If you sell products, sooner or later you will ask whether you can take your catalogue to ChatGPT the way you take it to Shopping. The short answer is yes: the platform accepts a product feed. The long answer is that it resembles Shopping less than you expect, and that much of what is said about its requirements is not published anywhere.
Is it for you?
A feed makes sense when the specific product is the answer: shops with dozens or thousands of SKUs, with prices and stock that move, and with enough margin to absorb discovery traffic. If you sell a service, a single product or software, you do not need a catalogue: ordinary ads and good context hints are enough, and you will keep far more control over what each ad promises.
The good news: your Google feed is the starting point
OpenAI has not invented a format. The official product file specification accepts a profile compatible with Google Shopping feeds, with the caveat that only the profile described in its documentation is supported. In practice: the file you already maintain for Merchant Center is the draft for this one, not a job from scratch.
What the official documentation does spell out (read on 18/09/2026):
- Format: UTF-8 CSV, one row per purchasable item or variant.
- Required base fields: item id, title, description, product URL, brand, image, price, availability and seller name.
- Price with its currency and current; page and image reachable over public HTTPS.
- An ads-eligibility column that must be set to "true" for the product
to be processed:
is_ads_eligible. Beware of the near-identicalis_ads_enabled, which the platform does not read (a warning from the Headcore manual, August 2026). - Final eligibility also depends on product data, availability, reviews, filters and the campaign.
⚠️ Trap: OpenAI's crawlers must be able to see the product page and the image. A flawless catalogue behind a bot-blocking wall, or with images served over HTTP, is a catalogue that does not get advertised.
What we could not verify
It is repeated everywhere that a feed takes "between 100 and 1,000,000 items". In the official documentation we could read on 18/09/2026 there is no minimum or maximum number of products, and third-party sources contradict each other: some say 100 items minimum, others 1,000 minimum and 2 million maximum. The Headcore manual (August 2026), which reviewed the documentation, says in so many words that the thousand-product minimum should be treated as unverified.
In practice: do not build your plan on that number. If your catalogue is small, prepare it properly and check in your own account whether the feed is accepted; if it is huge, assume there is a ceiling and prioritise what you upload.
How it compares with Shopping
| Google Shopping | ChatGPT Ads feed | |
|---|---|---|
| Source file | Merchant Center | The same feed, compatible profile |
| What triggers the product | The search and your data | The conversation |
| Search terms | Full report | Does not exist |
| Negatives | Campaign, group and list level | None at all |
| Value bidding (tROAS) | Yes | Does not exist |
| Levers for control | Priorities, labels, negatives | Campaigns, hints and which products you upload |
That last row is the key one: with no negatives, your real control is the catalogue and the context hints. Whatever you do not want advertised, do not upload it, or mark it as not eligible. In Shopping you correct afterwards; here you decide beforehand.
Prepare the catalogue for a conversation, not a search
Nobody tells ChatGPT "trail shoe size 8 black sale". They describe a problem. So titles and descriptions have to talk about use, not just the reference:
| Before | After |
|---|---|
| "Model XR-200 size 8 black" | "Lightweight trail running shoe for rocky ground, size 8" |
| "Ergonomic chair mod. 7B" | "Ergonomic office chair with lumbar support for long working days" |
Three rules follow:
- The title says what it is for, with the brand and the attribute people mention out loud (size, capacity, material, compatible with…). The model number goes at the end, not the start.
- The description answers questions: who it is for, what is included, what is not, lead times and returns. It is text a conversation can use.
- One product per row, variants included, with real price and real availability.
Typical mistakes that cost delivery
- Price or availability out of step with the product page: the classic rejection on any catalogue platform.
- Unreachable images or pages (HTTP, bot blocking, intermittent 404s).
- The eligibility column missing, misspelled or under the wrong name.
- Country and currency that do not match the campaign.
- A feed uploaded once and never refreshed: a live catalogue goes stale within days.
What to measure
With a catalogue, the number that rules is not the click: it is cost per sale and margin by product family. Since the platform gives you no search terms, build your diagnosis from what you do have: which products get impressions, which absorb the spend and which convert, cross-checked against your own analytics. Conversion value is yours to supply: the platform's reports do not carry it (see "Why there are no official performance benchmarks, and how to read the ones going around").
What to take away
- The feed reuses your Google Shopping one: UTF-8 CSV and a compatible profile, per the official documentation.
- What is published are the base fields, current prices and public URLs and images; minimum and maximum item counts are not: do not plan on third-party figures.
- No negatives and no search terms: control lives in what you upload and in your hints.
- Write titles and descriptions about use, not about references.
- Measure cost per sale and margin, and supply the conversion value yourself.