Most new ad platforms promise reach and deliver excuses. OpenAI is trying a different approach — building the measurement infrastructure before asking for serious budget commitments. That's a meaningful strategic shift, and performance marketers should pay attention to the mechanics, not the hype.
The gap between "experimental AI ad channel" and "line item in the media plan" has always been attribution. You can tolerate fuzzy brand metrics in an awareness channel. Direct-response budgets demand proof of conversion. With product carousels and an AppsFlyer attribution integration now live, OpenAI has at least addressed the most obvious structural objection to taking ChatGPT ads seriously.
What Actually Changed — and What Hasn't
Let's be precise about what OpenAI shipped. First, product carousels: ChatGPT can now display multiple products from a retailer's catalog at the bottom of a conversation, using the automated product feeds OpenAI introduced roughly three months ago. The format is algorithmically determined — OpenAI decides whether a user sees a single product ad or a carousel, not the advertiser. Second, AppsFlyer attribution: roughly 40 brands including Grubhub are now testing a direct integration that connects ChatGPT ad exposures to app installs, in-app purchases, and subscription conversions, comparable to how they'd measure any other paid acquisition channel.
Taken together, these additions give ChatGPT the functional skeleton of a performance advertising platform: a structured product data layer, dynamic creative assembly, and third-party conversion measurement. That's a significant infrastructure build in a short window.
What hasn't changed is equally important to flag. The AppsFlyer integration doesn't measure incrementality. It tells you whether someone who saw a ChatGPT ad eventually converted — it doesn't tell you whether that conversion would have happened anyway through organic search, a competitor's retargeting pixel, or any other touchpoint. For any performance marketer running multi-channel attribution models, that's a material limitation. Last-touch or even assisted-attribution data from a new channel with limited historical benchmarks is interesting signal, not decision-grade proof.
Advertiser control over creative presentation is another unresolved variable. The platform decides carousel versus single-product format. Until OpenAI publishes clear documentation on what signals drive that decision — user query type, product catalog structure, session context — performance teams are essentially optimizing blind on a key creative dimension.
The Performance Marketer's Checklist Before Shifting Spend
Here's where the practical question gets interesting: at what point does ChatGPT advertising graduate from test budget to meaningful allocation?
The infrastructure OpenAI has built mirrors what Google Shopping and Meta Advantage+ established as table stakes for performance channels. Product feeds powering dynamic creative is table stakes. Third-party attribution integration is table stakes. What remains unproven is the layer that actually justifies budget scale: cost-per-acquisition at volume.
A channel can attribute conversions and still be economically nonviable if CPAs run 3x higher than established channels or if daily conversion volume is too thin to power algorithmic optimization. Google Performance Max and Meta Advantage+ work at scale partly because their AI optimization loops have hundreds of millions of conversion signals to learn from. ChatGPT's ad inventory is comparatively nascent — the learning curves will be steeper, and early CPA data may not reflect long-term performance.
Before moving meaningful direct-response budget to ChatGPT, performance teams should demand answers to these specific questions:
- What are the attributed CPA benchmarks from the AppsFlyer beta? Grubhub and the other ~40 testing brands should have preliminary data by Q4. Push for category-level benchmarks, not just aggregate platform claims.
- What conversion volume is achievable at target CPAs? Scale constraints matter. A channel that converts efficiently at $500/day but degrades significantly at $5,000/day isn't a real acquisition channel.
- How does ChatGPT attributed conversion data hold up against incrementality testing? Run a geo-based or audience holdout test alongside any AppsFlyer integration. Don't let attribution substitute for incrementality measurement.
- What creative controls will be available? Before Q4, advertisers need clarity on how product selection for carousels is determined, what data surfaces in reporting at the product level, and what bid and budget controls exist.
- What are the audience intent signals? ChatGPT users in a shopping context may carry different purchase intent than Google Shopping users. That's potentially a quality advantage — or a reach limitation. You need enough data to distinguish between the two.
What the Q4 Test Will Actually Reveal
OpenAI is reportedly providing advertisers with more detailed guidance on feed-based campaigns ahead of the holiday shopping season. That makes Q4 2026 the real proof-of-concept moment. Holiday retail is when performance budgets are at their most aggressive and when CPA benchmarks matter most — it's the hardest test environment and also the most revealing one.
If ChatGPT ads can demonstrate competitive CPAs for e-commerce and app install campaigns during Q4, even within a limited advertiser beta, the platform's trajectory changes materially. AI-native channels have a structural advantage in high-intent conversational contexts that traditional search ad formats weren't designed for — when a user asks ChatGPT for "the best running shoes under $150 for wide feet," the purchase intent signal embedded in that query is exceptionally precise. The question is whether OpenAI's optimization systems can translate that signal quality into conversion efficiency at the CPAs performance budgets require.
The honest position for most performance teams right now: allocate test budget, instrument it properly with independent attribution verification, and don't move meaningful spend until you have your own CPA data — not OpenAI's. The infrastructure is now good enough to test rigorously. It isn't yet proven enough to scale confidently.
OpenAI has built the machinery. Q4 is where it either works or it doesn't — and performance marketers who set up clean measurement frameworks now will be the first to know which it is.
