OpenAI's Labrador index and ChatGPT shopping: US vs Europe, free vs paid

For most of 2025, the working assumption in AEO was simple: ChatGPT search runs on Bing, so rank in Bing and you are in ChatGPT. That assumption is now out of date. Researchers spent 2026 watching what ChatGPT actually fetches, and what they found is a layered system: OpenAI's own index, several outside search providers, merchant product feeds and the user's own context, mixed differently depending on where the user is and which plan they pay for.

This guide is for AEO and GEO engineers and the D2C teams they work with. It covers what is known about OpenAI's reported "Labrador" index, every source ChatGPT draws on when it shows products, how that changes between the US and Europe and between free and paid users, and what to change in your stack because of it. Where something comes from OpenAI's own documentation we say so. Where it comes from outside research, we say that too, because a lot of what is written about this topic is guesswork repeated until it sounds like fact.

Status as of October 2026. This area changes week to week. Check the sources at the end before you act on any single detail.

The short version

• Labrador is the name researchers found for OpenAI's in-house retrieval index, which appears to be a family of vertical indexes including one for shopping. OpenAI has not confirmed the name.

• ChatGPT blends sources. Its own index sits alongside scraped Google results, Microsoft's search APIs, licensed partners, merchant product feeds and Shopify Catalog.

• Free and paid users do not get the same retrieval. One analysis found free accounts get products from OpenAI's own index almost every time, while paid thinking mode drew about 75% of its results from scraped Google.

• The organic product layer is not paid. OpenAI says product results are selected independently and are not ads. Sponsored products are a separate unit shown only to Free and Go users.

• The US has more direct levers. OpenAI's standard product feed upload targets US shoppers. Elsewhere, including Europe and India, product data reaches ChatGPT mostly through crawling, third-party metadata and platforms like Shopify Catalog.

• In-chat checkout is gone. Instant Checkout launched in the US in September 2025 and was dropped in March 2026. Discovery happens in ChatGPT, and the purchase happens on your site.

What is the Labrador index?

Labrador is the label researchers found on results coming from OpenAI's own retrieval system. Between 21 May and 21 July 2026, ChatGPT's server-sent events carried a field called result_source that showed where each search result came from. It only ever carried four values: Labrador, Bright, Oxylabs and SERP. Labrador is the one attributed to OpenAI itself. The other three point to outside scraping providers. Around 21 July, OpenAI stopped exposing that field, so the clean signal is gone.

From the same event data, job postings and earlier testimony, researchers describe Labrador as a family of vertical indexes rather than one index. The verticals reported are general web, PDF, YouTube, news (with separate tiers for the last day, the last seven days and older), arXiv, Wikipedia, local, finance, legal, medical, images and shopping. Each record reportedly stores the full page content, the crawl date and the publication date.

This was not built overnight. In testimony during Google's antitrust trial, ChatGPT's head of product said OpenAI began building its own index in 2023, with a goal of answering 80% of queries from it by the end of that year.

What the shopping vertical seems to do

The most detailed public teardown of ChatGPT's shopping events reported a retrieval stack that will look familiar to anyone who has built search:

• A lexical first pass (BM25-style) that narrows candidate sources to around 10.

• A rerank stage over a pool of 400 candidates, and a separate product pool of 400.

• Approximate nearest-neighbour vector search, with embeddings reported at 4,096 and 12,288 dimensions.

• Hybrid ranking. Sparse and dense retrieval combined, most likely merged with reciprocal rank fusion. That last part is an inference, not something seen directly in the data.

The same research found OpenAI testing its own index against scraped search results. An experiment named "prefer-index-over-serp-v3" was running on about 8% of chats in mid-August 2026. As of 2 September 2026, five shopping experiments were live at once, including one named "chatgpt-shopping-noamazon". Experiment names tell you what is being tested, not what shipped, so treat them as direction, not rules.

Cached, not live

The index appears to cache pages rather than fetch them live for every answer. That makes sense for speed: in the HTTP Archive's data, 58% of pages take more than 0.8 seconds just to start responding. In one test, ChatGPT returned page content even with live browsing switched off. The practical effect: the version of your product page ChatGPT answers from may be days or weeks old. A price change on your store is not a price change in ChatGPT until the page is crawled again.

What we do not know. OpenAI has not announced Labrador, confirmed the name or published how much of ChatGPT's traffic it serves. Everything above comes from reverse-engineering by outside researchers, much of it from a single study. It is the best evidence available, not documentation.

Every source ChatGPT uses when it shows products

Labrador is one input. When a shopper asks ChatGPT what to buy, the answer can draw on six kinds of source:

1. OpenAI's own index (Labrador). Built by OpenAI's crawler, OAI-SearchBot, and cached. OpenAI documents OAI-SearchBot as the bot that surfaces websites in ChatGPT's search features.

2. Scraped Google results. Delivered through outside providers. In one test, a site with no organic traffic saw a Google Search Console spike on exactly the days it was queried in ChatGPT.

3. Microsoft. Bing appears in Deep Research events, and Microsoft says its Web IQ service powers ChatGPT alongside Copilot.

4. Merchant feeds and catalogues. OpenAI's help documentation says product results use "structured metadata from first-party and third-party providers," such as price and product description. Merchants can send feeds through OpenAI's Agentic Commerce Protocol, and OpenAI says Shopify merchants are covered automatically through Shopify Catalog.

5. Licensed and partner data. Yelp and Tripadvisor are reported as licensed partners. These matter more for local than for products.

6. The shopper's own context. OpenAI says a product appears in the carousel when ChatGPT judges it relevant to the request, considering the query and context "such as Memory or Custom instructions."

After ChatGPT picks a product, it picks the merchants to show for it. OpenAI says the merchant list comes from product and merchant metadata supplied by third-party providers or by merchants directly, ranked on availability, price, quality, and whether the merchant is the maker or primary seller. For a D2C brand that last factor is an advantage worth protecting: your own store should be clearly identifiable as the brand's store, not one reseller among many.

Free vs paid: what changes

Different plans see different things, and at least one study suggests they are served by different retrieval mixes.

Retrieval mix

Resoneo, an SEO consultancy, captured 1,249 ChatGPT answers in July 2026 across free accounts, paid accounts and logged-out sessions in several countries. On free accounts, product results came through OpenAI's in-house index almost every time. In paid thinking mode, about 75% of the 16,407 search results it captured came from scraped Google and about 24% from the in-house index. That is one study, with a small paid sample, and the source labels it relied on have since been removed. But the direction matters. A brand that only optimises for one pipeline is only visible to part of the audience.

Ads and sponsored products

OpenAI began testing ads with logged-in adult users on the Free and Go plans in the US on 9 February 2026. Plus, Pro, Business, Enterprise and Edu accounts do not see ads, and neither do Temporary Chats. Free users can switch to an ad-free experience with lower usage limits. OpenAI says ads run on separate systems from the chat model and that advertisers cannot shape, rank or alter answers. Sponsored product cards appear below the answer, are labelled, and come from a separate eligibility flag in the product feed.

For measurement, this means one thing. A sponsored card is not organic visibility. If your tracking counts both together, your "AI visibility" will move with your ad budget.

Memory and personalisation

Since product selection can use Memory and Custom instructions, the plan affects how personal the results are. Free users got a lightweight version of chat-history memory in June 2025 that recalls recent chats. Plus and Pro get a longer-term version. A shopper who has told ChatGPT about their sensitive skin or their budget will see different product picks from one who has not, and paid users are more likely to carry that context.

Shopping research

The longer, guided shopping research mode launched on 24 November 2025 for logged-in Free, Go, Plus and Pro users. It asks follow-up questions and builds a buyer's guide, which means it pulls in more sources and more product detail than a single carousel.

US vs Europe: what changes

Merchant feeds

This is the biggest difference. OpenAI's product feed specification says the standard upload "currently targets the US". Its target_countries field accepts US, CA and MX, but standard uploads are treated as US regardless, and other markets need OpenAI to confirm the integration and its allowed countries and currencies. In practice, a US brand can hand ChatGPT a structured, frequently refreshed product feed. A European or Indian brand selling at home mostly cannot yet, so its product data reaches ChatGPT through crawling, third-party metadata and catalogue platforms. OpenAI says Shopify merchants are covered through Shopify Catalog. Check whether that applies in your market.

Checkout

Instant Checkout launched on 29 September 2025 for US users buying from US Etsy sellers, with Shopify merchants to follow. It was a US feature from the start. In early March 2026 OpenAI dropped direct checkout inside ChatGPT after weak uptake, and moved purchases to merchants' own sites and apps. The Agentic Commerce Protocol it built with Stripe remains, now focused on discovery. Today the US and Europe look the same at the point of purchase: the shopper leaves ChatGPT to buy.

Ads

Ads started in the US in February 2026 and expanded to 31 European markets from late August 2026, again only for Free and Go users. One European difference matters for measurement: OpenAI's help documentation says personalised ads are not offered in the EEA or Switzerland. European sponsored placements are based on the current conversation rather than the user's wider history.

Memory

Chat-history memory reached the EEA, the UK, Switzerland, Norway, Iceland and Liechtenstein later than the US. In those regions, users see a prompt before it is switched on, rather than having it on automatically. Fewer European users may carry personal context into shopping prompts, so results there may be less personalised. That is a reasonable inference, not a measured fact.

Organic results

Product carousels rolled out worldwide in April 2025, including for logged-out users. A European store can appear and be cited without being in any merchant programme, as long as its pages are reachable and its product facts are readable. Without a feed, the quality of your pages, your structured data and your presence on third-party sources carries more weight.

What to change in your stack

For OpenAI's index

• Allow OAI-SearchBot in robots.txt, and check your CDN or firewall does not block it. OpenAI says sites that opt out of OAI-SearchBot will not appear in ChatGPT search answers. GPTBot, the training crawler, is a separate decision.

• Serve product facts in the initial HTML. Price, stock, variants, material and returns should not depend on JavaScript. The major AI crawlers fetch JavaScript files but do not execute them.

• Write the H1 and title for the snippet. Resoneo found the H1 appeared in 83.6% of the snippets ChatGPT pulled from pages that had one, and snippets were cut at about 200 characters. Put the product type and key attributes there, not just the product name.

• Keep pages fast and fresh. The index caches pages, and freshness is stored as a field. Show a real last-updated date on buying guides and keep it honest.

For the Google and Microsoft layer

• Do not drop classic SEO. If paid thinking mode leans on scraped Google results, your Google rankings still feed ChatGPT answers for your highest-intent shoppers.

• Keep Google Merchant Center and Bing clean. Both are inputs to the wider answer-engine ecosystem, and they are open to merchants outside the US today.

For the feed layer

• US brands: make sure your catalogue reaches ChatGPT, through Shopify Catalog or by applying to OpenAI's merchant programme. Fill the nine required fields: item_id, title, description, url, brand, seller_name, image_url, availability and price.

• Add the recommended fields. These include review_count, star_rating, return fields and shipping. OpenAI says recommended fields improve ranking, relevance and trust.

• Update the feed when a sale starts or ends and when stock changes. The specification is explicit that dates in the feed do not schedule changes for you.

• Treat eligibility flags separately. Search, checkout and ads eligibility are independent fields. Being eligible for ads does not make you more visible organically.

• Brands outside the US: until the standard upload opens in your market, your structured data has to do the feed's job. Mark up every product page with Product schema, including price, currency, availability and brand, and keep it in sync with the store.

For merchant ranking

• Make your store unmistakably the primary seller. Use the same brand name across your site, feed and marketplace listings, and set seller_name to the brand.

• Keep prices and stock consistent. If a reseller lists your product at an old price, ChatGPT may show that merchant, or that price, instead of yours.

For measurement

• Test from more than one vantage point. Use logged-out, free and paid accounts, and US and European locations. One account in one country shows you one slice of the system.

• Separate personal context. Use Temporary Chat or clean accounts for baseline tracking, because memory changes product picks.

• Keep sponsored and organic apart. Track sponsored cards separately from organic mentions, or your numbers will move with ad spend.

• Judge trends, not single answers. Track the same prompts weekly. Retrieval experiments run on slices of traffic, so the same prompt can return different products on the same day.

Where Eldor fits

None of the levers above help if you cannot see what ChatGPT says about your products, where it got that information and whether it is right. That is the gap Eldor is built for.

• Tracking by region and plan. Eldor runs your prompts from the markets you sell in, such as the US, Europe and India, and as logged-out, free and paid users. You see each slice of the system, not just the one your own account happens to show.

• Clean, comparable runs. Prompts run in fresh sessions with no memory, and sponsored placements are tracked separately from organic answers, so your numbers move with your visibility, not with someone's ad budget.

• Weekly tracking across engines. Eldor tracks 100 or 300 shopper prompts every week on ChatGPT, Google AI Overviews and Perplexity, and on more engines on higher plans. Retrieval experiments and caches mean a single check tells you little.

• The sources behind every answer. Eldor shows share of voice and the sources each engine cites. You see whether answers about your category come from your own pages, from marketplaces or from a listicle you have never heard of.

• Product truth. Eldor checks every SKU against your live store and flags every answer that quotes the wrong price, stock or policy. With cached pages and slow-moving third-party data, this is where D2C brands lose shoppers without noticing.

• Feed and schema readiness. Eldor checks your product data against what ChatGPT reads: the required feed fields, valid availability values and the Product schema on every page, market by market.

• Fixes at the source. Eldor enriches missing product attributes and rewrites thin product pages. The same clean data then serves your pages, your structured data and the catalogues engines read.

• An agent storefront. On the Growth plan, Eldor gives AI shopping agents a structured way to read your catalogue and act on your store, built on open protocols such as MCP and UCP.

The free Eldor AI audit is the starting point. It runs 25 shopper prompts for your category on ChatGPT and Google AI Overviews and checks 100 of your products against your live store, with the competitor and source behind every lost answer. To turn that into a weekly routine, our guide to running AEO like a performance channel covers the scorecard and the review rhythm.

FAQ

What is OpenAI's Labrador index?

Labrador is the name researchers found in ChatGPT's server events for OpenAI's own retrieval index. It appears to be a set of vertical indexes, including one for shopping, built from pages crawled by OAI-SearchBot and cached. OpenAI has not publicly confirmed the name or how much of ChatGPT's traffic it serves.

Does ChatGPT still use Bing?

Yes, among other sources. Research in 2026 found ChatGPT combining its own index with scraped Google results, Microsoft's search services and several other providers. Bing alone is no longer a complete ChatGPT strategy.

Are ChatGPT's product recommendations paid?

Not the organic ones. OpenAI says product results are selected independently and are not ads. Sponsored products are a separate, labelled unit shown only to Free and Go users in markets where ads are live.

Do free and paid ChatGPT users see different products?

They can. Free and Go users may see sponsored products and paid users do not. Memory, which can shape product picks, is fuller on paid plans. And one 2026 study found paid thinking mode relied much more on scraped Google results than free accounts did.

Can a European or Indian store appear in ChatGPT shopping?

Yes. Product results appear worldwide and a store does not need to be in a merchant programme to be shown or cited. What is US-only for now is OpenAI's standard product feed upload, so stores outside the US depend more on crawlable pages, structured data and third-party catalogues.

Can shoppers still buy inside ChatGPT?

Not in most cases. OpenAI dropped Instant Checkout in March 2026, and purchases now happen on the merchant's own site or app. The Agentic Commerce Protocol remains, focused on product discovery.

Sources: independent analysis of ChatGPT's server-sent events and shopping experiments (September 2026); Search Engine Journal on Resoneo's analysis of ChatGPT's search index (August 2026); OpenAI Help Center, "Shopping with ChatGPT Search" and "Ads in ChatGPT"; OpenAI developer documentation, Agentic Commerce key concepts and product feed specification; OpenAI, "Introducing shopping research in ChatGPT" (November 2025), "Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol" (September 2025), "Our approach to advertising and expanding access to ChatGPT" (January 2026), "ChatGPT Ads expands across Europe" (August 2026) and "Powering Product Discovery in ChatGPT"; Forrester and Shopifreaks on the end of Instant Checkout (March 2026); OpenAI, "Memory and new controls for ChatGPT"; TechCrunch on ChatGPT shopping results (April 2025); OpenAI, "Overview of OpenAI crawlers"; Vercel, "The rise of the AI crawler" (December 2024).