How does ChatGPT decide which products to recommend?
ChatGPT combines structured metadata from first-party and third-party providers, such as price and product description, with other third-party content and the model’s own knowledge. Among merchants selling the same item, it ranks on availability, price, quality and whether the merchant is the maker or primary seller (OpenAI Help Center).
The request shapes the weighting. OpenAI’s example: “If you specify a budget of $30, ChatGPT will focus more on price.” A shopper’s constraint becomes a filter, and a product whose price or stock is not machine-readable cannot pass it. The gravel bike use case shows a cheaper shop dropped for exactly that reason. So the first lever is data the assistant can read, and the second is what other sites say about you.
Are ChatGPT product recommendations paid ads?
No. OpenAI states that product results “are selected independently by ChatGPT and are not ads”. A merchant cannot buy a place in them, so visibility comes from the inputs above, which is why the work sits in product data and in third-party coverage.
Does a product feed help with ChatGPT shopping?
Yes, it is the most direct input a merchant controls. Shopify merchants are included through Shopify Catalog with no extra work, according to OpenAI, and other merchants can apply to send a direct product feed.
The feed specification marks brand, price and availability as required, gtin as optional, and a group ID as conditionally required for variants (OpenAI product feed spec). A feed that disagrees with the page creates two prices, and the GTIN guide compares its identifier rules with Google Merchant Center. On Google, AI Overviews and AI Mode draw on the same index plus Merchant Center data, and may run a “query fan-out”, several related searches across subtopics for one question (Google Search Central).
Why does ChatGPT recommend my competitors and not my shop?
Usually because the competitor is easier to read or is named by more third-party sources. AI search engines show “a systematic and overwhelming bias towards Earned media”, meaning third-party authoritative sources, over brand-owned content, according to Chen and colleagues (arXiv, September 2025).
Our July 2026 audit shows both causes on one shop. A chocolatier near Geneva was named in 58% of the Google AI Overviews we tested, with guides, press and Reddit among the cited sources, yet none of its 26 product pages declared a price an agent could read (chocolatier use case). Coverage got it named, and missing data stopped the sale. Check both: whether reviews, guides and press in your market name you, and whether your price, stock and identifiers are fields.
When several shops sell the same product, the comparison moves to the offer itself: availability, price and delivery. That is the case for agentic commerce for multi-brand retailers, where the same GTIN is listed by many shops at once.
Do reviews and ratings matter for AI recommendations?
They count as part of the third-party content assistants read, and their effect differs by model. A study of AI shopping agents found that “sensitivities to price, ratings, and reviews vary sharply across models”, and that agents concentrate demand on a few products (arXiv 2508.02630, August 2025). Publish ratings in structured data where you have genuine reviews, and do not expect one rating to move every assistant the same way.
What does the GEO research say actually helps?
Adding cited sources, quotations and statistics raised a source’s visibility in generative engine answers by 30 to 40% in the foundational GEO study, and by 115.1% for a source ranked fifth in search, while the top-ranked source lost 30.3% on average (Aggarwal et al., KDD 2024). Keyword stuffing scored below the unoptimised baseline.
That rank-five result is the useful one for a mid-market shop, since evidence on the page helps most when the page is not already the leader. The boundary comes from the 2026 survey by Olivier Martinez, which reviewed 45 studies and found no technique with a stable, cross-platform effect on discoverability or traffic, and that rewrites aimed at citation can reduce how often a page is retrieved when they dilute its topic (arXiv 2607.14035, 15 July 2026). The GEO gains were measured on passages already in the model’s context, so they say nothing about being retrieved in the first place.
What should a shop do, in order?
| Step | What it fixes | Where it is covered |
|---|---|---|
| Let OpenAI’s and Google’s crawlers in, with product content in the HTML | Being retrieved at all | llms.txt and schema guide |
| Declare price, stock, GTIN and delivery as structured data, and send a product feed | Passing the shopper’s constraints | Machine Readability Sprint |
| Keep one product or one question per page, with the answer in the first sentences | The passage that gets retrieved carries the answer | GEO vs SEO guide |
| Add sourced numbers and named examples to guides and category pages | Being cited once read | This guide |
| Earn mentions in reviews, guides and press | The earned media bias | Autobasket Discover |
| Measure across repeated prompts | Telling progress from noise | share of model guide |
Nobody, ourselves included, can promise a fixed place in an AI answer. What can be done is to remove the reasons a product is dropped and to measure whether the share of answers naming you moves over several months.