free. Its consumer help page also says Temporary Chats do not carry ads and accounts identified as belonging to minors are excluded. A weekly user total across every plan and setting therefore cannot be multiplied by an advertiser’s desired frequency to estimate paid inventory.

An eligible user is only the first filter. A conversation about a product may have no appropriate ad moment. A brand may not operate in that country. Its creative may fail policy review, its budget may be exhausted, or the auction may find the placement irrelevant. OpenAI explicitly describes returning no ad as an ordinary successful outcome. A campaign could enter a newly available country and receive few impressions without anything having broken.

The invoice can start from either of two buying units. OpenAI’s Ads basics offers a Reach objective billed per thousand impressions and a Clicks objective billed per valid click. Advertisers set a maximum CPM or CPC bid at the ad-group level. The company recommends a starting maximum CPC bid of $3 to $5 for click campaigns. That is bidding guidance, not an average price actually paid. The relevance-weighted, second-price auction can select another price, and a campaign with a weak landing page or a narrow set of suitable chats may deliver little at the chosen bid.

The billing rules matter to a finance team as much as the bid. A daily budget is an average over seven days, not a hard daily cap. Billed media spend may reach twice the selected daily budget on one day, while staying within seven times that budget over the applicable seven-day period. Alternatively, a campaign-total budget sets a maximum for the campaign. OpenAI also offers conversion-optimized campaigns with click or impression billing for eligible accounts. It does not bill per conversion. A marketer who calls an optimized campaign “pay for outcomes” would be assigning the wrong financial risk.

A company selling a $40 item could choose a $5 maximum CPC bid and a $100 daily budget, but those are configuration inputs in an illustrative calculation, not a forecast of traffic, cost or return. At that bid ceiling, twenty paid clicks would use $100 before any production cost, agency time or uncertain attribution. A 2% purchase rate would produce less than one sale on average in a group of twenty clicks. The arithmetic only explains why a bid and a click count cannot establish profitability. It does not describe the platform’s actual clearing prices or any advertiser’s conversion rate.

OpenAI’s $1 billion figure has a related denominator problem. Annualized run rate projects a recent revenue pace over twelve months. It is not a disclosed completed year of ad revenue, profit, cash collected or advertiser return. More than 100,000 advertisers is a count of organizations that have run ads since launch, not the number currently buying or renewing. The release documents a fast business expansion. It does not yet disclose the concentration of spending, the distribution of results across buyers or the share of available conversations that carry ads.

The ad that never reaches the screen

Ritendra Datta and Karthik Gopalratnam, the two engineers named on OpenAI’s October 21 account, describe an advertising service that runs beside the service writing ChatGPT’s answer. A custom model interprets conversation context, retrieves possible ads, filters ineligible candidates, evaluates relevance and quality, and calls an auction. That separation is a deliberate product choice: the ad process, the company says, cannot change the organic response, and an ad service that is slow or unavailable should not delay ChatGPT’s answer. Those are architectural claims from the operator, not results from an independent system audit.

Their decision also limits what the advertiser can demand from the product. An ad server could wait longer for a better candidate, but the person asked ChatGPT a question, not an ad network. Holding the answer while the auction finishes would make a commercial placement part of the response’s critical path. OpenAI chose to let the answer proceed. The missed placement is the price of that boundary. From a user’s standpoint, it may be a good trade. From an advertiser’s standpoint, it makes campaign forecasts less like a fixed number of searches multiplied by a bid.

The final timing constraint comes after the auction. The ad system sends a chosen creative and a model-generated preamble to the frontend. If the bundle arrives before the answer is fully generated, ChatGPT can show the unit below the answer with a sponsored label. If it arrives too late, the unit is not shown for that turn. The engineering post says an auction winner becomes a delivery only after the client approves and displays the placement and that delivery is acknowledged. It also distinguishes an earlier server-side ad delivery from a user-facing impression. The precise event labels in an advertiser export deserve confirmation in a live account; the public description establishes that selection, arrival and display are different events.

An advertiser’s Ads Manager reporting defines an impression as a time an ad was shown. That is the count a buyer should reconcile to billed CPM media, with the account’s own invoice and methodology. A selected auction candidate is not a shown ad. A server log of a returned creative is not, by itself, evidence that the user saw it. A click is another event again. Without access to a real campaign, Digidai cannot verify OpenAI’s impression instrumentation or infer how often the late-arrival branch occurs.

This does not establish overbilling. The public reporting definition points in the right direction: shown ads, not auction candidates, are impressions. A buyer who pays only for a valid click may care less about the selection gap than about whether the click reached a working page. The gap matters because the engineering article introduces more internal events than a campaign report necessarily exposes. It gives the buyer a sharper question to ask about delivery and forecasting, not evidence that a customer was charged for unseen inventory.

ChatGPT’s answer time changes from turn to turn. The ad service has to decide whether the current exchange creates a useful moment and finish enough work to return before the answer closes. An auction can clear with a valid creative while the placement misses that deadline. Faster ranking could increase display opportunities; it could also change which candidates receive attention. Neither effect can be calculated from revenue run rate or advertiser count.

For comparison, Google’s description of ads in AI Mode says it is testing labeled formats integrated into AI-generated responses. OpenAI describes its ad below an answer, with a separate ad-generation path. These are different product designs, and neither page supplies a common denominator for conversion rate, viewability, incremental sales or cost. Calling either system the superior inventory would require a controlled buyer test, not a side-by-side reading of launch pages.

OpenAI does not publish, for the 89-market rollout, the proportion of eligible turns that start an ad session, the share with retrieved candidates, the fraction filtered before auction, the late-arrival rate, or a market-level impression breakdown. That does not mean those numbers do not exist internally. It means the buyer should request the available reporting and avoid turning a country count into a delivery forecast.

A conversation keeps changing the query

The engineers’ example begins with a person asking how to train for a race. Several turns later, the person asks about shoes. An ad selected against the first message could miss the need. Waiting until every detail is settled may leave no placement opportunity. In OpenAI’s design, an ad session carries context across a turn and updates as the conversation develops. With personalized ads enabled, selected nonsensitive signals from past conversations may also enter the match, subject to the company’s stated controls.

That makes a conversation different from a search keyword. OpenAI’s advertiser guide allows context hints that describe topics or situations where a product may be relevant. It says those hints are not exact-match keywords and do not guarantee delivery in a particular conversation. The company evaluates the ad’s landing page, title, copy and advertiser-supplied context. A media buyer cannot simply move a list of search queries into the new channel and assume the same traffic, control or intent.

OpenAI says its orchestrating model turns a developing conversation into retrieval queries and semantic embeddings, then calls tools for candidates, filtering, ranking and the auction. It may decide that a user is speaking hypothetically or referring to a stale topic rather than shopping. The same word can signal a purchase in one turn and an example in another. The account is a description of OpenAI’s system; it does not publish false-positive or false-negative rates across real chats.

The company’s September 16 advertising update adds work on both sides of this market. It introduced a test of Sponsored Agents that a person could open after clicking an ad, natural-language campaign creation and analysis in ChatGPT Work, AI creative tools in Ads Manager, and integrations with HubSpot and Shopify. The announcement does not show that agencies need fewer employees. It does show that a campaign manager’s inputs can move from fixed keywords and hand-built creative variants toward context descriptions, review of machine-adapted copy, landing-page accuracy and downstream event definitions.

Creative adaptation creates a concrete quality problem. OpenAI’s current creative-enhancement help says enabled variations may run automatically and an advertiser need not preview every version. It advises keeping prices, promotions, product details and availability current. If a landing page says a product ships tomorrow but a generated variation suggests same-day delivery, an improved click rate would be a poor result. Human review needs a sample of actual served variants and the current destination, not only approval of the original brief.

A local merchant also faces a geographic problem hidden in the headline number. Availability in 89 countries does not mean every language, payment method, category and measurement integration works identically. OpenAI says advertisers can access the new markets through technology partners or self-service Ads Manager. A brand entering two countries should ask for account-level availability and supported objectives in each, then test a local landing page. The country announcement is a launch map, not a completed buying guide for every market.

Who sees the conversion

A buyer opening Ads Manager can see impressions, clicks, spend, click-through rate, average CPC, average CPM and attributed conversions when conversion measurement is set up. OpenAI’s measurement help says impressions, clicks and CTR usually refresh about every 15 minutes, spend insights can lag by seven to eight hours, and attributed conversions may take 24 to 48 hours. A dashboard with recent clicks and zero spend or sales may be showing different processing clocks. A buyer who pauses a campaign from that snapshot could be acting on incomplete data.

The conversion path needs more work than adding a tracking link. OpenAI’s conversion documentation describes a Pixel or Conversions API sending configured website or app events. The platform must receive a qualifying event, match it to the campaign’s configured conversion, and connect it to an eligible ad interaction within the relevant attribution window. The advertiser decides whether an order, registration or lead is the useful unit. A newsletter signup and a paid purchase cannot be treated as interchangeable just because both appear under a conversion heading.

The reporting page offers selectable click-through windows of 7, 14 or 30 days and a view-through option. Changing a window can change an attributed total without changing what customers bought. OpenAI describes its model as last-touch, not multi-touch. Attribution assigns credit under a rule; it does not estimate the sale that would have happened without the ad. Finance needs that distinction when a platform-reported cost per acquisition is compared with a channel’s incremental return.

OpenAI’s October 5 measurement release named integrations with Hightouch, Tealium and LiveRamp for conversion data, several web and app attribution partners, and early work with geo-experiment providers Haus, Measured and WorkMagic. The same release presented early advertiser examples. A partner reported WeightWatchers’ attributed acquisition cost on ChatGPT Ads 15.3% below its blended paid-search benchmark. Another partner reported incremental purchases for Dose, 67% of them from new customers. The figures are selected partner and customer accounts with different methods and populations; neither is a platform-wide return on ad spend.

DoubleVerify’s own October 5 release describes the WeightWatchers result against that customer’s established benchmark during a measured period. It does not supply a randomized comparison between equal audiences on two channels. DoubleVerify is a commercial measurement partner to OpenAI, so its report is useful evidence of available attribution work, not an independent census of advertiser results. The company also says a brand-suitability evaluation pilot is being developed. That pilot should not be described as an already deployed, universal prebid guarantee.

WeightWatchers growth vice president Jake Dmochowski called the result an encouraging early signal as his company considered scaling the channel, according to the partner’s release. His phrasing is narrower than a claim that the new channel has permanently beaten search. It leaves open whether the visitors differ, how long they retain a membership, whether the benchmark came from the same season, and which sales another channel would have won. A finance team can value an early signal while reserving the larger budget decision for a test built around those missing comparisons.

The advertiser faces a privacy constraint as well as a statistical one. OpenAI says advertisers receive aggregated reporting such as views and clicks, while conversations, memory, a person’s name, precise location and other personal details are not handed to them. That is a product policy claim, not a Digidai inspection of data flows. A brand that cannot inspect the exact chat where a placement appeared needs other ways to audit suitability, troubleshoot disputed attribution and test whether matched conversions represent new demand.

AI-generated editorial illustration. The panels separate display, visit and purchase; the drawing is not campaign evidence.

The brand-safety test happens inside the chat

The advertising policy excludes placements next to sensitive or emotionally vulnerable conversations and restrict categories including health claims, financial or legal services and political content during the initial program. The company says it uses safeguards to separate sponsored placements from ChatGPT’s answer. A safety policy is an operating promise. Its effectiveness would require observation of actual placements and appeals over time; a policy page cannot provide a violation rate.

The general legal line is older than conversational AI. The US Federal Trade Commission’s native advertising guide says commercial content should be identifiable as advertising and that necessary disclosures must be clear and prominent on the devices people use. The agency tells businesses to consider the overall impression of a placement, not just a label in isolation. This guidance does not find that ChatGPT Ads violate US law. It explains why the exact position, wording and visibility of a sponsored label matter when an ad sits below an assistant’s answer.

OpenAI’s engineering account adds an unusual piece: a model-generated preamble explains how the sponsored placement relates to the conversation, while the advertiser’s creative remains a separate object. The company argues this division keeps its assistant’s voice neutral. A user may nevertheless read the preamble and ad as one recommendation. Whether the visual separation is understood cannot be established from a diagram. A buyer could test recognition on the devices and languages it uses, then examine whether any translated or adapted creative implies endorsement or a claim the landing page cannot support.

DoubleVerify’s announced suitability pilot addresses another blind spot: context changes from one turn to the next, and advertisers are not supposed to receive private conversations. The partner says it intends to assess the environment in a controlled test without access to private user chats. That is a stated pilot objective. It leaves practical questions about categories, error rates, appeals and independent reporting open. Those are procurement questions for a brand with strict placement rules, not proof of a defect in every ChatGPT ad.

The wider measurement industry has similar concerns, although it has not audited this product. In an IAB/BWG Strategy survey of 430 respondents, 51% described legal, governance and compliance concerns as significant or critical challenges for AI in advanced measurement over the next one to two years. Accuracy and transparency, and data security, each drew 49%. These are reported concerns across marketing measurement. They cannot be read as the prevalence of privacy incidents or misplaced ChatGPT Ads.

Media buyers inherit an unfamiliar job

A conventional campaign manager already watches reach, bids, creative quality, conversions and brand safety. ChatGPT Ads adds a new set of decisions to each. A context hint is not a keyword. An auction selection may not become a visible placement. The same campaign can buy CPM or CPC. A seven-day average budget can spend twice its nominal amount on an individual day. A conversion can arrive in a report two days after a click, under a window the buyer selected.

For a small business, the extra work may make the self-service promise harder to realize. The owner can create an ad in Ads Manager, yet still needs someone to confirm local stock, reject inaccurate generated variants, configure the purchase event and answer customers who arrive expecting the wrong offer. A larger advertiser can assign those tasks to media, analytics, legal and product specialists. The small firm often has the same tasks concentrated in one person. That asymmetry is a plausible adoption constraint, not a measured estimate of how many staff the new platform creates or removes.

Those differences give the budget owner more to check. They also redistribute work across teams. A media buyer needs the product team to verify landing-page promises and the analytics team to define conversion events. Legal or brand staff may need to review creative variation and sensitive categories. Finance needs the billable unit and settlement timing. An agency may build the campaign, but the advertiser still owns the product claims and the decision to scale it. No current source shows the net hours or headcount saved by using AI tools in this process.

OpenAI built its own advertising stack quickly and credits Codex as a development force multiplier in the engineering account. The business had an annualized revenue run rate above $1 billion in less than 200 days, according to that same company account. Those facts describe OpenAI’s production pace and sales. They do not show that a small merchant can now run a profitable campaign without a skilled marketer, or that an advertiser has less creative, measurement or customer-service work after adopting the channel.

One operational test is more informative than a general claim about AI replacing a media planner. Ask a buyer to trace one campaign from context hints through eligible placements, billed impressions or clicks, attributed orders and returns. Have another reviewer check whether the creative and landing page made the same promise. If that task cannot be completed from the data available to the account, the next skill investment is in measurement and coordination, regardless of how quickly an agent drafted the ad.

From eligible turn to an incremental sale

A buyer can use seven checkpoints to decide whether the new channel has earned more money. The table is an editorial pilot design based on the public product documents. It is not a report from a campaign Digidai operated.

Stage	Evidence to request	Decision it supports	Boundary
Eligible opportunity	Markets, plan eligibility, policy and context settings	Is the intended audience reachable?	Weekly users are not eligible turns.
Candidate retrieval	Available diagnostic categories for matches and rejections	Are hints and creative finding the right context?	A retrieved ad was not shown.
Auction selection	Bid, pacing and auction-level definitions available to the account	Can the campaign compete within its budget?	An auction winner was not necessarily displayed.
Ad delivery	Server and client event definitions, where reported	Did the placement reach the frontend in time?	Server receipt is not a person seeing an ad.
Visible impression	Ads Manager impressions reconciled with the CPM invoice	What was shown and what was billed?	An impression is not attention or a click.
Click and visit	Valid clicks, tagged landing-page sessions and page health	Did interested people reach the promised offer?	A click is not a lead or a sale.
Outcome	Configured events, attribution window, net margin and a holdout or geo test	Did the campaign add profitable demand?	Attributed orders may have occurred anyway.

Some diagnostics in this table may not be exposed to a self-service advertiser. That is itself a finding to record, not a reason to fill the cells with estimates. The starting test can stay small: one product, two eligible markets, one buying objective, a verified landing page and a predeclared conversion event. Check the account’s actual budget terms before launch. Keep a separate log of creative variants, placement-policy exceptions, refunds and customer complaints. Compare the cost of this review work with the value of the orders it helps earn.

At the next budget meeting, the buyer can show an Ads Manager impression count and a paid invoice. OpenAI and its partners have begun offering attribution and exploring incrementality tests, but the public launch numbers cannot say what that particular brand gained from its next dollar. The disputed line on the budget sheet is the incremental sale: how many purchases would disappear if the campaign stopped, and what did the team spend to find out?