OpenAI 2 days ago updated detailed technical documentation confirming that ChatGPT Ads Manager blocks any Custom Audience from being used in a campaign until it reaches 25,000 matched users, with every uploaded file capped at 500,000,000 bytes, or roughly 500 MB.

The Help Center article, timestamped as updated two days prior to today's publication, lays out for the first time in an official OpenAI document the operational rules governing a feature that independent observers first spotted inside the Ads Manager interface in mid-May. It arrives at a moment when the feature itself appears to be moving through a second, broader stage of rollout, following screenshots that surfaced in early July showing an "Audiences" option appearing more widely across advertiser accounts.

For advertisers who have watched Custom Audiences exist only as scattered LinkedIn screenshots and third-party trade reports since May, the documentation is the first primary-source account of how the feature actually behaves: what file formats it accepts, how long processing takes, what happens when an upload fails, and how bid multipliers interact with audience membership at the ad group level.

What the documentation discloses

According to OpenAI's Help Center article, custom audiences are built from lists of customers or prospects using either email addresses or phone numbers, and they can be applied in three distinct ways once created. Campaign-level inclusion restricts delivery to people who appear in one or more selected audiences, a mechanism the documentation frames as useful for reaching known customers or qualified prospects. Campaign-level exclusion works in the opposite direction, preventing delivery to people already inside a chosen audience, which the article describes as a tool for avoiding existing customers or recent purchasers. The third mechanism, ad-group audience bid adjustments, lets an advertiser raise or lower an ad group's maximum bid specifically when the viewer matches a selected audience.

The file preparation requirements are specific. Advertisers must prepare a valid UTF-8 CSV or TXT file, and OpenAI's documentation states that a UTF-8 byte order mark is accepted. Uploads can contain up to 5,000,000 identifiers in a single file. Only one identifier type is permitted per upload; the documentation instructs advertisers not to mix emails, phone numbers, hashed emails, or hashed phone numbers within the same file. TXT files must carry one identifier per line, while CSV files can include a header row or none at all. If a header is used, according to the documentation, the column must be named exactly one of four values: email, phone_number, email_sha256, or phone_number_sha256.

Once uploaded, an audience cannot be edited. The documentation states plainly that to change an audience list, an advertiser must create a new audience and archive the old one rather than modifying the existing file in place.

Processing times and audience statuses

After a custom audience is created, Ads Manager processes the file before the audience becomes usable. According to the documentation, processing usually takes about 20 to 30 minutes, though the exact duration can vary depending on file size. The uploaded row count and the final matched audience size are not guaranteed to be identical. Invalid values, duplicate entries, or identifiers that fail to match will not count toward the total, meaning a file with several million rows may still resolve to a considerably smaller usable audience.

Five statuses can apply to an uploaded audience, according to the documentation. A Processing status means the file is still being handled and the audience is not yet ready. A Ready status confirms the audience has been processed, has cleared the 25,000-user minimum, and can be applied to campaigns or ad groups. A Too small status appears when the audience falls short of that 25,000-user floor, which the documentation states makes it unusable regardless of how many rows were originally uploaded. A Failed status indicates the file could not be processed at all, and the documentation advises checking that the file is valid UTF-8, uses only one identifier type, and meets the size and formatting requirements before attempting a re-upload. Intermediate statuses such as Upload pending, Indexing, or Publishing also appear during preparation, according to the documentation, and audiences in any of those states remain unusable until they resolve to Ready.

OpenAI's documentation separately recommends that advertisers work with audiences of at least 100,000 users, a figure higher than the hard 25,000 minimum, though the article does not explain the reasoning behind the larger recommended size.

Bid multipliers and audience layering

At the ad group level, advertisers can apply a bid multiplier tied to a specific custom audience. According to the documentation, the multiplier can range from 0.1x to 10x, and it adjusts the ad group's maximum bid upward or downward specifically when the viewer belongs to the selected audience. The documentation offers illustrative figures rather than fixed recommendations: a 2x multiplier for high-value customers, a 5x multiplier for people who have previously engaged with a brand, or a 0.5x multiplier for lower-priority audiences.

Where a viewer qualifies for more than one selected audience carrying different multipliers, OpenAI's documentation states that the highest matching multiplier applies. Bid multipliers, the article clarifies, do not affect eligibility. They change how competitively an ad group bids once a viewer is already eligible to see it; they do not determine whether that viewer can see the ad group's ads in the first place.

Audiences can also be combined for inclusion and exclusion within the same campaign. According to the documentation, ads can be restricted to people in at least one included audience, shown broadly except to people in an excluded audience, or both at once, provided the audience remaining after exclusions still clears the 25,000-user minimum. A single custom audience cannot be set as both included and excluded within the same campaign, the documentation notes.

Identifier formatting and retention

The documentation sets exacting formatting rules for each identifier type. Email addresses must contain exactly one @ symbol; uppercase letters are accepted but normalized to lowercase, leading or trailing spaces are stripped, and spaces within the address itself are rejected outright. Phone numbers must be submitted in E.164 format, including the plus sign and country code; spaces, dashes, parentheses, and periods are removed during normalization, but a phone number lacking a country code will not be accepted.

For advertisers using hashed identifiers, the documentation specifies that a SHA-256 hashed email address must be generated from the lowercase, trimmed version of the address, and the resulting value must be a 64-character SHA-256 hex digest. The same rule applies to hashed phone numbers, which must be hashed from the normalized E.164 format including the country code, again producing a 64-character hex digest. OpenAI's documentation adds a caution for advertisers exporting identifier lists from other tools: all uploaded files must be valid UTF-8, and a file exported from another platform should be re-saved as UTF-8 before it is uploaded to Ads Manager.

On retention, the documentation is direct. Uploaded audience files are deleted after processing completes, usually no later than 24 hours after the file was uploaded. The audience itself, once matched, persists for use in campaigns, but the underlying raw file an advertiser submitted does not remain on OpenAI's systems beyond that window.

Context: a feature independent observers spotted two months ago

Custom Audiences did not appear inside ChatGPT Ads Manager for the first time this week. PPC Land reported in Maythat entrepreneur Juozas Kaziukenas had shared screenshots on LinkedIn showing a gated rollout of the feature on May 14, 2026, nine days after OpenAI opened its self-serve Ads Manager beta to all United States businesses. At that stage, the feature allowed advertisers to upload CSV or TXT files of up to 512MB containing raw or SHA-256 hashed emails and phone numbers, a figure that aligns closely with the 500,000,000-byte ceiling OpenAI's documentation states today, once the difference between decimal and binary megabyte conventions is accounted for.

The May sighting placed the feature inside a rapid sequence of product releases. OpenAI's self-serve Ads Manager beta opened to all US businesses on May 5, 2026, the same announcement that introduced cost-per-click bidding alongside the existing cost-per-mille model and formally launched a Conversions API and pixel-based measurement tool. PPC Land's coverage of that launch noted that the pixel and Conversions API addressed what had been, as PPC Land reported in April 2026, one of the most persistent gaps in the platform: the absence of any standard attribution infrastructure that would let advertisers measure conversions with precision. Nine days after that opening, Custom Audiences became the next layer added to the same interface.

The pace did not slow after May. PPC Land documented that on May 22, 2026, OpenAI added daily budget pacing, geo-targeting by US state, designated market area, and ZIP code, plus aggregate totals in list views and an early test of dynamic call-to-action buttons. By June 5, conversion-optimized campaigns began rolling out to advertisers who had at least one conversion event configured through the Pixel or Conversions API by June 1, a hard eligibility cutoff OpenAI communicated to beta advertisers by email on May 30. Within a single 26-day window, the platform had accumulated self-serve access, CPC bidding, a Conversions API, custom audience targeting, daily budgets and geo-targeting, and conversion optimization, six distinct additions in less than a month.

Geographic expansion followed the feature buildout. PPC Land reported that ChatGPT advertising went live in the United Kingdom on June 6, 2026, the first European market to receive the pilot, and subsequently in Japan and South Koreaalongside a UK self-serve Ads Manager expansion. By early June, OpenAI had also migrated product feed managementinto a dedicated section of the Ads Manager, supporting up to 1 million SKUs per advertiser, a feature carried over from the platform's now-defunct instant checkout infrastructure.

A second wave of visibility in July

The Custom Audiences feature resurfaced in trade coverage roughly a month after its initial May sighting. On July 7, 2026, Craig Graham, CEO of Grayvault Consulting, identified what PPC Land's coverage of the Criteo takeover biddescribed as a new ChatGPT Ads feature letting advertisers upload custom audience lists under an "Audiences" tag, a sighting separately reported by Search Engine Roundtable the same day. Graham's observation carried a caveat: match rates on uploaded lists, he cautioned, remain partial rather than complete, a limitation that mirrors what any deterministic list-matching system faces when a raw email or phone number fails to correspond to an active account on the receiving platform.

Separate PPC Land coverage published this week framed the July sighting in comparative terms, describing customer-list targeting as the connective tissue of mature ad platforms and citing Google's Customer Match and Meta's Custom Audiences as the direct precedents. That coverage noted that the demand side has not caught up everywhere the supply side has expanded: Adthena logged zero paid ChatGPT placements across 169,560 result scrapes in the United Kingdom during June, even as Google's AI Overviews appeared on 23 percent of tracked UK queries over the same period, a gap PPC Land reported drawing on Adthena's underlying data.

Today's Help Center documentation is best understood against that backdrop. It does not announce a new capability. It formalizes, in OpenAI's own words, the mechanics of a feature that has existed in some form since mid-May and that a second cohort of advertisers appears to be discovering only now, in July, as the wider rollout continues. The specific figures disclosed today, the 25,000-user minimum, the 5,000,000-identifier cap, the 20-to-30-minute processing window, and the 0.1x-to-10x bid multiplier range, had not previously appeared in OpenAI's own public documentation, even though the feature itself had been visible in screenshots for two months.

Why the details matter for advertisers

The 25,000-user floor is a meaningful constraint for smaller advertisers. A local or regional business with a customer list well under that threshold cannot use Custom Audiences at all, regardless of how well its file matches OpenAI's formatting rules. The documentation's separate recommendation of 100,000 users as a practical working size widens that gap further, suggesting OpenAI expects reliable matching and delivery to require a considerably larger seed list than the bare technical minimum.

The distinction between the 25,000-user floor and match-rate variability, the caveat Graham raised in July, also matters in practice. A file containing 30,000 raw rows is not guaranteed to produce a 30,000-person matched audience. Invalid entries, duplicates, and identifiers that fail to resolve against an active ChatGPT account all fall away before the count that determines Ready, Too small, or Failed status is calculated. An advertiser working close to the minimum threshold has less margin for a high failure rate in their source data than one working with a list well above it.

The absence of an editing mechanism, an audience cannot be modified after creation and must instead be recreated and the old version archived, differs from list-management workflows on some other platforms and adds a small operational step every time a customer list changes. Combined with the file deletion window, uploaded files are typically gone within 24 hours, advertisers effectively get one attempt per version of a list before needing to prepare and re-upload a fresh file if changes are required.

The bid multiplier documentation clarifies a distinction that had not been explicit in earlier trade reporting: multipliers affect competitiveness, not eligibility. An advertiser cannot use a multiplier to make an ineligible viewer eligible, nor can a low multiplier remove an eligible viewer from consideration. That separation between targeting and bidding logic follows a structure familiar from other platforms' audience systems, though the specific mechanics, including the rule that the highest matching multiplier applies when a viewer qualifies for several audiences, are OpenAI's own implementation choices as described in today's documentation.

For a comparatively young advertising platform still building out its measurement and targeting stack, publishing granular, rule-by-rule documentation of a feature two months after independent observers first spotted it suggests OpenAI is treating the July resurgence of interest as an opportunity to formalize expectations before a broader set of advertisers begins relying on the feature at scale.

Timeline

  • May 5, 2026 - OpenAI opens its self-serve Ads Manager beta to all US businesses, introducing CPC bidding, a Conversions API, and pixel-based measurement, while removing the minimum spend requirement entirely.
  • May 14, 2026 - Custom Audiences targeting is first spotted inside ChatGPT Ads Manager in a gated rollout, documented through screenshots shared by entrepreneur Juozas Kaziukenas.
  • May 22, 2026 - Daily budgets, US geo-targeting by state, designated market area, and ZIP code, and a dynamic call-to-action test are added to the Ads Manager.
  • May 30, 2026 - OpenAI notifies beta advertisers by email that conversion-optimized campaigns will begin rolling out June 5, with a June 1 cutoff for early-access eligibility.
  • June 5, 2026 - Conversion-optimized campaigns begin rolling out to eligible advertisers.
  • June 6, 2026 - ChatGPT advertising goes live in the United Kingdom, the first European market activation.
  • July 7, 2026 - Craig Graham, CEO of Grayvault Consulting, and Search Engine Roundtable separately document a wider sighting of the Audiences feature appearing under an "Audiences" tag, with match rates described as partial.
  • July 12, 2026 - OpenAI's Help Center article on Custom Audiences shows a most-recent update timestamp, according to the document's own metadata.
  • July 14, 2026 - OpenAI's Custom Audiences documentation, detailing file limits, processing times, audience statuses, and bid multiplier rules, is examined in this report.

Summary

Who: OpenAI published the documentation, addressed to advertisers using ChatGPT Ads Manager. The feature itself was first identified publicly by entrepreneur Juozas Kaziukenas in May and again by Craig Graham, CEO of Grayvault Consulting, in July.

What: OpenAI's Help Center article details the operating rules for Custom Audiences, including a 25,000-matched-user minimum before an audience can be used, a 500,000,000-byte file size cap, a 5,000,000-identifier limit per upload, a 20-to-30-minute typical processing window, five distinct audience statuses, and ad-group bid multipliers ranging from 0.1x to 10x.

When: The documentation carries an update timestamp of two days before today, July 14, 2026. The underlying feature was first spotted in a gated rollout on May 14, 2026, and again in a wider sighting on July 7, 2026.

Where: The feature applies within OpenAI's ChatGPT Ads Manager, the self-serve advertising interface that opened to all US businesses on May 5, 2026, and has since expanded to the United Kingdom, Japan, and South Korea.

Why: The documentation matters because it is the first primary-source account of technical rules that had previously circulated only through screenshots and secondhand trade reporting, arriving as a second wave of advertisers appears to be encountering the feature for the first time, roughly two months after its initial appearance.