ChatGPT Ads Now Eats a 2-Million-Item Feed. Here's the One Claude Builds for You Tonight

Until last week, running a ChatGPT ad meant building objects one product at a time. On June 11 OpenAI switched on product-feed campaigns in Ads Manager, and the math changed: you can now pour an entire catalog, anywhere from 1,000 to 2,000,000 items, into the system with a single upload. ChatGPT ads only went live February 9, so this is a four-month-old channel that just grew the one feature performance marketers actually run on.

Here is the part that rewires the workflow. The ad title and description are pulled straight from the feed. There is no separate creative step. Your feed copy is the ad copy. Whoever ships the cleanest, best-written catalog wins the high-intent shopping conversation, and the whole job is feed hygiene you can finish tonight with Claude instead of waiting on a developer.

The feed is the creative, and it does not behave like Google or Meta

OpenAI's file-upload spec runs about 67 fields across 15 groups: 15 required, a dozen conditional, the rest optional. The required core is familiar to anyone who has touched a Shopping feed: item_id, title, description, url, brand, image_url, price, availability, seller_name, seller_url, and country codes. Two OpenAI-specific flags gate everything: is_eligible_search makes a product discoverable, and is_eligible_checkout (which needs the first) makes it buyable inside chat.

The quirks are where feeds get rejected. Title caps at 150 characters and cannot be all-caps. Description caps at 5,000 characters and must be plain text and factual: OpenAI explicitly does not want marketing language, and HTML gets stripped on ingestion. Availability is an enum, not free text (in_stock, out_of_stock, pre_order, backorder, unknown). Set a sale_price higher than price and the item is rejected silently, with no error returned. A single 404 on your url or image_url kills that product's eligibility. None of these are style notes. They are pass-fail.

That plain-text-factual rule is the trap for marketers raised on keyword-stuffed titles. The model ranks on relevance to a real question ("what trail running shoes handle wet rock"), so a precise, attribute-rich description beats a salesy one. That is exactly the rewrite Claude does well, at catalog scale.

Do it tonight: the build is feed hygiene, not an engineering ticket

Export what you already have (Shopify, Google Merchant Center, your Meta catalog) as CSV. The actual work is mapping your columns to OpenAI's schema, rewriting every title and description to spec, normalizing the enums, and catching the rows that would fail before you upload. That is a single pass for Claude, not a dev ticket.

We packaged that pass as a ready Claude skill. It maps your export into a spec-compliant feed and hands back a QA report that flags every blocker by row: missing required field, bad enum, sale_price above price, non-HTTPS URL, title over the limit. Take it, run your own catalog through it, then upload. Get the free ChatGPT Ads feed skill

However you build the file, validate on a small sample, around 100 items, before pushing the full catalog. Delivery is a daily full snapshot over SFTP. Parquet is preferred, but csv.gz, tsv.gz, and jsonl.gz are accepted, so a compressed CSV is a fine start. Keep the same filename and overwrite it each day. To pull a product, flip is_eligible_search to false or drop the row from the next snapshot.

After the feed validates, ChatGPT builds an ad per item and you set an ad template for the ad group to control the framing. Those ads surface two ways: a shopping carousel, and a conversational unit a shopper can question in line, so your description doubles as the answer to their follow-up prompts. If you also want in-chat checkout, is_eligible_checkout pulls in two more required fields, seller_privacy_policy and seller_tos, so add those URLs while you are already in the file.

Why bother this week instead of next quarter: OpenAI says feed ads are already among its top performers in beta. One global DTC brand doubled its click-through rate and cut its cost-per-click in half after moving from standard ChatGPT ads to feed ads. Right now the auction is thin and most catalogs are messy. Both close fast.

The work that decides whether you show up is not bidding. It is whether your feed is clean, factual, and live before your competitor's is. That is an evening with Claude, not a sprint with engineering.

Sources: OpenAI Developers (commerce file-upload overview and product feed spec); OpenAI Ads Manager rollout email via Marketing4eCommerce, Jun 12 2026; AIToolDiscovery, Jun 11 2026; aishoppingfeeds ACP field walkthrough.