Ask ChatGPT for a good toiletry bag that hangs up in a small hotel bathroom
, and you get a short list of brands with a reason beside each one. A shopper who trusts that list may never look at a store that was left off it.
The AI model runs its own web searches while it writes, reads some of the pages those searches turn up, and names brands from what those pages say. The reason beside each brand usually comes from the same pages. For a store, that means what other sites say about your products counts for a great deal, and so does how plainly your own pages say who each product is for.
The AI models in this piece are ChatGPT, Gemini, Perplexity, Claude and Google’s AI Overviews, and where one works differently the piece says so. The examples follow Tarrowby, a made-up store that sells travel bags. The other brands are real. What they are shown saying is invented.
Where does an AI model get its answer?
An AI model answers from two places. The first is what it learned in training. It was trained on a vast amount of text collected up to a certain date, and from that it already has a picture of which brands are known for what. OpenAI lists this (opens in a new tab) among the things ChatGPT considers when it picks products, describing it as the answer the model gives before it looks at any new search results.
That picture favours brands that have been written about for years. Researchers at the University of Toronto (opens in a new tab) asked ChatGPT and Perplexity fifty questions about the best and most popular colas. Coca-Cola, Pepsi and the other big names took more than half of ChatGPT’s mentions, and craft sodas took about one in eight. The researchers put this down to how prominent the big brands are in the pages the AI models read, together with what the models already believed. They tie it to questions that name no particular need.
The second place is the web as it stands when the shopper asks. For a shopping question the AI model usually searches, because prices and stock keep changing after training ends. It opens some of the pages it finds and writes its answer from them. This is the part a store can change, and the rest of this piece is about it. Three ways an AI model learns about your products covers the training side and the software AI companies send to read your site.
What does it search for?
It runs searches that are narrower than the shopper’s question. OpenAI says (opens in a new tab) ChatGPT rewrites a question into one or more targeted queries
, then sends more specific ones after reading the first results. Google calls the same thing query fan-out (opens in a new tab) and describes it as running several related searches for one question.
Google has given a shopping example of its own (opens in a new tab). A shopper tells Google’s AI Mode they want a bag for a trip to Portland, Oregon, in May. The AI model runs several searches to work out what makes a bag good for rainy weather and long journeys, then uses what it learns to suggest waterproof bags with pockets that are easy to reach. The shopper never typed waterproof
. The AI model worked out that it mattered and went looking for it.
The toiletry-bag question works the same way. A toiletry bag that hangs up in a small hotel bathroom
could turn into one search for toiletry bags that hang and another for bags that stand up on a narrow shelf. Each search brings back its own pages. A page that answers one of those narrow needs well can be read even if it would never rank for the whole question.
The shopper’s own words decide which searches run. OpenAI gives the example of a budget (opens in a new tab). A shopper who says they want to spend $30 gets an answer that pays more attention to price. A trip or a skin type changes the searches in the same way, and so can what ChatGPT remembers about the shopper from earlier chats. Google says (opens in a new tab) people using its AI features ask longer and more specific questions, with follow-up questions on top.
Those specific questions are where a smaller brand has its best chance. Each one leads to searches for one exact need, and any page that meets that need well can be read. The guide’s first chapter follows one answer from the question to the finished list.
Which pages does it believe?
It leans on the pages that other pages back up. Google says (opens in a new tab) its AI Overviews are integrated with the same systems that rank ordinary search results, and are designed to show only information that is backed up by top web results. For Google’s AI features, a page that already does well in ordinary Google search is in a far stronger position to be used.
Most of the weight sits with sites the brand does not own, such as review sites and magazines. The Toronto researchers (opens in a new tab) sorted every source cited in answers to which brand is best
questions. When the questions were about smaller, less famous brands, 95 of every 100 sources ChatGPT cited were independent sites like these, and about 5 were the brands’ own websites. Gemini cited brands’ own sites more often, about one source in five, and independent sites for about two in three.
Ahrefs looked at 75,000 brands (opens in a new tab) and found the same pattern in Google’s AI Overviews. Of everything it measured, the number of times other websites mention a brand tracked most closely with how often AI Overviews named that brand. Links tracked it less closely. Ahrefs notes that this shows the two go together without proving that one causes the other.
In the example, Osprey’s claim that its toiletry kit hangs from a hook appears on osprey.com and is repeated by a review site, a retailer and a video. Tarrowby’s claim that its kit also stands upright on a shelf appears on tarrowby.com and nowhere else yet. An AI model weighing the two finds one claim backed up three times over and the other resting on the seller’s word alone.
What does your own store count for?
Independent sites need something to repeat, and the facts about your products start with you. If you sell on Shopify, your product data already reaches ChatGPT through Shopify, and OpenAI says (opens in a new tab) the store has nothing extra to do. ChatGPT weighs that data, including the price and the product description, when it picks products.
Being among the products ChatGPT can show is the start. For any given question, your pages need to say in plain words that your product fits it.
Microsoft explains why the plain words matter (opens in a new tab). It says Copilot breaks a page into small pieces and judges each one, and that the sentences that get used are ones that still make sense when lifted out of the page. Its example is a product page, which should say 42 dB dishwasher designed for open-concept kitchens
instead of just quiet dishwasher
. The words designed for open-concept kitchens
tell an AI model which buyer the dishwasher suits.
Tarrowby’s page could say the kit hangs from a hook or stands upright on a shelf, for bathrooms with no counter space. Any product has a sentence like that. A face cream made for skin that reacts to fragrance tells a shopper more than gentle
, and a dog food for large dogs over seven tells them more than senior formula
. An AI model can lift sentences like these straight into an answer.
Reviews count as well, including the ones on other websites. OpenAI says (opens in a new tab) the review summaries in ChatGPT are drawn from reviews on public websites, and that a label such as Budget-friendly
can mean reviewers often mention good value.
Can you pay to be recommended?
The companies say no. According to OpenAI (opens in a new tab), the products ChatGPT shows are chosen by ChatGPT, are not ads and are not influenced by its partnerships, and any ads it shows are kept separate from the answer and clearly labelled (opens in a new tab). Perplexity (opens in a new tab) describes its product cards as not sponsored.
Google says (opens in a new tab) a page needs no special AI file or extra markup to appear in its AI features. It needs to be in Google’s index and able to show a snippet, the short preview under an ordinary search result.
Why does the answer change every time you ask?
An AI model writes a fresh answer each time it is asked, and the brands in it change from one answer to the next. SparkToro and Gumshoe (opens in a new tab) asked about 600 volunteers to put the same 12 questions to ChatGPT, Claude and Google’s AI, close to 3,000 times in all. Each question asked for recommendations, such as the best chef’s knives for an amateur cook with a budget under $300. The odds of getting the same list of brands twice were less than one in a hundred. The odds of the same list in the same order were closer to one in a thousand.
Shoppers also word the same need in very different ways. In a second round, the volunteers each wrote their own question asking for headphones for a family member who travels. Of 142 questions, barely two looked alike. ChatGPT can also use what it remembers (opens in a new tab) about a shopper when it searches, so two people asking the same thing can set off different searches.
What stays steady is how often a brand is named. Across the 994 answers to those 142 headphone questions, Bose, Sony, Sennheiser and Apple each turned up in 55 to 77 per cent of them. Rand Fishkin, who ran the study, concluded that how often a brand appears, across dozens to hundreds of questions each asked several times, is a reasonable thing to measure. He dismissed a brand’s position in the list as a measure, since the order changes almost every time.
In the example, Tarrowby was named in 2 of the 12 answers. A single screenshot would have shown it either winning or missing, depending on which answer happened to be on the screen.
What can your store do about it?
There are four things to work on, in roughly this order.
- Make sure AI models can read your pages. Three ways an AI model learns about your products covers which visitors to let in, and the guide’s first chapter shows how much of a page reaches them.
- Say plainly on each product page who the product is for and what it does well, in the words your shoppers use. Chapter 5 of the guide shows how to rewrite a page so an AI model can quote it.
- Get your products talked about on the sites AI models read, such as review sites, magazine roundups, retailers that stock you, creators’ videos and the forums where people in your category ask for advice. Make the claim you most want repeated easy for them to find.
- Keep count. Find the questions your shoppers really ask, which chapter 2 of the guide covers, then put each one to several AI models more than once and note how often you are named and which pages the answers cite.
Keeping all four going, for every product and every question your shoppers ask, is a job of its own. Indaeo AI and the Indaeo team take it on for your store. Indaeo AI finds the questions your shoppers ask, keeps putting them to five AI models, and records who each answer names and which pages it cites. Its agents draft the product page rewrites, the articles and the pitches to the sites AI models quote, and the Indaeo team works through them with you.