ChatGPT now runs its own product index and is testing it against scraped search results on live shopping traffic, which means Bing and Google rankings no longer predict whether a Shopify store gets recommended. Catalog data quality, review depth, and crawlable product pages do.
Only 1.5% of the URLs in ChatGPT’s own index appear in Bing’s top 20 for the same query. Two years of “rank on Bing to show up in ChatGPT” advice was built on a pipe that the free tier barely touches anymore.
On September 4, a researcher at Peec AI published something the industry had been guessing at for two years: the receipts. Between May 21 and July 21, 2026, ChatGPT’s own server stream carried a field called result_source that named the pipeline behind every search result it returned. Four values came back. Three were external scraping providers. The fourth was Labrador, OpenAI’s own index. And Labrador turned out to be a family of vertical indexes, one of which is shopping.
I spent six years at Shopify watching merchants pour money into whichever discovery channel was fashionable that quarter, and the pattern that repeated most often was the same one I see in the AI visibility space right now: someone publishes a finding, an entire cohort of agencies turns it into a rule, and the rule outlives the finding by about 18 months. “Rank on Bing to show up in ChatGPT” was a reasonable rule in 2024. It is not a reasonable rule in September 2026, and the evidence for that is now public.
This piece is about what that evidence says, what it does not say, and what a Shopify merchant at three different revenue stages should do with it this month.
ChatGPT operates its own retrieval index, internally called Labrador, and as of early September 2026 it is running at least five separate experiments that test a product index against scraped search results on real shopping traffic. That is the core finding in Tomek Rudzki’s investigation for Peec AI, and it rests on three kinds of evidence.
The first is the server events themselves. Labrador is not one index. It breaks out into general web, PDF, YouTube, news (with separate tiers for the last day, last week, and older), arXiv, Wikipedia, local, finance, legal, medical, shopping, and images. That is the same shape Google spent two decades building: a general index with vertical indexes layered on top, each tuned to a content type. The shopping vertical is the one that matters for anyone reading this.
The second is the experiment log. In mid-August, an A/B test named prefer-index-over-serp-v3 was affecting 8% of chats. By September 2, five shopping experiments were live at once: prefer-index-over-serp-v3, chatgpt-shopping-noamazon, shopping-index-q2qb, shopping-hqi-v2, and shopping-hq-v1. Two of them are confirmed to run on ChatGPT’s own index rather than scraped results. One is testing what shopping looks like with Amazon removed. The ranking stack visible in a related experiment includes a BM25 lexical pass that narrows to 10 sources, a 400 candidate product pool, a 400 candidate reranker, and vector search at two dimensions. That is a search engine, not a summarizer bolted onto someone else’s results.
The third is the paper trail. Nick Turley, who runs ChatGPT, testified under oath in the Google antitrust case that OpenAI had “significant quality issues” with non-Google search data, that Google refused to license its index, and that OpenAI began building its own index in 2023 with an original target of answering 80% of queries from it by the end of that year. OpenAI job postings describe indexing systems at exabyte scale. And a Peec researcher published a one billion page test site that ChatGPT’s crawler had worked through six million pages of by early September, at 35,000 requests per hour. Nobody crawls at that rate to throw the pages away.
What Peec does not claim matters too. OpenAI has not confirmed any of this publicly. The experiments are slices, not defaults. And Rudzki closes by saying the system is being rebuilt week to week and some of it may already look different. Treat this as a direction of travel with strong evidence behind it, not a finished map.
Bing rankings do not predict ChatGPT visibility because only 1.5% of the URLs served from ChatGPT’s own index appear in Bing’s top 20 for the same queries, and that index is what the free tier runs on. Resoneo’s analysis of 88,000 ChatGPT search results, published in Search Engine Land on August 17, measured that overlap three ways and found the same answer each time. No Labrador snippet matched a Bing snippet. Bing caps titles at 75 characters; 24% of Labrador titles run past that.
The routing is economic, and once you see the economics the strategy writes itself. In instant mode, which is what a free user gets, ChatGPT has to answer in a few seconds for someone paying nothing. So it queries only what OpenAI already owns. Resoneo found that zero pages were opened in 93% of instant answers. The model grounds itself on a title and a snippet of roughly 200 characters cut from your page at crawl time. The expensive tools, scraped Google results and live page fetches, are reserved for paid Thinking mode where the user pays and waits. More than 90% of ChatGPT users are on the free tier. Free means instant, and instant means Labrador.
For a Shopify merchant the implication is sharper than it is for a publisher. Resoneo’s write-up says plainly that shopping and local never touch web search at all. Product cards are served by internal pipes labeled P1, P2, and P3, drawing on OpenAI’s own merchant feeds, with Google kept around as an oracle for prices and reviews. Your battle for a product card is fought in feeds, catalog data, and business listings, on top of whatever classic citations you earn in the text narrative.
Here is the thing I want to be honest about, because I gave the Bing advice myself in early 2025. It was not wrong then. Microsoft still supplies grounding to ChatGPT through its Web IQ platform, Bing still shows up as a source in Deep Research, and Peec confirmed with a zero-traffic canary site that ChatGPT still queries Google while grounding answers. The pipes still exist. What changed is that they are no longer the pipe the majority of your shoppers flow through, and OpenAI is actively testing how much less it can rely on them. A tactic that works on a shrinking slice of traffic is a tactic on a clock.
ChatGPT selects product cards by fanning your shopper’s prompt out into internal product sub-queries, retrieving roughly four candidate cards per prompt, and then filling each card’s price and merchant from the top ranked offer it can find. Profound’s breakdown of 201,137 shopping prompt runs in late June 2026, covering 812,190 product cards, is the clearest published picture of that pipeline, and its numbers are worth sitting with.
Of those 812,190 cards, 87.3% were retrieved from web crawl and 12.7% came from direct merchant product feeds. When a card came from web crawl, the commercial information on it was pulled from three sources: ChatGPT’s own fallback product search, Google index signals, and live merchant pages. Live merchant sources accounted for just under 25% of those retrievals. The remaining three quarters came from ChatGPT’s internal index and Google’s commercial signals, which is the shopping equivalent of what Peec found on the web side. Bing is not in that list.
The ranking correlations are the actionable part. Profound split the sample into 200,975 rank 1 cards and 214,301 cards at rank 4 or lower. The two biggest gaps were GPT tag presence (2.39% of top cards carried a machine-assigned label like “best value” or “premium” versus 0.98% of bottom cards) and median review count (787 reviews for top ranked cards versus 352 for bottom ranked). Promotional pricing showed a modest 13% lift. URL length and product name length showed no lift at all. If you have ever debated whether to shorten a product title to help AI, the data says the debate is not worth your afternoon.
One more finding that should reset how you think about this channel: Reddit accounted for roughly one third of all shopping citations. The text narrative that wraps your product card is being influenced by forum threads, and those threads also shape which candidates get selected. I will come back to what that means at each stage, but the short version is that a $200K per month brand with 40 detailed reviews per SKU and two honest Reddit threads about it has a structural advantage over a $2M brand with beautiful photography and 12 reviews.
Shopify Catalog gets your products into ChatGPT’s merchant feed layer automatically if you sell to US buyers, but it does not decide whether your product wins the card, and it does not fix product data you have not written. OpenAI’s own help documentation on ChatGPT shopping states that Shopify product data is already integrated through Shopify Catalog and no additional work is required from individual merchants. That sentence is true and it is also the most misread sentence in this whole channel.
Here is what it actually buys you. When OpenAI expanded the Agentic Commerce Protocol to product discovery on March 24, 2026, it named Shopify Catalog alongside direct ACP integrations from Target, Sephora, Nordstrom, Lowe’s, Best Buy, The Home Depot, and Wayfair. Shopify’s agentic commerce documentation confirms the ChatGPT channel is live for any Shopify Catalog merchant selling to US buyers regardless of store location, with purchases completing on your storefront through an in-app browser. The same post reports that AI-driven traffic to Shopify stores grew 8x year over year in Q1 2026 and orders from AI-powered searches grew nearly 13x, with new buyers ordering through AI channels at close to twice the rate of other channels. I covered the March switch-on in the week agentic commerce stopped being theoretical, and the architecture has held: discovery in the chat, conversion on your turf.
Where it stops is the part that determines your outcome. First, Catalog is the feed path, and feeds were 12.7% of product cards in Profound’s sample. The other 87.3% came from crawl, which means your product pages themselves, and how they read to Labrador, are still the majority case. Second, when a shopper clicks a card and sees the merchant list, OpenAI ranks merchants on availability, price, quality, and whether you are the maker or primary seller. Catalog delivers those fields; it does not make them accurate. A product marked available that ships in three weeks damages your standing with the index over time. Third, Catalog syncs what you wrote. If your title is “Classic Crew Tee,” Catalog faithfully delivers “Classic Crew Tee” to a product query fanout that is looking for “organic cotton crew neck men’s regular fit.”
I keep a mental list of the moments in my Shopify years when a merchant assumed a platform integration had done the work for them. Google Shopping feeds were the classic case: the feed was connected, the products were disapproved for missing GTINs or mismatched prices, and nobody looked for four months. Shopify Catalog is a better default than any of those integrations were, but it is a delivery mechanism, not a merchandising decision. The delivery is handled. The decision is still yours.
The right first move at every stage is to rewrite the titles and attributes on your top 20 revenue SKUs in the language a shopper types into ChatGPT, because that is the input to product query fanout and nothing downstream matters if you miss the candidate pool. Beyond that, the priorities split by stage, and the split matters because the failure mode I see most often at $500K to $2M is buying a tool before fixing the data the tool would report on.
If you are doing $20K to $50K per month, spend your 90 minutes on three things. Run the self-test from the Shopify AI Visibility Audit: describe your product category with constraints and no brand name, in ChatGPT’s free tier, and see who appears. Fix the 20 titles. Then change your post purchase review request from “how did we do” to two specific questions, one about fit or sizing and one about how the product held up, because 787 versus 352 median reviews was the second largest rank gap in Profound’s data and detailed reviews are the only version that compounds. Do not buy monitoring software yet. You cannot monitor your way out of a data problem.
If you are doing $100K to $500K per month, add the technical layer that Resoneo’s findings make concrete. The snippet ChatGPT reads in instant mode is roughly 200 characters starting at your H1, and it ignores your meta description entirely. On most Shopify themes those 200 characters are eaten by a breadcrumb, a vendor name, and a star widget. Move your one sentence product answer directly under the H1. Confirm OAI-SearchBot is allowed in robots.txt alongside GPTBot. Check that your product pages stay under 4 MB and render their core facts without JavaScript, because the fetcher rejects heavier pages outright and does not execute scripts. And note that the cache strips JSON-LD before the model reads a page: keep your Product schema for Google, but write the facts in visible text too. Then open ChatGPT’s lockdown mode and ask for your top product pages. If it serves them from cache, you are in the index. If it cannot, you have a crawl problem before you have a ranking problem.
If you are doing $1M per month or more, you can afford to think about the 12.7% and the one third. On the feed side, OpenAI now accepts direct product feeds by SFTP, file upload, or hosted URL, and a direct feed with 15 minute refresh keeps price and availability accurate at a freshness Catalog sync may not match for high velocity SKUs. On the community side, one third of shopping citations coming from Reddit is not an invitation to seed threads; it is a reason to answer real ones under your own name and to make sure the threads that already exist are accurate. This is also the stage where the AEO monitoring tools I compared in the 18 best AEO tools for ecommerce earn their subscription, because you now have enough SKUs and enough prompts that manual testing stops scaling.
Yes, and I would put the odds that OpenAI’s own index is the default shopping retrieval path by early 2028 well above even, because every piece of evidence points the same direction and the economics only push harder as the free tier grows. The 8% slice becomes 20%, the 20% becomes the default, and the scraped Google pipe becomes a fallback for live prices the way Resoneo already describes it. I am willing to be wrong about the timeline. I am not willing to bet a merchant’s channel strategy on the direction being different.
Here is what I think most coverage of Peec’s piece will get wrong. The headline is “ChatGPT built its own index,” and the temptation is to treat that as a new platform to optimize for, with a new set of tricks. But look at what the index actually rewards in the shopping data: complete attributes in shopper language, deep specific reviews, accurate availability, honest community presence, and a page that states its one important fact immediately. Every one of those was worth doing in 2019 for Google Shopping and worth doing in 2023 for a human buyer. The index did not invent a new game. It removed the shortcuts from the old one.
That is why the 18 month filter cuts hard against most of what is being sold right now. llms.txt files, AI-specific schema, “GEO agencies” promising Labrador placement: none of it appears in the ranking correlations, and Resoneo found the cache strips your structured data before the model reads it anyway. What survives the filter is boring. Fix the top 20 SKUs. Ask for reviews that say something. Keep the catalog honest. Watch Google’s own shopping AI Overviews with the same discipline, because the merchants who do this work once are showing up in both.
The merchants I watched win at Shopify were rarely the ones who found the channel first. They were the ones whose product data was ready when the channel found them. The ChatGPT shopping index is being built in public, at 35,000 requests per hour. It is going to find you. The only question you control is what it reads when it does.
No, ChatGPT does not use Bing to select shopping results in 2026. Resoneo’s analysis of 88,000 ChatGPT search results found that product cards are served from internal pipes drawing on OpenAI’s own merchant feeds, with Google kept as a reference for prices and reviews, and that shopping never touches the web search pipeline at all. Profound’s separate analysis of 812,190 product cards found commercial information sourced from ChatGPT’s own fallback search, Google index signals, and live merchant pages, with Bing absent from the list. Microsoft still supplies some grounding to ChatGPT through its Web IQ platform, and Bing appears as a source in Deep Research, but for product cards specifically, Bing rankings are not a useful proxy for visibility.
If you are on Shopify and sell to US buyers, your products are already delivered to ChatGPT’s merchant feed layer through Shopify Catalog with no additional integration work. What you still have to do is make the data worth surfacing: rewrite titles and attributes for your top revenue SKUs in the language shoppers actually type, keep availability and pricing accurate, and confirm that OAI-SearchBot and GPTBot are allowed in your robots.txt so the crawl side of the index can read your product pages. Roughly 87% of product cards in Profound’s June 2026 sample came from web crawl rather than feeds, so the product page itself matters as much as the Catalog sync. You can verify indexing by asking ChatGPT in lockdown mode to show your product page; if it serves a cached copy, you are in.
The two strongest correlations with top ranked product cards are review count and GPT tag presence. In Profound’s analysis of 200,975 rank 1 cards versus 214,301 cards at rank 4 or lower, top cards had a median of 787 reviews against 352 for bottom cards, and 2.39% of top cards carried a machine-assigned label such as “best value” or “premium” versus 0.98% of bottom cards. Promotional pricing showed a modest 13% lift. URL length and product name length showed no lift. At the merchant list level, OpenAI states it ranks merchants on availability, price, quality, and whether the seller is the maker or primary seller. Product results are not ads and are not influenced by OpenAI partnerships.
Your product shows up in Google but not ChatGPT most often because ChatGPT’s free tier grounds answers on its own index, which overlaps very little with traditional search rankings and reads your page differently. Resoneo found only 1.5% of URLs from ChatGPT’s own index appear in Bing’s top 20, that the meta description is ignored entirely, and that in instant mode the model sees only your title plus roughly 200 characters starting at the H1. If those characters are a breadcrumb and a review widget, the model learns nothing about the product. Other common causes are blocked crawlers in robots.txt, product pages over 4 MB which the fetcher rejects outright, and core product facts rendered only through JavaScript, which the fetcher does not execute.
A direct product feed is worth it for Shopify merchants doing roughly $1M per month or more with high velocity SKUs where price and availability change faster than Catalog sync reflects. OpenAI accepts feeds by SFTP, file upload, or hosted URL and supports refresh intervals as short as 15 minutes, which keeps a fast moving catalog accurate at the freshness a shopping surface expects. For most merchants below that scale, Shopify Catalog already delivers the feed and the better use of time is improving the data inside it: complete attributes, accurate stock, and detailed reviews. Feeds accounted for 12.7% of product cards in Profound’s June 2026 sample, so a direct feed complements crawl visibility rather than replacing the work on your product pages.
