Product feed prep for conversational ad surfaces

Product-feed campaigns are one of the reported formats for ChatGPT Ads in India, described as always-on. That turns the file your store already exports into raw material for a surface nobody has run yet, and most brands have not opened it since the day they connected it.

Your feed already exists and you have not looked at it

A product feed is a machine-readable table of your catalogue: one row per sellable item, with columns for id, title, description, landing page link, image link, price, availability, brand, category and product identifier. It is a table, not a webpage, and it is read by a machine.

For most Indian D2C brands it already exists in two or three places. Shopify and WooCommerce generate one and push it into Google Merchant Center for Shopping. A near-identical file sits in the Meta commerce catalogue, driving dynamic product ads and Instagram Shopping. Marketplace uploads are separate files on separate templates, which is why the versions drift apart.

What is not known yet: OpenAI has not published a feed specification, a required field list or creative specs, and has not said whether an existing Merchant Center or Meta catalogue can be pointed at it. That becomes checkable when self-serve access opens on 4 September 2026. Everything below is worth doing anyway.

The attributes that carry meaning when a machine is matching

A shopper on a listing page fills in the gaps from the photograph, the reviews and the page around it. A machine matching a stated need has the fields and nothing else. A blank field is not neutral: it is a product that cannot be matched to a need it satisfies.

The fields that carry real meaning are the ones answering which one: size or volume, material, formulation or key ingredient, power rating, compatibility, age band, skin or hair type, fit. Then the structural ones that look boring and are not: a correct brand value, a precise category, and a real product identifier, a hygiene problem of its own covered in our note on GTIN, EAN and UPC hygiene.

The check is simple. Read one row aloud with nothing else in front of you and see whether you could recommend that product to someone with a specific problem. If you cannot, nothing else can.

Keyword-stuffed titles fail when the match is conversational

The marketplace habit is to cram head terms into a title, because there a retrieval engine matches a typed query token by token. Carry that habit into a feed and you get rows like Premium Organic Cold Pressed Coconut Oil for Hair Skin Body Combo Pack Natural.

That is a keyword pile, not a product. When the match starts from a need stated in a person’s own words rather than a typed token, a heap of modifiers describes nothing precisely. Cold pressed virgin coconut oil, 1 litre is a product, and every word in it is true.

Keep four things in the feed title: the brand, what the product is, the one attribute that separates this variant from its siblings, and the size or quantity. Everything else belongs in the description and attribute fields. That is a rule about the feed title field, not about marketplace listing copy, which is a different discipline covered in marketplace keyword research and in the catalog mistakes that kill conversion.

A browse feed and a stated-need feed are not the same file

A feed built for category browse assumes the shopper already arrived at a category and is narrowing down: titles can be short variant labels, descriptions can be brand voice, and the category tree carries the meaning. A feed built to answer a stated need has to survive with no category context at all, so the description has to say who the product is for and what problem it solves, in plain sentences.

The fix is a pattern, not a rewrite of every row. First sentence: the use case and the person. Second: the one differentiator. Third: the constraint, meaning size, quantity, or what it is explicitly not for. Write the pattern once per product family and start with the families carrying the revenue.

Images and prices are an eligibility problem, not a design one

The image in a feed is a URL, and URLs break quietly. Stale CDN paths, silent 404s, an image showing a discontinued pack, an image of a three-pack on a row that sells a single unit. This is not a photography brief, and shoot standards belong to the marketplace image guidelines. The only question here is whether the row points at a live image of exactly this item.

Price is worse, because a wrong price costs you twice. The feed price has to equal the price a click actually lands on, after every automatic discount your store applies. A mismatch is the quickest route to disapproval on the surfaces that publish rules, and to a click that bounces on arrival.

Sync discipline is most of the job

Availability is the field that decides whether you pay for clicks on things you cannot ship. The standard failure: the feed refreshes overnight, a hero SKU sells out by mid-morning, and you buy clicks on it all day. In a festive week that is not a rounding error.

Four things to fix. Set a refresh cadence that matches how fast your best sellers move. Pull out-of-stock rows rather than leaving them at zero. If you serve limited geographies, make sure the availability state reflects what a real pincode can receive. And name the person who owns the file, because an unowned feed rots.

Then audit four numbers weekly: feed rows against live SKU count, rows with a missing or broken image, rows where feed price does not equal site price, and rows marked in stock with no inventory.

This work outlives whatever happens to this platform

The reason to do this now is not that ChatGPT Ads will work, because nobody knows whether it will, and whether to test it at all is a separate decision. The reason is that the same file feeds Google Shopping, Meta dynamic product ads, Instagram Shopping, your own on-site search and a growing set of AI shopping surfaces that read structured product data rather than pages. A clean, honestly synced feed is growth work that cannot be wasted, because there is no version of the next few years where machines read less of your catalogue than today.