What a look inside ChatGPT Shopping says about your product feed

Most advice about getting into ChatGPT's shopping results is guesswork dressed up as method. So when a research team publishes what they actually observed the system doing, it's worth slowing down for.

That's what Peec AI did this month. Their team, led on this piece by Metehan Yesilyurt and presented at brightonSEO by Malte Landwehr, studied the debug data behind ChatGPT's shopping results in the US and in Europe and wrote up what they found. Masters of Search summarized it a few days later with a useful warning, which we'll repeat: these are one team's observations of a system OpenAI can change at any time. Nothing here is an OpenAI specification.

With that said, the findings are specific, and most of them point to fixes a CPG brand can make in a week.

Two markets, two systems

The first finding explains a lot of confusing results people have reported. ChatGPT Shopping isn't one system.

	United States	Europe
Where products come from	OpenAI's own product index, built from feeds	Google Shopping results, fetched through a search API
Seller requirement	Verified-seller flag required	Not observed
How candidates are ranked	Text matching, embeddings and GTIN matching	Round robin across the queries it ran
Time to retrieve results	About 0.6 seconds	About 4 seconds

The US path runs on an index Peec says OpenAI calls Labrador. In Europe, ChatGPT ran several searches against Google Shopping and stitched the results together. In one European example, 5 searches returned 113 products and 18 were shown.

For a brand selling in both places, that means two different jobs. In the US, your [product feed](term:Product feed) and how OpenAI reads it decide a lot. In Europe, it's your Google Merchant Center listing doing the work, which is another reason tip 16 still matters.

Where the US feeds come from

Peec counted 669 feeds feeding the US index at the time, 606 of them active. About 70 came from direct partnerships, where a brand sends its own feed to OpenAI. Around 500 were feeds OpenAI had built itself by scraping retailer websites through Bright Data, a data collection company. The rest came through Shopify, plus one from Salesforce and one from a Stripe integration. Peec also notes that when a brand signs a direct partnership, OpenAI switches off the feed it had built by scraping that brand's site.

Show the data

Feed origin	Approximate number of feeds
Built by OpenAI from scraped retailer sites	500
Direct merchant partnerships	70

Source: Peec AI, ChatGPT Shopping optimization in 2026. Peec gave both as approximate counts.

That one chart should change how a brand thinks about this. If you sell through a large retailer, there's a good chance your products are already in ChatGPT's index, assembled from that retailer's pages. The description, price, size and image a shopper sees may be the retailer's version, not yours. That's fine when the retailer's page is accurate, and a problem when it's two packaging changes out of date. It's the same consistency issue from tip 5, with a new reader.

From 41 candidates to 5 on screen

Peec's clearest example was a search for running shoes. The system found 41 candidate products. It turned 32 of them into detailed prompt blocks, with title, description, ratings, GTIN and more, and showed the shopper 5. For each candidate the debug data showed standard search scores: embeddings, a BM25 text-match score and rank fusion.

Show the data

Stage	Products
Candidates found	41
Turned into prompt blocks	32
Shown to the shopper	5

Source: Peec AI, ChatGPT Shopping optimization in 2026

Two details from that search stand out.

The verified-seller flag was decisive. One shop had 25 of the 32 candidates, more than three quarters of them, and placed none. Peec says it had no verified-seller flag, an attribute each candidate carries. Peec says it hasn't yet found what determines the flag, and we don't know either. But visibility in the candidate pool clearly wasn't enough on its own.

More listings meant more slots. Dick's Sporting Goods had five candidate products, and four made the final five. Nike had two, and one made it. De-duplication only removed the same shoe in different sizes, so each model and color could take its own slot. For a CPG brand, that's an argument for listing each flavor and pack size as its own product.

Peec also saw personalization built in: the system knew the user's home and current location, an estimated gender with a confidence score, and which merchants the user had already seen or clicked. And the debug data listed an A/B test called "No Amazon," US shopping with zero Amazon results. Peec calls it only a test. We'd treat it as a reminder that the merchant mix can change quickly, and that a brand sold only on Amazon has one point of failure here.

What a CPG brand should take from a running shoe test

Running shoes aren't groceries, and it's fair to ask how much of this carries over. We think most of it does, for one reason: the system that picked those shoes is the same one that will pick your oat milk. The scoring is generic. Text matching rewards titles and descriptions that use the words shoppers use. Embeddings reward descriptions that are clearly about the thing being asked for. GTIN matching rewards products whose barcodes are present and consistent.

Where CPG differs is in the number of near-identical products. A search for "high protein yogurt" or "fragrance-free laundry detergent" returns dozens of products that look the same on paper. In that crowd, the details that separate you (the protein number, the certification, the pack size) need to be stated plainly and early, or the system has nothing to tell you apart with.

The first 40 words

This is the finding we expect to travel furthest, so it needs the most care.

Peec observed product descriptions stopping at 50 tokens at ChatGPT's scoring stage. In English that's roughly 40 words, or about 250 characters. Anything after that, in their observation, was ignored when scoring how well a product matched the search. Masters of Search was right to flag that this is one team's measurement, not a documented spec, and it may change.

Even so, it costs almost nothing to act on, and it's good writing practice anyway. Most CPG product descriptions open with brand voice: a line about the founder's kitchen, the morning ritual, the feeling. The facts a shopper compares on (what it is, size, the claims that matter, who it's for) come later. If only the first 40 words count, that order is backwards.

Here's a made-up example to show the difference. These are illustrative figures, not real data.

Version	First 40 words, roughly
Brand-voice first	"It started in a tiny kitchen in Vermont, where our founder wanted a breakfast that felt like a hug. Years later, we still toast every batch slowly, by hand, because some things are worth waiting for. Every..."
Facts first	"Maple almond granola, 12 oz bag. 8g protein and 6g added sugar per serving. Gluten-free certified, made without peanuts. Slow-toasted in small batches in Vermont. Good for breakfast bowls, yogurt and snacking..."

The second version still has room for the Vermont story. It just comes after the facts. Tip 3 covers how to describe a product so an assistant can match it to a question; this finding makes the order of those words matter too.

What to do this quarter

Peec's recommendations were short: send a feed or sell on Shopify, put key information in the first 40 words, list variants as separate products, test your titles and descriptions against simple text-matching, and in Europe keep Google Shopping strong. Here's how we'd turn that into a list for a CPG brand.

1. Find out if you're already in. Search ChatGPT for your products by name and by need ("gluten-free granola under $8"). If they show, check whose description and price appear. If they don't show, read tip 32 on getting in.
2. Apply for a feed if you sell DTC. OpenAI's merchant page says merchants can apply to share product feeds, Shopify and Etsy catalogs are already integrated, and there are no fees on purchases that start in ChatGPT. Shopping is live in the US. OpenAI says it is building a self-service portal for later this year.
3. Rewrite the first 40 words of your top 20 product descriptions. Facts first: product, size, key claims, use. Do it on your own site and in any feed you control.
4. List variants separately. A 12 oz and a 24 oz bag, or three flavors, should be distinct products with distinct titles. The model can only recommend what it can tell apart.
5. Include GTINs. Peec saw GTIN matching used in the US ranking. Your UPCs are an easy, exact signal that ties the scraped retailer copies of your product back to you.
6. Check your retailer listings. If OpenAI built a feed from a retailer's site, that listing is your product in ChatGPT. Fix the stale ones first.
7. Tag and watch the traffic. Peec saw clicks from US feed results carrying utm_medium=feed. Check your analytics for it; it's the cleanest way to see what this channel actually sends.

How much to trust this

We think the structural findings (two systems, scraped feeds, the verified flag mattering) are the most durable, because they describe how the thing is built. The specific numbers, the 50-token cut and the 41-to-5 funnel, describe a moment. Build habits from the first and run your own checks on the second. Tools such as XENA Intelligence's SearchPanel now record what ChatGPT shows for each product over time, and a manual monthly search works for a short list. Either way, the point is to watch what the system does, not what anyone, including us, says it does.