ChatGPT Shopping no longer decides what to recommend by reading your product page. On July 10, 2026, the share of product recommendations served from integrated product feeds jumped from 8.26% to 61.54% in a single day, and it has since settled at roughly 65%. If your catalog is not in a structured feed that OpenAI ingests, you are competing for a shrinking third of the shelf.
That is the headline. The more useful finding sits underneath it, and it cuts both ways: feed-sourced products land in the first offer slot almost every time, but even merchants who already have feed integration still see three quarters of their offers pulled from ordinary web product pages. Feeds are not a replacement for a good PDP. They are a second, higher-priority lane that most merchants cannot yet enter.
Key takeaways
- Feed-integrated retrieval went from 8.26% to 61.54% of tracked product recommendations in one day (July 10, 2026), measured across 1,757,723 prompt runs, and sat near 65% by early September.
- Of 687 sampled brands, 450 lost at least a third of their ChatGPT Shopping visibility across that break. 67 gained by the same margin.
- The shelf consolidated hard: the top 10 merchants went from 22.5% to 41.8% of all references, and unique merchants referenced fell from 13,524 to 10,607 — a drop of more than 20%.
- When a product is cited from a direct feed, roughly 99.9% of the time it appears as the first product offer.
- Feeds are gated. OpenAI's own documentation says feed onboarding "is currently available to approved partners," which means most merchants' only lever right now is still the product detail page.
What actually changed inside ChatGPT Shopping?
Until mid-2026, ChatGPT assembled product recommendations much the way a search engine would: it crawled and read product detail pages. In a sample of 201,137 shopping prompt runs from June 18–25, 2026, covering 812,190 product cards, 87.3% of cards were retrieved from web crawl and only 12.7% came from direct merchant product feeds.
Three weeks later that ratio inverted. Analysis by Profound of 1,757,723 daily prompt runs found feed-integrated sources climbing from 8.26% to 61.54% on July 10 — a roughly 6.5x move inside twenty-four hours — and holding around 65% of tracked product recommendations by September 3.
A one-day step change of that size is not gradual model drift. Profound attributes it to the ChatGPT 5.6 model release it dates to July 9, while being careful about the claim: it describes that release as "the only event-based factor on this exact date that has such explanatory power." That is correlation with a very tight timestamp, not a confirmed cause, and it is worth holding it that way.
The practical read does not depend on the attribution being right. Whatever triggered it, the retrieval path changed, and the change stuck.
How much did merchant visibility actually move?
Enough that most affected brands would have registered it as an unexplained drop with no obvious cause.
Comparing the three days before the break (July 7–9) with the three days after (July 10–12) across 687 sampled brands:
| Outcome | Brands | Threshold |
|---|---|---|
| Visibility fell | 450 | at least one-third |
| Visibility rose | 67 | at least one-third |
| Roughly unchanged | 170 | below one-third either way |
Across the 517 brands that moved materially, a joint regression model using only the change in web-search retrieval and the change in feed retrieval explained 83% of the observed variation in visibility. In plain terms: for five out of six brands that swung, the swing is accounted for by which retrieval lane their products were in. Not content quality. Not brand strength. Not price.
The concentration numbers tell the same story from the other end. Top-10 merchant share of all references nearly doubled, from 22.5% to 41.8%, while the number of distinct merchants referenced fell from 13,524 to 10,607. About 35% of feed-integrated retrieval traced to Shopify storefronts, which is the clearest available signal of who was positioned to benefit: merchants sitting on a platform that could deliver a compliant catalog at scale, rather than merchants who happened to write better product copy.
Does a product feed guarantee a better position?
Close to it — for the products that make it in. Across roughly 1,000,000 shopping offers tracked over 30 days, when a product citation came from a direct feed, about 99.9% of the time it occupied the first product offer slot.
That number is the strongest argument for feed access anyone has published. It is also the one most likely to be over-read, because of the counterweight sitting right beside it:
Even among merchants that already have feed integration, 75.81% of their offers still come from web product detail pages.
And in that same 30-day snapshot, 88.29% of all product offer instances still originated from web PDPs.
Both things are true at once. Feed coverage is partial — it applies to the items you submit, in the categories and regions where it is enabled, when the data validates. Everything outside that boundary falls back to the crawler. A merchant who treats feed onboarding as permission to stop maintaining product pages will win the first slot on their feed-covered SKUs and quietly lose everything else.
Why do the feed percentages seem to contradict each other?
They do, and this is where most retellings of this data go wrong, so it is worth being explicit.
You will see feed share reported as 12.7%, as 20%, as 61.54%, and as 65% — all from the same research operation, all correct. They measure different things:
| Figure | What it counts | Window |
|---|---|---|
| 12.7% | Product cards retrieved from feeds | June 18–25, 2026 |
| ~20% | Share of shopping retrievals from feeds, end of an 8-month trend | through mid-2026 |
| 61.54% | Share of tracked product recommendations | July 10, 2026 |
| ~65% | Same measure, later reading | September 3, 2026 |
Cards, retrievals, offers, and recommendations are not interchangeable units, and the sampling windows overlap unevenly. If you are reporting this internally, cite the unit alongside the number — otherwise someone will build a forecast on a comparison that was never valid.
What does OpenAI actually require in a product feed?
This part is documented rather than inferred. OpenAI's Agentic Commerce Protocol feed specification lists nine fields required on every record:
| Field | Notes |
|---|---|
item_id |
Stable, unique per item or variant within your feed |
title |
Max 150 characters |
description |
Factual overview, max 5,000 characters |
url |
Product detail page with the variant already selected |
brand |
As shown on the product page |
seller_name |
The entity supplying the offer |
image_url |
Main image showing that specific variant |
availability |
One of in_stock, out_of_stock, pre_order, backorder, unknown |
price |
Money format, e.g. 79.99 USD |
Beyond the required nine, the spec defines conditional and recommended fields covering variants (group_id, variant_dict, offer_id), attributes (condition, product_category, color, size, material, gender, age_group), commerce terms (sale_price, shipping_price, accepts_returns, return_deadline_in_days), and social proof (review_count, star_rating).
Accepted formats are JSONL, CSV/TSV in UTF-8, and Google-compatible .txt/.tsv/.csv, with .gz compression supported. If you already maintain a Google Merchant Center feed, you are most of the way there.
The gate is access, not format. OpenAI's getting-started guide states that "onboarding product feeds in ChatGPT is currently available to approved partners," with applications submitted through an OpenAI form. There is no self-serve path as of this writing.
That gating is the part of this story worth sitting with. A retrieval system that strongly prefers structured feeds, combined with feed access granted by application, produces exactly the consolidation the data shows: 2,917 fewer merchants on the shelf and the top ten taking nearly twice the share. This is not a meritocratic reshuffle. It is a distribution bottleneck, and small merchants are on the wrong side of it by default.
What should you do if you can't get feed access yet?
Work the lane that still carries most of the volume. Web PDPs remain the source of 88.29% of product offer instances overall, and 75.81% of offers even for feed-integrated merchants — so PDP quality is not a consolation prize.
The attributes that separate top-ranked from bottom-ranked product cards are measurable:
| Signal | Observed difference |
|---|---|
| Median review count | 787 on top-ranked cards vs 352 on bottom-ranked — a 123.58% lift |
| Explicit availability info | Present on 79% of top-ranked PDPs |
| Promotional / sale pricing | 13.33% lift on top-ranked cards |
| Best-price positioning | 21% prevalence gap between top and bottom offers |
Concretely, that means: publish live stock status in machine-readable form rather than a JavaScript-rendered badge; surface review counts and ratings as text on the page; keep sale pricing and its end date explicit; and make sure the variant a shopper would actually want has its own crawlable URL with its own image — which is, not coincidentally, exactly what the feed spec demands. Build your PDP to satisfy the feed schema and you are both ranking better today and pre-validated for the day access opens.
Two things are worth checking before anything else: that OpenAI's crawler can reach your product pages at all, and that your titles and descriptions survive truncation at the limits the spec sets. You can test both in a couple of minutes with the AI crawler checker and the meta tags checker.
Frequently asked questions
Did ChatGPT stop using web search for shopping?
No. Feed-integrated sources account for roughly 65% of tracked product recommendations, which leaves about a third still served from web retrieval — and web product detail pages remain the origin of 88.29% of product offer instances in the broader offer data. Web search was demoted, not removed.
Is the July 10 shift confirmed to be caused by GPT-5.6?
Not confirmed. Profound's analysis dates a ChatGPT 5.6 release to July 9, 2026 and calls it "the only event-based factor on this exact date that has such explanatory power," but stops short of asserting causation. Treat the timing as strong circumstantial evidence and the mechanism as unverified.
How do I get my product feed into ChatGPT?
OpenAI's documentation states that feed onboarding "is currently available to approved partners," with access granted through an application form rather than self-serve signup. Prepare a compliant feed in advance — nine required fields, JSONL or CSV/TSV, UTF-8 — so you are ready when access widens.
Does having a product feed guarantee the top slot?
For products actually covered by the feed, close to it: roughly 99.9% of feed-sourced citations appear as the first product offer. But feed coverage is partial, and 75.81% of offers from feed-integrated merchants still come from web pages. The feed wins the slots it covers; the PDP has to win the rest.
Why did the number of merchants referenced fall?
Unique merchants referenced dropped from 13,524 to 10,607 across the July 10 break, while the top 10 merchants went from 22.5% to 41.8% of references. When retrieval favors a gated data source, the merchants inside the gate absorb the share that merchants outside it lose.
Can I reuse my Google Merchant Center feed?
Largely, yes. The specification accepts Google-compatible .txt, .tsv, and .csv files alongside JSONL, and the required fields map closely onto Google's. The main additions to check are seller_name, the constrained availability enum, and variant-specific url and image_url values.
Where this leaves you
The measurable lesson of July 10 is that in AI shopping, how your data reaches the model now outweighs how well your page is written. That is uncomfortable, because feed access is currently allocated by application rather than earned by performance — and the consolidation numbers show precisely who that arrangement rewards.
The response is not to wait for the gate to open. It is to make your product data feed-shaped now, in public, where the crawler can still see it, and to track your visibility by retrieval lane so that the next step change arrives as a diagnosis rather than a mystery. If you are building that measurement, start with the KPIs that actually track AI search visibility, and see what happened to Shopping ad CTR and impressions under AI Overviews for the paid-side version of the same squeeze.
Run the free AI crawler checker against your product pages first. If OpenAI's crawler cannot fetch them, nothing else on this page matters yet.