ChatGPT Product Feeds Now Gate Shopping Results (September 4, 2026): Should Michigan Retailers Fund a Feed or More Google Shopping?

On September 4, 2026, OpenAI emailed advertisers a batch of ChatGPT Ads changes and Search Engine Roundtable published the list the same morning. One item buried in the coverage matters more than the rest: analyst Juozas Kaziukėnas reported on LinkedIn that ChatGPT shopping results now appear to come only from brands submitting a product feed. OpenAI has not confirmed that gate publicly, so treat it as a well-sourced observation rather than policy — but plan as if it holds, because the cost of being wrong runs one direction.

What actually changed in ChatGPT Ads on September 4, 2026?

Several things shipped at once, and they are all plumbing for performance advertising rather than new ad surfaces.

• Ads Manager plugin — build, edit, troubleshoot and report on campaigns from inside ChatGPT or Codex in plain language, with a confirmation step before changes apply.
• Custom audiences — edit members without rebuilding the audience, mix identifier types in one request, use audiences under 25,000 matched users for exclusions, build past 5 million members, and pass Google Advertising ID (GAID).
• Conversion matching — the Measurement Pixel now accepts hashed phone, first name, last name, region and postal code; the Conversions API accepts plural list fields plus Android GAID.
• Carousel reporting — feed campaigns in the multi-product carousel now report card impressions and clicks, complete from August 20, 2026 on a rolling 30-day basis.
• Total budget pacing and geographic expansion — total-budget campaigns now spread spend instead of front-loading, and ChatGPT Ads began rolling out to select countries in Europe, India, the Middle East and North Africa that week.

Why does a product feed now decide whether you appear organically?

Because ChatGPT uses one structured file for two jobs. The catalog that generates your paid ad units also supplies the product data ChatGPT draws on when it answers a shopping question without an ad. No feed, no authoritative price, availability or attribute data — so the model has nothing solid to name you with.

Internalize that mechanism, because it predicts cases nobody has written up yet. Google keeps organic and paid shopping on separate rails — you can rank in free listings without ever funding a Shopping campaign. ChatGPT has collapsed both onto one input. Every feed decision you make to control ad spend, including which SKUs you filter out of ad eligibility, simultaneously edits your visibility on a surface you are not paying for. Suppress a slow-moving product line and you may also remove it from ChatGPT’s answer when a customer asks for it by name.

The same release hides a second trap. Carousel card impressions count when an individual product card becomes viewable, and OpenAI states plainly that they are separate from billable ad impressions and excluded from billable counts. You now have two impression denominators in one account. Divide clicks by card impressions and the CTR looks catastrophic next to Google Shopping; divide by billable impressions and it is comparable. Export both into one dashboard without labeling the denominator and you will produce a chart arguing to shut off a campaign that works.

Does ChatGPT ad traffic convert well enough to justify the build?

The published numbers are encouraging and should be discounted heavily. Criteo, OpenAI’s first ad tech partner, has reported AI-referred conversion rates approaching twice those of traditional search in consumer electronics, lifestyle and wellbeing, and home and garden, per PPC Land.

Read that claim carefully before moving budget on it. A 2x conversion rate on a channel this young describes which queries currently qualify for ad inventory, not the channel’s ceiling. ChatGPT ads serve against long conversational questions, and someone describing a problem in a paragraph is deep in consideration by definition. Compare that against “traditional search,” a bucket holding every broad-match impression a Shopping campaign ever bought, and the ratio is close to guaranteed. The honest comparison is ChatGPT ads against your own long-tail non-brand traffic, and nobody has published that.

The platform’s own commerce reads have been wrong before, too: OpenAI shut down instant checkout in March 2026 after Walmart disclosed that in-chat conversion rates ran three times lower than click-outs to Walmart’s site. Treat vendor performance framing as a hypothesis you test with small amounts of your own money.

How should a Michigan retailer choose between a ChatGPT feed and more Google Shopping budget?

Use catalog complexity as the tiebreaker, not budget size. If your products are fitment-critical — the buyer must land on the exact right SKU or the sale fails — build the feed now. If your products are interchangeable, add the budget to Google Shopping instead and revisit in Q1.

That rule follows from how conversational shopping works. A customer asking ChatGPT for a wheel bearing for a 2019 Silverado 2500HD is handing the model a fitment problem to solve. If a Warren or Sterling Heights aftermarket supplier has no feed, the model answers from whatever it can reconstruct off web pages — and reconstructed fitment is how a customer ends up with a part that does not fit, a return, and a review that says so. The same logic covers powersports dealers in Grand Rapids, marine parts sellers around Muskegon and Holland, and industrial MRO distributors in Livonia selling by spec table.

The stakes are concrete. A Michigan parts seller doing $40,000 a month online at a 12% return rate already absorbs roughly $4,800 in returned goods monthly. Fitment errors introduced by an AI surface you never configured land on top of that, and they arrive attributed to nothing in your analytics. We covered the underlying discipline in our fitment audit for Michigan parts sellers, and the same title and attribute hygiene carries straight over.

What does the Ads Manager plugin change about who does the work?

It moves campaign construction from a specialist task to a review task. The plugin turns a brief or a website into launch-ready ads, creates variants and surfaces prioritized recommendations, with a confirmation step before anything applies.

For a three-person team, that confirmation step is the whole product. The reviewer needs enough context to say no, not enough skill to build the alternative themselves. Name one reviewer and write down what they check — budget caps, geo boundaries, excluded SKU groups — before you connect the account. A plugin that proposes and a human who approves everything is just an unsupervised plugin.

How do you stand up a ChatGPT product feed in one week?

Reuse the Google Shopping file you already maintain. Per Digiday’s reporting, the format is close enough that most retailers can repurpose it, which is what makes this a one-week project rather than a quarter-long one.

1. Export your current Google Merchant Center feed and pull a 100-product sample — OpenAI has been requiring new ecommerce partners to submit that sample before connecting a full catalog.
2. Pick your 100 by margin, not by volume. The sample is what the platform learns your catalog structure from, so send products with complete attributes rather than your bestsellers with thin data.
3. Fix titles before you upload, not after. Front-load year, make, model and the distinguishing spec. Conversational queries carry the qualifier; your title has to contain it.
4. Set storage residency and currency deliberately in the Feeds section, and decide consciously whether to toggle organic product listing participation on. That toggle governs the unpaid surface.
5. Wire the Measurement Pixel or Conversions API before you spend. The new matching fields only help if they are populated at implementation.
6. Load a suppression audience under 25,000 users — existing customers, warranty claimants, wholesale accounts. The exclusion floor was just removed, so a small clean list finally works.
7. Start at the $3 to $5 recommended CPC with a total budget and let the new pacing spread it. Minimum spend went from $250,000 at the February 2026 pilot launch to $50,000 to zero on May 5, 2026.
8. Label your reporting denominators so carousel card impressions never sit in the same column as billable impressions.

What dates belong on your Q4 calendar?

Three of them, and they interact. Build the feed now if fitment matters; otherwise watch the carousel data through Q4 and decide in January on your own numbers, not Criteo’s.

• August 20, 2026 — the earliest date with complete carousel card data, on a rolling 30-day window. Pull your baseline before that window rolls past.
• Now through early October — the sample-and-approval cycle for a new feed. Start in November and your first ChatGPT impressions land after the holiday peak has priced itself.
• Your Q4 feed freeze — whenever your team stops structural catalog changes. Google’s Content API for Shopping deprecation carries the same deadline pressure, mapped in our Content API 410 error checklist. Do not run both feed projects in the same week.

Frequently asked questions

Is the product feed requirement for organic ChatGPT shopping results official?

No. It comes from analyst Juozas Kaziukėnas on LinkedIn and was described by Search Engine Roundtable as a suspicion, not confirmed policy. OpenAI’s September 4 email covered the Ads Manager changes without addressing organic eligibility. Treat it as credible and unconfirmed — submitting a feed costs little either way.

Can I use my existing Google Merchant Center feed file?

Largely, yes. Digiday reported that retailers can mostly repurpose the structured file they already send to Google, and one brand completed the process through Criteo faster than expected. Expect to adjust currency and storage residency, plus the 100-product sample step before your full catalog connects.

How large can a ChatGPT product catalog be?

Up to one million SKUs per advertiser, according to an ad executive cited by Digiday. That ceiling is well above what most Michigan distributors carry, so catalog size is not the constraint. Attribute completeness is. A 40,000-SKU catalog with half-filled fitment fields will underperform a 4,000-SKU catalog with complete ones.

Should a service business without products do anything here?

Not with feeds. But the audience changes apply to everyone: removing the 25,000-user exclusion floor lets a multi-location HVAC or dental group finally suppress existing patients and customers from prospecting, which was impossible at their list size. That one change is worth an afternoon of implementation work.