How AI assembles an answer
When a shopper types “best oversized cotton T-shirts for a hot summer under ₴2,000” into ChatGPT, the model does not open a single website. It fans the query out into 5–15 related sub-queries — “how to choose cotton for hot weather”, “what oversized fit means”, “brands under ₴2,000 in Ukraine” — and assembles the answer from dozens of sources. The mechanism is called query fan-out. Kyle Risley, senior SEO lead at Shopify, describes it in Shopify’s official AEO guide as the key principle underpinning how modern AI engines produce results.
For each of those 15 sub-queries the AI looks for the page that covers it best. Not the most authoritative domain — the most precise answer. This is where Amazon loses: its page shows 200 T-shirts and is optimised for a click, not for a direct answer. Your page, which honestly explains “why this particular cotton does not cling in 30-degree heat”, wins that specific sub-query. The final citation is assembled out of small wins like that.
Where you genuinely win, and where you do not
AI answers from two sources at once. You have to grasp this, or nothing that follows makes sense.
The training data layer. What GPT-5, Claude and Gemini were trained on up to their cutoff date. Here Amazon is a giant, and that recognition cannot be bought in a month.
The RAG layer (search-augmented retrieval). This is where the model goes at the moment of the query: it searches for fresh pages in Google and Bing, reads them and cites them. Domain age does not decide anything here. Three things do: how precisely you answer the specific sub-query, whether you have structured data (schema), and freshness.
Today’s AI engines — particularly Perplexity and ChatGPT with web search — weight the two layers roughly equally. That is why a niche store that wins on RAG lands in the final answer alongside the giants. It is the most quickly monetisable advantage small e-commerce has ever had.
What to do today
Open ChatGPT and Perplexity. Ask the three questions shoppers in your niche ask most — not “buy a T-shirt” but “which T-shirts do not cling in summer”. Look at who the AI cites. If a Reddit thread or a small blog is in there, the RAG layer is currently open to stores like yours.
Next: why this will not last. That is part two.