ChatGPT Product Feed Setup: Fields, Formats and SFTP Delivery

What a ChatGPT product feed is and who can submit one

A ChatGPT product feed is a structured catalog file that you send to OpenAI so ChatGPT can read your products, prices and stock status. Delivery is a push: you deliver files on a schedule instead of waiting to be crawled, and the quality of that file is what everything else depends on.

Access is the first gate. As of September 29, 2026, OpenAI's Get Started guide says product feed onboarding is available to approved partners only, and applications go through the merchant form at chatgpt.com/merchants. Apply first, then use the waiting time to clean your data, because most of the setup work is catalog hygiene that any feed will need.

There are two delivery routes, file upload and an API. OpenAI generally recommends sending the entire feed once a day by file upload and pushing updates through the day via the API. Small feeds can do everything through the API, and promotions data can only be sent through the API. This guide covers the file-upload route, checked against OpenAI's file upload overview.

ChatGPT product feed setup in 7 steps

1. Apply for access through OpenAI's merchant form and confirm your delivery details (SFTP location and, if you use them, Ads Manager or API options) during onboarding.
2. Pick the route: file upload, API, or both. A daily full snapshot plus intraday API updates is the pattern OpenAI recommends.
3. Map your catalog to the nine required fields (next section). One row per purchasable item or variant.
4. Export in a supported format: Parquet is preferred, and jsonl.gz, csv.gz and tsv.gz are also accepted. Encode everything as UTF-8.
5. Send a sample of roughly 100 items with every required field present, and read the upload results before scaling up.
6. QA the first full snapshot, not just the sample. Sample-only testing hides problems that appear at catalog scale, such as duplicate IDs and odd variants.
7. Automate the daily snapshot, overwriting the same file name each time.

The nine required fields for a ChatGPT product feed

OpenAI's product spec lists nine fields as required for a useful discovery feed. Use a real brand and seller name, never placeholders. Details below follow the OpenAI products spec.

A valid one-line JSONL record looks like this:

{"item_id":"GRIND-HAND-BLK","title":"Hand coffee grinder, matte black","description":"Manual burr grinder with a stainless steel burr and a 30 g glass jar.","url":"https://example.com/products/hand-grinder-black","brand":"Brewline","seller_name":"Brewline Coffee Gear","image_url":"https://example.com/images/hand-grinder-black.jpg","price":"64.00 USD","availability":"in_stock"}

One expert caution: the spec says format guidance is not a guarantee that every invalid value is rejected at upload. A file that uploads without errors can still carry bad data, so validate before you send, not after. A small checker is included further down.

File formats, SFTP delivery and file rules

Two behaviors matter in practice. First, the SFTP root directory represents your entire catalog and there is no completion marker: OpenAI processes every file in the root once uploads stop changing for a short time. Upload all shards back to back, because a long pause can trigger processing of an incomplete catalog (OpenAI then reprocesses once the rest arrive). Second, if several brand feeds share one location, prefix file names with the brand.

Some third-party guides describe delivery as an HTTPS push. OpenAI's own overview says SFTP, so follow that and confirm the exact endpoint with OpenAI during onboarding.

The heavy part is rarely the export itself. It is keeping one clean catalog feeding several channels at once, which is the point where handing over product feed management usually costs less than maintaining custom scripts per channel.

Variants, group IDs and GTIN validation

Send one row per selectable variant. Each row has its own item_id, the same group_id (which must differ from every item ID), listing_has_variations=true and a variant_dict of the selected options. Each row carries its own price, availability, URL and images.

{"item_id":"TEE-NAVY-M","group_id":"TEE","listing_has_variations":true,"variant_dict":{"color":"Navy","size":"M"},"offer_id":"brewline-TEE-NAVY-M","title":"Cotton tee, navy, size M","description":"Midweight cotton tee.","url":"https://example.com/products/tee-navy-m","brand":"Brewline","seller_name":"Brewline Coffee Gear","image_url":"https://example.com/images/tee-navy.jpg","price":"24.00 USD","availability":"in_stock"}
{"item_id":"TEE-NAVY-L","group_id":"TEE","listing_has_variations":true,"variant_dict":{"color":"Navy","size":"L"},"offer_id":"brewline-TEE-NAVY-L","title":"Cotton tee, navy, size L","description":"Midweight cotton tee.","url":"https://example.com/products/tee-navy-l","brand":"Brewline","seller_name":"Brewline Coffee Gear","image_url":"https://example.com/images/tee-navy.jpg","price":"24.00 USD","availability":"out_of_stock"}

Rules that catch teams out: use identical option names across a group, keep top-level color or size consistent with variant_dict because neither reconciles conflicts for you, and never put the price inside an offer_id. Legacy Custom_variant1_category and option pairs still work, but do not mix them with variant_dict.

For identifiers, gtin must be exactly 8, 12, 13 or 14 digits with a valid check digit, kept as a string so leading zeros survive. Submit mpn together with brand, and do not invent an MPN to fill a missing GTIN. This pre-upload checker covers the required fields, the availability values, the money format, GTIN check digits and duplicate IDs, and writes a gzipped JSONL snapshot:

import gzip
import json
import re

REQUIRED = ("item_id", "title", "description", "url", "brand",
            "seller_name", "image_url", "availability", "price")
AVAILABILITY = {"in_stock", "out_of_stock", "pre_order", "backorder", "unknown"}
PRICE = re.compile(r"^\d+(\.\d+)? [A-Z]{3}$")


def gtin_ok(value):
    if not value.isdigit() or len(value) not in (8, 12, 13, 14):
        return False
    body, check = value[:-1], int(value[-1])
    total = sum(int(d) * (3 if i % 2 == 0 else 1)
                for i, d in enumerate(reversed(body)))
    return (10 - total % 10) % 10 == check


def problems(row):
    out = [f"missing {f}" for f in REQUIRED if not str(row.get(f, "")).strip()]
    if row.get("availability") not in AVAILABILITY:
        out.append("availability would reject the row")
    if not PRICE.match(str(row.get("price", ""))):
        out.append("price must look like 79.99 USD")
    if row.get("gtin") and not gtin_ok(str(row["gtin"])):
        out.append("gtin fails the check digit")
    return out


def write_snapshot(rows, path="products.jsonl.gz"):
    seen, failed = set(), 0
    with gzip.open(path, "wt", encoding="utf-8") as f:
        for row in rows:
            issues = problems(row)
            if row.get("item_id") in seen:
                issues.append("duplicate item_id")
            seen.add(row.get("item_id"))
            if issues:
                failed += 1
                print(row.get("item_id"), issues)
                continue
            f.write(json.dumps(row, ensure_ascii=False) + "\n")
    return failed

Search and checkout eligibility flags

Setting a checkout flag does not complete checkout onboarding, since checkout requires a separately enabled integration. Legacy aliases such as enable_search and enable_checkout are still accepted, but send only one name per value: the enable_ version wins over the is_eligible_ version, which produces confusing results if an old export script is still writing both.

Also check the prohibited products policy before uploading. OpenAI excludes categories such as adult content, alcohol, nicotine, gambling, weapons and prescription-only medications, and it may remove products or ban a seller that violates the policy. Filter them out of the feed or set search eligibility to false.

Optional fields worth sending, and the US-only default

The market default deserves attention. The standard OpenAI-format upload currently targets the US, and row-level columns such as target_countries do not change that. Other markets need OpenAI to confirm the integration first, and a currency, product URL or sizing system does not select a destination. If you sell mainly outside the US, ask about market support during onboarding before you invest in localized feeds.

How often to update the feed and how removal works

Publish full snapshots at least daily. A product missing from a processed snapshot is not deleted immediately: OpenAI keeps its last processed record for up to 14 days, which protects you from a delayed shard. That has two consequences. To pull a product from search on the next processing run, set is_eligible_search=false. To remove it by omission, leave it out of every shard in later full snapshots and expect the retained record to expire within 14 days.

Date fields do not schedule anything. Sale windows, expiration dates and availability dates do not switch prices or stock automatically, so update the price when a sale starts and ends, and update availability when an item sells out. Because products are matched by item_id, generate IDs from your stable SKU or database key, never from a row number or a title slug that changes when someone edits the name.

Can you reuse a Google Merchant Center feed for ChatGPT?

Only when OpenAI confirms the Google-compatible format for your registered feed. It is a separate profile with different rules, and the OpenAI format is checked first, with one format applying to the whole upload.

The practical lesson is that a Google export is a starting point, not a drop-in. The spelling of preorder, the missing opt-out and the registered seller name are exactly the differences that surface late.

Common ChatGPT feed mistakes and a pre-upload checklist

• Price in the wrong shape. Write 79.99 USD, not 7999 or 79,99.
• Bad availability values. An unrecognized value rejects the row. Use unknown explicitly rather than leaving it blank.
• Placeholder strings. Do not write null, n/a or unknown into optional fields. Omit the field.
• Invalid GTINs. Check the digit and keep leading zeros by treating the value as a string.
• New file names each run. Overwrite one stable name or shard set instead.
• Two shipping representations. Do not send both shipping_price and a shipping tuple for the same charge, and note that the tuple needs feed-specific setup.
• Private or blocked URLs. Product and image URLs must be publicly accessible.

Getting the feed built and kept current

Getting a feed accepted once is the easy part. Keeping IDs stable, prices and stock fresh and variants grouped correctly across every update is what protects your listings over time. Netalith builds and maintains product feeds for Google Merchant Center, Meta, Pinterest, Microsoft and OpenAI (ChatGPT). If you want a second opinion on your catalog or a feed built for you, request a free quote and describe your platform and catalog size.