OpenAI is preparing to add custom audience targeting to its ChatGPT advertising platform, allowing advertisers to upload raw or SHA-256 hashed emails and phone numbers to filter which users see their campaigns - or, equally, to block them from seeing ads at all. The feature is currently in a gated rollout, according to entrepreneur Juozas Kaziukenas, who shared screenshots of the interface on LinkedIn on May 14, 2026.

The development places a familiar piece of infrastructure - the kind that Meta built into Facebook's ad platform years ago - into a context that did not exist for advertisers until February 9, 2026, when OpenAI formally launched its ChatGPT advertising pilot. The speed of feature accumulation since that launch has been notable. In roughly three months, OpenAI has moved from a closed pilot with a $200,000 minimum spend to a self-serve Ads Manager open to all US businesses, cost-per-click bidding, a Conversions API, pixel-based measurement, and now the beginnings of a custom audience layer.

What the interface shows

According to Kaziukenas, the audience creation flow appears inside the campaign creation interface at the "Create Ad Group and Ads" step. Within that flow, a new panel allows advertisers to upload a CSV or TXT file of up to 512MB containing identifiers - either email addresses or phone numbers - and to name the resulting audience segment.

The interface shows two distinct functions: an "include audience" option and an "exclude audience" option. The include path restricts ad delivery to users within the selected audience. The exclude path filters those users out, preventing them from seeing ads in that campaign. According to the LinkedIn post, OpenAI normalizes the identifiers during processing - meaning raw emails or phone numbers are processed on ingestion, and the platform also accepts pre-hashed values using SHA-256.

No lookalike audience functionality appears to be present yet. Kaziukenas noted that point directly, drawing a historical parallel: according to his LinkedIn post, "that's what FB/Meta brought next after first introducing custom audiences."

The targeting also supports landing page query parameters. The interface displays template variables including {campaign_id}, {ad_group_id}, {ad_id}, and {ad_account_id}. These can be appended to destination URLs to pass campaign-level identifiers through to landing pages, enabling downstream attribution and analytics matching.

Why this matters now

Custom audiences are the mechanism by which advertisers close the loop between their own customer data and paid media. Without them, a brand running ads on ChatGPT has no way to suppress existing customers, retarget people who have already shown intent, or protect a campaign from reaching audiences it has no interest in converting. With them, the platform starts to look like infrastructure advertisers can integrate into existing CRM and data workflows.

The feature arrives nine days after OpenAI opened its self-serve Ads Manager beta to all US businesses on May 5, 2026. That launch dropped the minimum spend requirement entirely - it had started at $200,000 to $250,000 at launch, dropped to $50,000 in April, and was removed altogether on May 5. It also introduced cost-per-click bidding alongside the existing $60 CPM model, with OpenAI recommending starting CPC bids of $3 to $5 per click. The Conversions API and pixel-based measurement landed in the same announcement.

The identity layer, as Kushal Khandelwal noted in comments under the LinkedIn post, may matter more than the ad units themselves. "Interesting part is not ads inside ChatGPT," he wrote. "It's the identity layer quietly forming underneath." The comment captures a structural dynamic that has characterized platform advertising for more than a decade: the ad format is often less consequential than the data infrastructure that surrounds it.

The context: a platform being built at speed

OpenAI confirmed its advertising plans on January 16, 2026, ending months of speculation. The company had issued a "code red" directive as recently as December 2, 2025, pausing advertising development to redirect resources toward product quality. The reversal was swift. By February 6, 2026, OpenAI had begun accepting advertisers. By February 9, the pilot was live in the United States, initially targeting ChatGPT Free and Go tier users.

Early participants included Target, Ford, Mrs. Meyer's, and Adobe. Holding companies WPP Media, Omnicom, and Dentsu were among the launch partners. The minimum commitment was set at $200,000, and the initial CPM rate landed at approximately $60 - a figure comparable to NFL broadcast inventory and more than three times the prevailing CPMs on Meta's platform.

Criteo became the first formal ad tech partner on March 2, 2026, connecting roughly 17,000 advertiser clients to ChatGPT inventory. By March 26, 2026, the pilot had crossed $100 million in annualized revenue within six weeks - a milestone accompanied by an announcement of international expansion to Canada, Australia, and New Zealand. David Dugan, a former Meta advertising executive, was named to lead OpenAI's global advertising solutions team that same week.

CPMs have since moved. By April 17, 2026, rates had fallen from $60 to as low as $25, reflecting the broader pool of advertisers introduced as minimum thresholds dropped and self-serve access expanded.

On April 30, 2026, OpenAI updated its US privacy policy to formally disclose - for the first time in binding legal language - that it receives purchase data from advertisers and their partners to measure ad effectiveness, that user information is shared with marketing partners for third-party targeting, and that personal data is used to promote OpenAI's own products. That legal foundation arrived five days before the self-serve Ads Manager opened to all US businesses. The sequencing mattered: the policy update created the legal basis for inbound advertiser data before the technical capability that depends on it was switched on at scale.

StackAdapt announced on May 5, 2026, that it was joining the technology partner roster alongside Criteo, Kargo, Adobe, and Pacvue. The StackAdapt announcement confirmed that OpenAI supports two parallel buying pathways: self-serve through the Ads Manager, and managed access via technology partners who handle budgeting, bidding, and creative while OpenAI retains control over delivery decisions.

Technical architecture of the ad platform

The campaign structure in the self-serve Ads Manager follows a three-tier hierarchy. A Campaign defines objective, budget, dates, and targeting parameters. An Ad Group organizes ads around a theme or intent area, and it is at this level where advertisers supply "context hints" - phrases describing the types of conversations or topics where their products may be relevant. An Ad is the individual creative unit: title, description copy, image asset, and landing page URL.

The platform supports two buying models simultaneously. The CPM model - with a default maximum bid of $60 - targets reach objectives. The CPC model - where OpenAI recommends starting bids of $3 to $5 per click - targets clicks objectives. The underlying auction uses a relevance-weighted, second-price mechanism. OpenAI has said it will add further bidding models over time.

The Conversions API, launched May 5, allows advertisers to send server-side signals back to OpenAI from their own infrastructure. The pixel-based tool handles client-side tracking. Together, they provide a mechanism to close the attribution loop between a ChatGPT ad click and downstream actions on advertisers' own properties. A community-maintained Google Tag Manager template for the OpenAI Ads Measurement Pixel was published on May 7, 2026, two days after the pixel launched, to reduce the implementation burden for teams without dedicated engineering.

The self-serve interface is accessible at ads.openai.com. Businesses sign up, complete a verification process that requires a Business ID and physical address, and wait for OpenAI's team to approve the account before campaigns can go live. According to Jon Loomer, writing in his Meta Advertiser Field Notes blog on May 7, 2026, demand for new accounts is running high: "I've been waiting several hours so far," he wrote, describing his own verification experience. OpenAI's own description of the process - "quick and easy" - appears to hold for the steps themselves. The review queue is the variable.

How custom audiences fit into the broader stack

The custom audience upload flow Kaziukenas documented is separate from the contextual targeting that has been the primary differentiator for ChatGPT as an ad platform. Since launch, OpenAI has argued that its inventory carries a premium because users encounter ads during active, decision-oriented conversations. The platform matches ads based on the subject of the current conversation, past chat history, and previous interactions with ads - a contextual signal set that differs structurally from keyword matching or behavioral profiles accumulated across browsing sessions.

Custom audiences sit on top of that layer. They introduce a deterministic filter - either "only show ads to these specific people" or "do not show ads to these specific people." That has two immediate practical uses. The first is retargeting: reaching users who have already interacted with a brand, visited its website, or expressed interest through a previous touchpoint. The second is suppression: excluding existing customers, users in active CRM flows, or audiences a campaign is not designed to convert.

The suppression case is arguably the more operationally significant in the near term, particularly for advertisers who have already built exclusion logic into campaigns on other platforms. One practical gap on those platforms is that custom audience windows have limits. On Meta, for instance, the standard maximum retention for website custom audiences based on the purchase event has been 180 days - a constraint Meta is extending to 730 days starting May 18, 2026, as noted in the same week's field observations. Longer windows make exclusions more complete and therefore more reliable.

For ChatGPT, the equivalent constraint set is still being defined. The upload limit of 512MB for the audience file is technically generous - a CSV of SHA-256 hashed emails at roughly 65 bytes per row can hold approximately 7.8 million identifiers within that limit. Whether processing happens in real time or in batch, and what match rates against ChatGPT's logged-in user base look like in practice, remain open questions.

The absence of lookalike audiences is worth noting. On Meta, lookalike audiences were the feature that followed custom audiences historically and dramatically expanded the addressable reach for performance advertisers. They allow a platform to find users statistically similar to a seed list - extending the custom audience logic from a finite set of known users to a probabilistically matched population. OpenAI's platform does not offer this yet, though the progression is visible in the documented history of how similar platforms have developed.

Meta's parallel changes

The same week that OpenAI's custom audience feature surfaced, Meta announced that the maximum audience retention for purchase events would increase from 180 days to 730 days, effective May 18, 2026. Existing custom audiences built on the purchase event with a 180-day window will automatically update to 730 days unless advertisers opt out before that date.

Separately, Meta has been rolling out its Ads MCP server - a tool that allows Claude and ChatGPT to connect directly to advertisers' Meta ad accounts and perform campaign management tasks through natural language. The open beta launched April 29, 2026. According to Jon Loomer's May 7 field notes, access to the Ads MCP is being rolled out gradually and is not yet enabled for all ad accounts, even for advertisers who have completed the setup process. The message Claude returns when attempting to use it reads: "Ads MCP is gradually being rolled out. Please check back at a later date to use Ads MCP with this Ad Account."

These two threads - OpenAI building audience infrastructure from scratch, and Meta opening its ad system to AI agents - reflect how the platforms are moving toward each other. OpenAI is adding the targeting capabilities that Meta has had for years. Meta is integrating the AI interfaces that OpenAI represents. The outcome for advertisers depends on how quickly each platform builds the parts it is currently missing.

Partner request scams targeting advertisers

Also circulating in the same week: a wave of partner request scams targeting Meta advertisers. Jon Loomer, in his May 7 field notes, described receiving emails sent legitimately from Meta's infrastructure - not spoofed addresses - requesting that he add "Agency Pro Network" as a partner to his business portfolio. The emails include a hyperlinked URL embedded in the business portfolio name field, which Meta then renders as a live link in its outbound emails.

The mechanism exploits a gap in Meta's partner request system. The emails are genuine - they originate from Meta's servers - but the requests themselves are initiated by scammers who have registered a business portfolio with a URL embedded in the name. According to Loomer, Meta includes warnings about fraud within the same emails, clarifying that the request is not affiliated with Meta, but the embedded link remains active. He described the fix as straightforward: "The first thing Meta could do is deactivate links that appear in business portfolio names." Advertisers who receive such emails are advised to navigate directly to their Meta Business Suite rather than clicking any link in the email.

Timeline

Summary

Who: OpenAI, through its ChatGPT advertising platform, is rolling out a custom audience feature for advertisers. The development was observed and shared publicly by Juozas Kaziukenas, an entrepreneur. Jon Loomer, a Meta advertising specialist writing in his Field Notes blog on May 7, 2026, documented related updates to Meta's ad infrastructure in the same week.

What: OpenAI's ChatGPT Ads platform is adding a custom audience feature that allows advertisers to upload CSV or TXT files - up to 512MB - containing raw or SHA-256 hashed emails and phone numbers. The feature supports both include (show ads only to listed users) and exclude (suppress ads from listed users) configurations. OpenAI normalizes identifiers during processing. No lookalike audience functionality is present yet. Separately, Meta is extending the maximum retention window for purchase event custom audiences from 180 days to 730 days starting May 18, 2026.

When: The custom audience feature was observed as of May 14, 2026, in a gated rollout. The Meta purchase retention change takes effect May 18, 2026. The broader context of OpenAI's self-serve Ads Manager opening to all US businesses occurred on May 5, 2026.

Where: The OpenAI Ads Manager is accessible at ads.openai.com and currently operates in the United States for ChatGPT Free and Go tier users. Meta's changes apply globally to advertisers using website and app activity custom audiences with purchase events.

Why: Custom audiences allow advertisers to bring their own first-party data into the targeting layer - enabling retargeting of known prospects and suppression of existing customers. For OpenAI, the feature closes a gap relative to established platforms and makes the ChatGPT ad channel viable for performance advertisers whose workflows depend on CRM integration and audience matching. The broader rollout of self-serve tools, CPC bidding, and measurement infrastructure signals that OpenAI is building the foundational stack required to compete at scale in a market currently dominated by Google, Meta, and Amazon.

Share this article
The link has been copied!