Audiences too small to target can now be used to exclude, list ceilings have moved, and a plugin lets campaigns be built and edited from a chat window. The product update also puts carousel card impressions into reporting while keeping them off the invoice.

OpenAI has loosened the operating rules that have governed custom audiences inside ChatGPT Ads Manager since July, in a product update circulated to advertisers in the platform's beta during the week beginning Monday, August 31, 2026. The same email adds a plugin that runs campaign work from inside ChatGPT or Codex, new identifier fields for conversion matching, carousel reporting that sits outside billable impression counts, pacing for campaigns holding a total budget, and confirmation that sponsored placements are now rolling out across select countries in Europe, India, the Middle East and North Africa.

The regional item lines up with the announcement OpenAI published on August 31, 2026, which put ChatGPT Ads in more than 40 countries and stated that self-service purchasing through Ads Manager was opening the same day across those four regions. The email does not enumerate the countries.

A plugin that operates the ad account

The first item in the update is a tool rather than a metric. According to OpenAI, the ChatGPT Ads Manager plugin allows advertisers to "Create, manage, and analyze campaigns from ChatGPT or Codex." The described scope is unusually wide for a first release: turning a website or a written brief into ready-to-launch ads through plain language instructions, generating creative variants, updating live campaigns, troubleshooting delivery problems, reviewing performance, and returning prioritised recommendations.

One sentence in the description carries more weight than the feature list. The plugin, according to OpenAI, "previews recommended changes and asks for confirmation before applying them." That is a human-in-the-loop gate on write actions, and it places the product on the cautious side of a line the advertising industry has been drawing all year. Setup runs through the Plugins directory, followed by an account connection.

The move puts OpenAI's own advertising platform inside the same category of agentic AI tooling that rival platforms reached first, and by a different route. Meta opened its ad system to Claude and ChatGPT through connectors carrying write access from the first day, on April 29, 2026, then opened the underlying ads MCP server to any developer holding a Meta app on July 16. X shipped an Ads MCP server with 23 tools, ten of them writes, on August 24, 2026. Microsoft took the opposite posture at Cannes Lions in June 2026, launching an Advertising MCP server in read-only form.

The architectural distinction matters for agencies. Those releases published a protocol endpoint that any compliant agent could call, which is why a platform owned by one AI company ended up documenting setup instructions for a rival's coding agent. OpenAI's plugin runs inside OpenAI's own surfaces. The email does not state whether the plugin exposes a Model Context Protocol server, whether third-party agents can reach the same functions, or which operations are available as reads against writes. Nothing in the update answers those questions, and the absence is worth noting rather than filling in.

The 25,000 floor moves, but only in one direction

Six changes land on custom audiences at once, and one of them rewrites a rule PPC Land documented in detail. OpenAI's help centre documentation, published in July 2026, blocked any custom audience from being used in a campaign until it reached 25,000 matched users, with uploaded files capped at 500,000,000 bytes. That floor now has an exception: audiences below 25,000 matched users can be used for exclusions.

The asymmetry is deliberate and follows the logic of the original restriction. A small inclusion audience narrows delivery onto a handful of identifiable people, which is the privacy problem a match floor exists to prevent. A small exclusion audience removes those people from delivery, which does not concentrate targeting on anyone. Suppression lists are the practical beneficiary. An advertiser holding a few thousand recent purchasers, refund cases or churned accounts previously had no way to keep them out of a campaign, because the list could not clear the threshold in either direction.

Four further changes alter how lists are built and maintained. Members can be added, removed or replaced without recreating the audience, ending the rebuild-and-reupload cycle that list refreshes previously required. Identifier types can be mixed within a single request rather than submitted separately. Audiences can be built beyond 5 million members, wording that implies a previous ceiling at that figure, though the email does not state what the earlier limit was. Audience-size ranges are reported at finer granularity, which affects how confidently a buyer can estimate reach before spending.

The sixth addition is an identifier. Google Advertising ID, the resettable device identifier Android assigns for advertising purposes, joins the supported types. Until this update, custom audience matching in ChatGPT Ads ran on emails and phone numbers, raw or hashed with SHA-256. Adding a mobile advertising identifier gives app advertisers a matching key that does not depend on holding a customer's contact details, and it follows the mobile install attribution integrations with AppsFlyer and Adjust that reached seven markets in late July 2026.

More fields for the pixel and the Conversions API

Conversion matching gains identifier fields on both measurement paths. The Measurement Pixel now supports hashed phone, first name, last name, region and postal code. The Conversions API supports plural list fields for hashed email, phone, external ID, first and last name, country, city, region and postal code, plus Android GAID for mobile.

Plural list fields are the more consequential half. A user record that can carry several hashed email addresses or several phone numbers in one payload will match at a higher rate than one restricted to a single value per field, because customers accumulate addresses and numbers that a CRM stores in parallel. The change is a match-rate lever rather than a new capability.

It also widens the identifier surface at a moment when that surface is already under pressure. Automatic advanced matching, which collects hashed customer data from website forms and sends it back for attribution, became the default on existing ChatGPT web pixels on August 17, 2026 after a ten-day notice period. Advertisers who left that setting untouched are now sending hashed form data by default, and the field list that data can populate has just grown to include names, regions and postal codes.

That combination has a compliance dimension in the newest markets. Sponsored placements began appearing across 31 European markets on August 24, 2026, and OpenAI Ireland Limited is named in the European privacy policy as the controller for users in the EEA and Switzerland, which routes any complaint through the Irish Data Protection Commission. Personalisation was withheld at the European launch, leaving conversation context, general location and device type as the signals in play. Advertiser-side identifier collection sits on a separate legal footing from platform-side personalisation, and the update does not address how the expanded field set interacts with consent requirements in those markets.

The multi-product carousel unit, which commerce specialists observed running in early August 2026 and which OpenAI confirmed as a test in the product update circulated at the time, now has its own measurement layer. When a product feed campaign serves in the carousel format, Ads Manager reporting includes carousel card impressions and clicks, and the public Insights API carries product-level fields.

Two sentences define the metric. A carousel card impression is, according to OpenAI, "counted when an individual product card becomes viewable." Those counts are "separate from billable ad impressions" and are excluded from billable impression totals. Complete data is available from August 20, 2026 on a rolling 30-day basis.

The counting rule is the part that will generate reconciliation questions. Viewability in display advertising rests on published thresholds: the Media Rating Council standard requires 50% of an advertisement's pixels in view for at least one continuous second, with a reduced 30% threshold for large formats. OpenAI's wording states that a card becomes viewable without stating what pixel share or exposure duration triggers the count inside a horizontally scrolling unit rendered in a conversation thread. An advertiser comparing carousel card impressions against viewability figures from any accredited vendor has no published basis for the comparison.

The billing separation is the more advertiser-friendly half of the design. A carousel showing several product cards in one placement could, under a naive counting model, multiply billable impressions by the number of cards rendered. Keeping card impressions in a reporting column and out of the billable count means the format raises measured density without raising the invoice for the same placement. Sensor Tower estimated that roughly 23% of ChatGPT advertisements placed on desktop in the United States between August 15 and August 30, 2026 were carousels, which makes the accounting question material rather than theoretical.

The rolling 30-day window carries its own constraint. Data outside that window is not described as available, so any analysis of carousel card performance over a longer period depends on the advertiser exporting and retaining it. Reporting that ages out of the platform is a familiar operational burden, and it lands on a format that is weeks old.

Pacing arrives for total budgets

Campaigns configured with a total budget now spread spend across the flight rather than consuming the budget at whatever rate delivery allows. According to OpenAI, actual spend may still vary day to day based on available delivery opportunities.

The change closes a gap opened earlier in the summer. Fixed daily budget caps were converted into seven-day averages nine weeks after they were introduced, removing the hard ceiling advertisers had been given and replacing it with a rolling average. Total budgets had no equivalent smoothing mechanism at all. A campaign set to spend a fixed sum across a date range could exhaust that sum early, which matters for promotional windows tied to a specific calendar period.

No numbers accompany the pacing description. The email does not state a permitted daily variance, a ceiling expressed as a multiple of the prorated target, or whether the mechanism front-loads or evens out delivery. Google's equivalent controls publish a ceiling at twice the average daily budget. OpenAI's does not publish one.

An impression-billed campaign type is next

The forward-looking item reverses a decision OpenAI made two weeks earlier. According to the email, a new conversion optimisation model is coming that "optimizes through the inclusion of view-through conversions," allowing optimisation against conversions from both clicks and views. Billing on that campaign type is impression-based while delivery optimises toward the selected conversion goal.

Set against the platform's current posture, that is a significant shift. View-through conversion reporting appeared in Ads Manager on August 19, 2026 with a window fixed at one day, presented as a standalone column and deliberately walled off from the headline conversions total, from cost-per-acquisition calculations, from post-click conversion rate, from bidding, from billing and from conversion optimisation. OpenAI described the data as supplemental reporting that fed none of those systems.

The coming campaign type folds post-view credit into optimisation and pairs it with impression-based billing, which is the industry norm the earlier design pointedly avoided. Two weeks separate the conservative version from the announcement of its opposite. The email gives no launch date, no eligibility conditions, no window length for the view-through component, and no statement on whether the fixed one-day window carries over. It states only that more detail will follow as the campaign type rolls out more broadly.

The trajectory around it explains the pace. The pilot began on February 9, 2026 with no conversion measurement of any kind. A Measurement Pixel and Conversions API arrived on May 5, 2026 alongside self-serve access and cost-per-click bidding. Conversion-optimised campaigns began rolling out on June 5. Automated bidding became the default in new ad groups in August. Roughly seven months separate a platform with no attribution from one preparing to bid on modelled post-view outcomes.

Where the ads now run

The final item confirms the geographic position. Sponsored placements are rolling out to select countries in Europe, India, the Middle East and North Africa, according to OpenAI, "expanding reach for advertisers with business in those markets."

That sentence sits at the end of the fastest expansion sequence the platform has run. Brazil and Mexico brought the market count to nine on August 17, 2026. Advertising went live across 31 European markets on August 24. India followed on August 29, closing a five-day burst that carried the pilot into 32 new markets and taking the publicly counted total to 41. The Middle East and North Africa were the missing block, and this update fills it.

Access route has been the recurring variable. India went live without self-serve access for local buyers, and a campaign-level platform targeting control documented in August listed a minimum daily spend of 725 Indian rupees for a market that had not yet activated. The August 31 announcement resolves that pattern by opening Ads Manager purchasing directly across all four regions.

What the update means for buyers

Read together, the seven items describe a platform closing the distance to established search and social tooling on the mechanics that determine whether budget can be administered at scale. List maintenance without rebuilds, mixed identifiers in one request, plural conversion fields, budget pacing and an exclusion path for small audiences are not headline features. They are the operational plumbing that decides how much spend an agency team can reasonably manage inside an interface.

Three gaps remain visible in the disclosure. The carousel viewability threshold is unstated. The pacing tolerance is unquantified. The coming impression-billed campaign type has no date, no window and no eligibility criteria attached. Each of those is a number a media buyer would need before committing budget, and none appears in the update.

The plugin is the item with the longest reach. Every major platform that has opened its advertising stack to conversational control this year has done so through a published protocol, which lets agencies standardise on one agent across several platforms. A first-party plugin confined to ChatGPT and Codex does not offer that. Whether OpenAI follows the protocol route, and how much write authority it grants when it does, is the question the update leaves open on the surface that matters most to the buy side.

Timeline

Summary

Who: OpenAI, communicating with advertisers enrolled in the ChatGPT Ads Manager beta. The changes affect performance buyers running custom audiences, product feed campaigns and pixel or Conversions API measurement, agencies administering campaigns at scale, and advertisers in the newly opened markets across Europe, India, the Middle East and North Africa.

What: A product update covering seven items. A ChatGPT Ads Manager plugin that creates, manages and analyses campaigns from ChatGPT or Codex with a confirmation step before changes apply. Six custom audience changes, including the use of audiences below 25,000 matched users for exclusions, membership edits without recreating a list, mixed identifier types in one request, audiences beyond 5 million members, finer audience-size ranges and Google Advertising ID as a supported identifier. Additional conversion matching fields on the Measurement Pixel and the Conversions API. Carousel card impression and click reporting in Ads Manager and product-level fields in the public Insights API, counted separately from billable impressions. Pacing for campaigns with total budgets. A forthcoming conversion optimisation model that includes view-through conversions and bills on impressions.

When: The update was circulated during the week beginning Monday, August 31, 2026. Complete carousel card data is available from August 20, 2026 on a rolling 30-day basis. OpenAI's public announcement of the regional expansion and the $1 billion annualised revenue run rate was published on August 31, 2026.

Where: ChatGPT Ads Manager, now reported by OpenAI as available in more than 40 countries, with self-service purchasing opening across India, Europe, the Middle East and North Africa. The plugin operates inside ChatGPT and Codex.

Why: Advertiser demand for list control, higher match rates and predictable spend distribution has outpaced the tooling built during the pilot's first six months. The exclusion exception addresses a documented restriction that left small suppression lists unusable, while the plugin and the coming impression-billed campaign type extend the platform toward the automation and buying models already standard on established search and social platforms.