OpenAI has added a platform targeting control to ChatGPT Ads Manager Beta, letting advertisers restrict campaign delivery to the iOS app, the Android app or the web. The company documented the control in a help centre article revised within the previous 24 hours, and listed it in a product update email circulated to beta advertisers during the week beginning Monday, August 17, 2026. Reporting has not moved with it: Insights still sorts device data into two categories, and mobile web sits inside the mobile one.

For six months, an advertiser buying a placement inside a ChatGPT conversation bought every surface the assistant runs on. Advertising began in the United States on February 9, 2026, and the rollout covered mobile and desktop from the first day. There was no way to buy one and not the other. The campaign object exposed an objective, a budget, a set of dates, a country list and, later, a geographic refinement inside the United States. Where the conversation was physically happening - an app on a phone, a browser tab on a laptop - was not an advertiser-controllable variable.

That has changed. According to OpenAI, the campaign dialog now carries a multi-select dropdown that governs which ChatGPT surfaces a campaign is eligible to deliver on.

Three values, one of which is compound

The supported platforms are iOS App, Android App and Web. The documentation is explicit that the third value is not a desktop value: Web covers both desktop web and mobile web.

That construction matters more than it first appears. An advertiser who wants to reach only people using a laptop cannot express that intention through this control. Selecting Web admits a phone browser as readily as a desktop one. The only clean separations available are between the two native applications and everything else, and between iOS and Android within the application layer.

The dropdown is a multi-select, so a campaign can name one, two or all three values. It sits alongside custom audiences in the documentation as an optional element of the campaign object, which places the default at unrestricted delivery. A campaign created without touching the field behaves as every campaign has behaved since February.

There is a naming inconsistency between the two places the control is described. The help centre article calls the field the Platforms multi-select dropdown. The product update email circulated to advertisers calls it the Eligible platforms multi-select dropdown. Whether the interface label matches one, the other or neither is not resolved by either document.

The reporting does not match the buying

The most consequential sentence in the documentation is a note appended to the platform targeting section. According to OpenAI: "Insights currently group device reporting into Mobile and Desktop, and mobile web is counted under Mobile. More detailed platform reporting is planned."

Read against the control it annotates, that note describes a mismatch. The buying interface offers three values. The reporting interface returns two. And the mapping between them is not one-to-one in either direction.

An impression served in the iOS app lands in the Mobile bucket. An impression served in the Android app lands in the same bucket. An impression served in a mobile browser - which the buying side classifies as Web - also lands in Mobile. Only a desktop browser impression reaches the Desktop row. A campaign that targets Web exclusively will therefore return Mobile rows, and a campaign that targets both native apps will return only Mobile rows and nothing else.

The practical result is that the new control can be set but its effect cannot be read back out of the platform's own reporting at the same resolution. An advertiser cannot use Insights to confirm how spend divided between the iOS app and the Android app, because both collapse into one figure. Nor can that advertiser separate mobile web performance from in-app performance, because the reporting boundary runs along the device axis while the targeting boundary runs along the application axis. The two lines cross.

OpenAI states that more detailed platform reporting is planned. No date accompanies that statement.

Where the control sits in the campaign object

Platform targeting is a campaign-level setting, not an ad group setting. It sits in the same tier as the objective, the budget, the start and end dates, the country list and the custom audience configuration. Context hints - the phrases that carry the entire targeting burden inside ChatGPT, since the system matches advertiser-supplied language against conversation subject rather than against keywords - remain one level down, at the ad group.

The placement has structural consequences that the campaign setup FAQ makes visible. According to the documentation, a campaign objective cannot be edited into a different objective after creation; the company directs advertisers to build a new campaign instead. Budget type changes are partially irreversible, with a campaign-total budget converted to a daily budget unable to convert back. The documentation does not state whether platform selection is editable after a campaign goes live, which leaves an open question about whether a surface split requires duplicate campaigns or can be applied to an existing one.

Duplicate campaigns carry a cost that is not merely administrative. The same article lists a Minimum Campaign Spend table with daily floors that vary by billing currency: 25 USD in the United States, 15 GBP in the United Kingdom, 25 CAD in Canada, 25 AUD in Australia, 25 NZD in New Zealand, 2,500 JPY in Japan, 25,000 KRW in South Korea, 40 BRL in Brazil, 150 MXN in Mexico and 725 INR in India. Splitting one campaign into three to isolate iOS, Android and Web means clearing that floor three times over rather than once. The India entry is itself notable, since OpenAI's activation of Brazil and Mexico earlier in August brought the publicly counted market total to nine, and India was not among them.

Measurement paths diverge by surface

The platform control lands on top of a measurement stack that already treats applications and browsers as separate problems.

Web conversions reach OpenAI through the JavaScript pixel and the Conversions API, both of which arrived on May 5, 2026 alongside the self-serve Ads Manager and cost-per-click bidding. LiveRamp connected its Conversions API Hub on June 10, the first third-party server-side route into the platform. App conversions travel a different road entirely: AppsFlyer and Adjust went live as integrated mobile measurement partners across seven markets in late July, carrying install attribution, re-engagement attribution and in-app event postbacks.

Until this week, an advertiser with only one of those two measurement paths in place was still buying both surfaces. A retailer running a pixel and no mobile measurement partner had impressions delivering inside the iOS and Android applications with no instrumented route back. A mobile-first business running AppsFlyer and no web pixel had the inverse gap. The platform control makes it possible to align the surface being purchased with the measurement infrastructure that actually exists behind it.

That alignment question sharpened on August 19, when OpenAI switched on view-through conversion reporting with a window fixed at one day, walled off from the main Conversions total and excluded from bidding, billing and cost per acquisition. A metric that credits impressions rather than clicks is sensitive to which surface produced the impression, and the reporting split does not currently expose that.

A non-personal lever arrives as Europe opens

Timing gives the control an additional dimension. The same product update email that documented platform targeting told advertisers that ChatGPT Ads would expand to 31 European countries the following week, and confirmed that Brazil and Mexico were already live. According to OpenAI, the week's updates were oriented toward "helping advertisers reach more markets, optimize toward campaign goals, manage delivery across surfaces, and better understand campaign performance."

European delivery operates under constraints that do not apply in the United States. OpenAI emailed European users on Saturday, August 15, 2026 confirming that advertising would reach the Free and Go plans later in the month, and the accompanying help centre material stated that personalised advertising is not initially available in the European Economic Area or Switzerland. Selection at launch draws on the current conversation topic plus general location and device type. Past chats and stored memories are excluded from that initial process.

Device type, in other words, is already among the handful of signals the selection system uses in Europe. Platform targeting hands part of that axis to the advertiser as an explicit setting rather than an inferred one - and it does so without relying on behavioural profiling, which places it inside the narrow set of levers that function identically on both sides of the personalisation line. For European buyers who were told that creative and context hints would carry the entire targeting burden at launch, the dropdown is one more variable that does not depend on a consent posture.

What the documentation does not say

Several things are absent from both source documents, and the absences are as informative as the additions.

There is no bid modifier. The control is an inclusion list, not a multiplier. Nothing in the documentation describes adjusting a bid upward or downward by platform, which is the mechanism Google and Microsoft advertisers have used for device management for well over a decade.

There is no operating system version targeting, no device model targeting and no distinction between tablet and phone. The three values are the whole surface taxonomy.

There is no statement about how platform targeting interacts with the three campaign objectives. CPM, CPC and oCPCcampaigns are all governed by the same campaign dialog, and conversion-optimised delivery depends on a tracked conversion event flowing through the pixel or the Conversions API. Whether an oCPC campaign restricted to the two native applications can optimise against a web conversion signal is not addressed.

And there is no reporting date. The note promising more detailed platform reporting carries no timeline, which leaves the mismatch between three buying values and two reporting values open for an unspecified period.

Why this matters for the marketing community

The channel is being assembled in public, one control at a time, and each addition removes a stated reason not to test it. Since the pilot began, ChatGPT Ads has acquired a self-serve interface with cost-per-click bidding, custom audience uploads, daily budgets and United States geographic targeting by state, designated market area and ZIP code, product feed campaigns supporting up to one million SKUs, conversion-optimised campaigns from June 5, and automatic advanced matching switched on by default for existing web pixels from August 17. Platform targeting is the surface-selection piece of a control set that most performance buyers would have considered table stakes on any other channel.

Demand data remains contested, which is what makes the control set worth watching rather than the spend figures. Similarweb measured sponsored placements in 26 percent of ChatGPT responses with a 0.50 percent click-through rate in a dataset published on August 17, 2026, alongside 7,378 distinct advertisers identified in a single week. Adthena counted the same 7,378 advertisers during the week of July 13 to 20, with the United States accounting for 60.1 percent of them, while recording zero paid ChatGPT placements across 169,560 United Kingdom result scrapes during June. Criteoreported more than 2,000 brands running campaigns through its API connection by June 2026.

Against that background, the specific value of platform targeting is negative rather than positive: it is a way to stop buying something. A brand with no mobile application and a checkout flow that behaves poorly in an in-app browser can now decline the two application surfaces. An application developer with no web funnel can decline the browser one. Neither exclusion was available a week ago.

What the exclusion cannot yet be paired with is proof that it worked. Three values go in. Two come back. The gap between them is the part of this release that has a plan attached but no date.

Timeline

Summary

Who: OpenAI, operator of ChatGPT Ads Manager Beta, communicating through a product update email to advertisers in the beta programme and through its public help centre. The change affects performance advertisers, media agencies and measurement teams running campaigns inside ChatGPT, and is most consequential for advertisers whose conversion infrastructure covers only one of the two surface types.

What: A platform targeting control added at campaign level, exposed as a multi-select dropdown with three supported values - iOS App, Android App and Web, where Web covers both desktop web and mobile web. The setting is optional, so campaigns created without it deliver across all surfaces. Reporting has not been extended to match: Insights groups device data into Mobile and Desktop, with mobile web counted under Mobile, and OpenAI states that more detailed platform reporting is planned without giving a date.

When: The control was listed in a product update email circulated to beta advertisers during the week beginning Monday, August 17, 2026, and documented in the Create Campaigns for ChatGPT Ads help centre article, which carried a revision timestamp placing the update within the previous 24 hours as of Friday, August 21, 2026. The same email confirmed Brazil and Mexico as live markets and scheduled a European expansion to 31 countries for the following week.

Where: ChatGPT Ads Manager Beta, reachable at ads.openai.com, across the markets where ChatGPT advertising currently operates. The documented minimum daily spend table covers Australia, Brazil, Canada, India, Japan, Mexico, New Zealand, South Korea, the United Kingdom and the United States.

Why: Advertising inside ChatGPT has been sold as a single undifferentiated surface since February 9, 2026, while the measurement infrastructure beneath it split along application and browser lines - pixel and Conversions API for web, AppsFlyer and Adjust for applications. Platform targeting lets advertisers align the surface they buy with the measurement they hold, and supplies a delivery lever that does not depend on behavioural personalisation, which is the constraint European campaigns operate under at launch. The reporting mismatch means the effect of that lever cannot yet be verified inside the platform's own numbers.