You can’t improve what you haven’t measured. For AI answers, the usual measurement is one question typed into ChatGPT one afternoon and a screenshot of whatever came back. It may show your store or leave it out, and either way it is one answer to one wording on one day.
A better measure starts from about twenty questions your shoppers have already typed. Ask each one more than once, in a fresh chat, on the AI models your shoppers use, and count how often your store is named and which pages the answers cite. That count is where you stand. Your analytics shows part of it, and the last section says which part.
The AI models in this guide are ChatGPT, Gemini, Perplexity, Claude and Google’s AI Overviews. The examples follow Tarrowby, a made-up store that sells travel bags. The other brands and websites are real, and what they are shown saying is invented. The product is the Overnighter 30, a backpack for work trips of one to three nights.
How is a question to an AI model different from a search?
A search is a few words, and a question to an AI model is a sentence with the shopper’s situation inside it. Search had keywords, short phrases that tools list with a monthly count of how many people typed them. Google says (opens in a new tab) people using its AI features ask longer and more specific questions, then follow-up questions to dig deeper.
The situation usually decides the answer. A parent with a bad back asks which baby carrier takes the weight off their back on a long walk, and the bad back is the part a useful answer has to meet.
No AI company publishes how often any question is asked, so there is no list of the biggest ones to work from. Aim for a small set that covers the different ways your shoppers ask. This chapter calls it your question list, and builds one of about twenty.
Why does the wording change who is named?
Each wording sends the AI model off on different searches, so three wordings a shopper would call the same question can bring back different brands. Ask for a backpack for a work trip, then for business travel, then for one that looks professional.
Asking the same wording twice changes the answer too. A study by SparkToro and Gumshoe (opens in a new tab) found that the same list of brands came back less than once in a hundred times, and How AI search decides which brands to recommend goes through what else they found.
So write several wordings for each subject. Ask each one more than once before reading anything into it, because a single answer is a single sample.
Where are the questions shoppers have really typed?
Most of them are in places you already own, and Shopify now shows you some of the questions AI shopping tools are asked about your store. These are the places to look.
- Your support inbox and chat. Search them for question marks.
- Your reviews, read for what buyers were unsure of before they bought.
- Your store’s own search box. Shopify’s Search & Discovery app (opens in a new tab) lists what shoppers searched for most, and the searches that found nothing.
- Google Search Console, Google’s free tool for site owners, which lists the searches your pages appeared in.
- The People also ask box, the list of related questions Google shows when you search for your kind of product.
- Reddit and Quora threads where shoppers in your category ask each other.
- Shopify’s Knowledge Base app (opens in a new tab), which keeps a log of the questions shoppers asked about your store in AI shopping conversations, and marks the ones that went unanswered.
Shopify’s log is the closest you can get to the questions shoppers put to AI about your store, in their own words. The searches that found nothing are worth a slow read too. A store selling running shoes for wide feet may find them full of wide toe box
, a phrase its own pages never use.
Copy each question exactly as typed, lower case and misspellings included, and note where it came from. Tarrowby’s Search Console turned up whats a good backpack for a 3 day work trip
. Nobody at Tarrowby would have written it that way, and the untidy version is the one shoppers send.
Which kinds of question belong on the list?
Four kinds of question do four different jobs, so give each question on your list a label. Each kind tells you something different when the answers come back.
- An open request asks for a recommendation and names no brand, like the work-trip question. A useful answer to it names products. If the answers give buying advice and name nothing, replace it with a wording shoppers use that does bring back products.
- A request with a condition adds something the shopper needs, such as a bag that opens like a suitcase but looks fine in a meeting. The condition should change who is named, and if the same brands come back with it and without it, no page the AI model read states that need plainly. Take each condition from something a shopper typed, so the list tests what shoppers need.
- A comparison sets your brand against one other. A useful answer gives a reason for each side, and the reasons it gives for yours show which of your facts got through.
- A fact check names your product and asks one thing with a right answer. It is the first of the two questions in chapter 1, and it tells you whether your own pages are being read.
How do you word them so the answers mean something?
Three mistakes spoil a question list. Each produces answers that look fine.
The first is a vague word in a fact check. Words like best
and good quality
mean something different to every reader, so the AI model answers a different question from the one you meant. Keep them for open requests, where shoppers really use them.
The second is writing the answer into the question. Ask whether the Overnighter 30 is the best work backpack and you get a polite yes with nothing in it.
The third is asking in a chat that already knows you. OpenAI says (opens in a new tab) ChatGPT can use what it remembers about a person when it decides what to search for, so a chat where you have talked about your own store can steer the answer toward it. Ask each question in a fresh chat, meaning a new conversation with memory switched off, or one opened while signed out. For Google’s AI Overviews, search in a private browser window.
How do you run the first count?
Put about twenty questions to each AI model your shoppers use, ask each one at least twice, and write down what comes back. A spreadsheet is enough.
- Choose five or six subjects your shoppers ask about, and write three or four wordings of each, across the four kinds.
- Ask each question in a fresh chat on each AI model, then ask it again in another fresh chat.
- For each answer, note which brands it named and which pages it cited. Cited pages are the pages an answer links to as the places it read something.
- Check what each answer says about your own products against your documents, such as the design spec or the supplier’s page, and note which document you had to open.
The last step is the reason to do this by hand once. Each time you reach for a document, you learn where a fact lives and whether any of your pages says it.
Tarrowby ran twenty questions on five AI models, asking each twice, which made 200 answers. Tarrowby was named in 14 of them. Peak Design was named in 118, Osprey in 97 and Patagonia in 64. The pages cited most were Outside’s 2024 review of work backpacks, YouTube reviews, Amazon listings and a Quora thread asking whether a backpack is fine for business trips.
This guide calls that number, named in 14 of 200 answers, the count. It is how many answers named your store out of how many you collected, and asking the same questions again later shows whether it has moved.
Keep two things from the first count besides the number. The list of cited pages is where chapter 6 starts. The reasons the answers gave for the brands they named are what chapter 3 works from. In Tarrowby’s count, Peak Design’s bag was named for opening like a suitcase, and Osprey’s for shoulder straps that tuck away.
Is AI already sending you shoppers?
You can see some of what AI sends you, and the count of answers stays the firmest measure. There are three places to look, and each shows only part of what happens.
Google Search Console has a Generative AI performance report (opens in a new tab), rolled out to every site by 31 August 2026. It shows how often links to your site appeared in AI Overviews and AI Mode, by page, country and device. These are impressions, the number of times a link was shown, and the report has no clicks. Clicks from Google’s AI features (opens in a new tab) are counted in with ordinary search clicks.
Shopify shows orders placed through AI channels (opens in a new tab) in your admin, with the channel or the referring site they came from.
Visits are the least complete. OpenAI has said (opens in a new tab) that ChatGPT adds utm_source=chatgpt.com to the links it sends, a tag that lets analytics tools tell those visits apart. If your store uses Google Analytics, visits from the AI assistants it recognises, such as ChatGPT, Gemini and Claude, are grouped in a channel called AI Assistant (opens in a new tab). Some visits from AI apps arrive with nothing to say where they came from, and they are counted as direct, alongside people who typed your address.
Other shoppers are never joined up at all. Someone who reads an answer on their phone and buys two days later on a laptop arrives as a new visitor, with nothing to connect the order to the answer.
So judge the work on the count of answers, and treat the visits and orders you can see as the least AI is sending you. Set no target for visits from one AI model, because those numbers miss too much to steer by.
Twenty questions on five AI models, asked twice, is two hundred answers to read, and a routine that tedious stops being done. Indaeo AI keeps it going. It finds the questions shoppers ask on your store, on Google, in AI models and on social media, and keeps asking five AI models the ones you confirm, saving every answer.
In chapter 3 you take these questions, and the reasons the answers gave for the brands they named, and decide what each of your products should be known for.