OpenAI's feed pivot: two pipelines for ChatGPT product visibility

OpenAI’s commerce infrastructure now runs two architecturally separate pipelines: an organic feed for unpaid product discovery in ChatGPT, and a paid ads pipeline inside Ads Manager. Understanding how they differ and where they do not overlap is the starting point for any merchant thinking seriously about ChatGPT as a distribution channel.

Two pipelines, one lever: is_eligible_search for organic discovery and is_ads_eligible for the new product-feed ads beta.

When OpenAI moved product feed management into a dedicated Feeds section inside ChatGPT Ads Manager on June 2, 2026, it formalized what had been building quietly since the Agentic Commerce Protocol (ACP) spec was published in late May. The two pipelines now have separate submission paths, separate gating fields, and separate eligibility rules. Products submitted only through the Ads Manager feed do not appear in organic ChatGPT conversations. That separation is not a bug in the rollout; it is the design.

What actually happened with Instant Checkout, and why did the feed infrastructure grow from it?

OpenAI launched “Buy it in ChatGPT” Instant Checkout around February 16, 2026, enabling U.S. ChatGPT Plus, Pro, and Free users to purchase products directly inside the chat interface via Stripe. The launch touched a real problem: ChatGPT was already processing roughly 50 million shopping-related queries daily⁷, according to figures OpenAI circulated at the time, against a base of 700 million-plus weekly active users.⁸

Walmart disclosed conversion rates running at approximately 3x lower than click-out purchases. Shopify president Harley Finkelstein publicly identified the structural problem⁹: taxes, inventory, and subscriptions are difficult to route through a single in-chat checkout layer. OpenAI also had not built a system for collecting and remitting U.S. state sales taxes, a live regulatory gap. By March 2026, Instant Checkout was retired. At the point it was pulled, approximately 30 Shopify merchants had gone live with it.¹⁰

The pivot was not a retreat from commerce. OpenAI moved to dedicated retailer apps inside ChatGPT, naming Target, DoorDash, Instacart, and The Knot as initial partners, with Etsy and Shopify among the early ACP participants. The ACP infrastructure itself, the merchant-push feed model and the open protocol, expanded rather than contracted. The checkout complexity problem was handed to app-level integrations; the product discovery problem was handed to the feed pipelines.

What is the ACP organic pipeline, and how does it differ from the ads pipeline?

The Agentic Commerce Protocol is an open standard co-developed with Stripe for connecting buyers, AI agents, and merchants to complete purchases. It was released under the Apache 2.0 license in September 2025.¹ PayPal joined as a second ACP-compliant payment provider in October 2025, and Stripe shipped its Agentic Commerce Suite in December 2025.

The ACP organic feed is a merchant-push model: SFTP, direct file upload, or hosted URL. Product data is not crawled. Merchants push a structured file to a secure OpenAI endpoint, with updates accepted as frequently as every 15 minutes.⁵ Supported formats include JSONL (gzip), CSV (gzip), TSV (gzip), and Parquet (zstd); UTF-8 encoding is required. This pipeline powers unpaid product discovery inside ChatGPT conversations.

The Ads Manager feed pipeline is architecturally separate. It was first publicly observed by entrepreneur Juozas Kaziukenas when OpenAI moved it into a dedicated Feeds section of Ads Manager on June 2, 2026.¹¹ The ads beta requires SFTP upload (no API currently), a minimum of 1,000 products, a maximum of 2 million, and a 100-product sample before full catalog activation.⁴ Conversion-optimized campaigns rolled out June 5, 2026 for accounts with at least one conversion event flowing.¹² Cost-per-action bidding began rolling out in early June 2026.

Dimension	Organic ACP Feed	Ads Manager Feed
Gating field	is_eligible_search	is_ads_eligible
Submission method	SFTP, file upload, hosted URL	SFTP only (beta)
Update cadence	Up to every 15 minutes	Not specified for beta
Catalog range	Not publicly floored	1,000 min / 2 million max
Appears in organic chat?	Yes, when field is true	No, ads-only pipeline
Developer docs path	developers.openai.com/commerce	developers.openai.com/ads/product-feeds

What do the three ACP boolean flags actually control?

The ACP product feed specification gates product visibility through three boolean fields. is_eligible_search is required and controls whether a product appears in organic ChatGPT search discovery. is_eligible_checkout is required and controls in-chat purchasing eligibility. is_eligible_ads is optional within the ACP spec itself and controls paid ad placement on that side of the protocol.

The Ads Manager pipeline uses is_ads_eligible as its canonical field name.³ The legacy alias is_eligible_ads is also accepted per the ads-specific developer documentation at developers.openai.com/ads/product-feeds. The ACP spec documents is_eligible_search as requiring lowercase true/false strings; a merchant serializing Python booleans without explicit string conversion will produce True with a capital T, which is a silent eligibility failure.²

My read on the field-naming split: the Ads Manager pipeline’s canonical field name diverges from the ACP organic spec’s naming pattern deliberately, reinforcing the architectural separation between the two pipelines. Whether that naming was planned or accumulated through parallel team development, the practical effect is the same: a merchant treating these as the same feed with the same field names will submit products that are invisible in one pipeline or the other.

What does feed hygiene specifically buy you inside ChatGPT?

ChatGPT Shopping Research, which launched November 24, 2025⁶ for all logged-in users including the free tier, uses a reinforcement-trained mini variant of GPT-5 tuned for retail product evaluation. OpenAI’s internal evaluation at launch put it at 52% product accuracy on multi-constraint queries versus 37% for standard ChatGPT Search.¹³ The model is making structured comparisons. What it compares is what the feed provides.

The organic pipeline is a merchant-push model, not a crawl. The model does not supplement missing feed fields by scraping a product page. If a product’s price field is stale, the comparison surface is stale. If availability does not reflect actual inventory, the model may surface out-of-stock items. If is_eligible_search is set to false across a product category because a feed export misconfigured defaults, that category does not appear in ChatGPT organic discovery at all, regardless of how well the products would rank.

The inference here, and I label it as one: feed hygiene in the ACP context carries higher stakes per error than in Google Shopping, because there is no crawl fallback and no editorial override. A Google Shopping feed error affects a product’s bidding eligibility or attribute completeness; the underlying page may still rank organically. In the ACP pipeline, the feed is the product record. There is no other layer.

What should merchants actually do with this information?

The two pipelines require separate strategic decisions. For organic visibility, the ACP feed is the mechanism; is_eligible_search must be explicitly set to true (lowercase string) at the product level. A feed that carries default or null values for this field is not an organic presence in ChatGPT, regardless of catalog size. Update frequency matters: the 15-minute cadence is available; whether it is necessary depends on how frequently pricing and inventory move.

For the ads pipeline, the beta threshold of 1,000 products with a 100-product sample review means the entry point is a deliberate catalog submission, not an automatic extension of an existing Shopping feed. The is_ads_eligible field governs paid placement; it does not backfill into the organic ACP pipeline. These require separate file submissions to separate endpoints.

The ACP spec published May 25, 2026 carries no version number. OpenAI’s developer documentation for both pipelines is actively evolving through the beta period. Any implementation decision made against the current spec should carry a review date, not an assumption of permanence.