What ChatGPT, Perplexity and Google say about your store: a check you can repeat

To check what ChatGPT says about your store, ask 10 to 20 questions in your customers’ words, 3 times on each assistant, and log the date and country. Score 2 numbers, not a rank: how often your store is named, and how many of the claims in the answers match your own pages.

In November and December 2025, SparkToro and Gumshoe.ai asked ChatGPT and Google’s AI each of 12 brand recommendation questions, 60 to 100 times per platform. The same list of brands came back fewer than 1 time in 100 (Search Engine Journal, 30 January 2026).

The screenshot you took of what ChatGPT says about your store is 1 draw from that spread. It may come from the kind end or the harsh end, and you can’t tell which. A shopper who is quoted a price you stopped charging doesn’t email you to say so.

What holds still is a rate: how often your store is named across many answers, and how many of the claims in them match your own pages. Both can be counted by hand, free, in an afternoon.

Adobe forecast in September 2026 that AI driven traffic to US retail sites would rise 130% year on year over the 2026 holiday season (Adobe, a vendor forecast built on Adobe Analytics). That is the category figure, not yours.

Answers vary, but they are not mostly wrong

The spread is real and it can be measured. In SparkToro’s headphone test, 142 people each wrote their own prompt for the same need. The prompts shared little wording, and each of Bose, Sony, Sennheiser and Apple still appeared in 55% to 77% of the 994 responses. How often a brand appeared held steadier than where it came in the list.

The person asking changes the answer too. In November 2025, Lily Ray of Amsive spot checked prompts from tracking tools against her own ChatGPT. It rewrote a generic electronics prompt into a New York City search for her (PPC Land).

Prices are the claim shoppers act on, so the Product.ai study is worth reading closely. It put 220 shopping questions to the free and paid versions of ChatGPT, Claude, Gemini and Perplexity, 5 times each, and collected 8,794 responses (Business Insider, via DNYUZ, 25 September 2026).

The figure that travelled, 86%, is the share of questions that produced a repeatable factual conflict: a checkable disagreement, such as a different price or specification, that appeared in more than 1 response. It shows assistants disagree with each other. It doesn’t show which one was wrong. Product.ai is a startup that checks product claims against evidence, and the report doesn’t say whether the study was peer reviewed. It did find large gaps between the assistants, and between the free and paid versions of the same one.

Ask your customers’ questions, 3 times each, and log every answer

It takes an afternoon, a sheet and 3 assistants. Nothing to buy.

Write the questions in your customers’ words

Take 10 to 20 questions from where customers already put them: the queries in your Search Console, your support inbox and the wording in your reviews. Cover each type below.

Only the discovery questions test whether you are named, because every other type contains your name. Score those on accuracy.

Write down what is true today

Before you ask anything, copy the claims from your live pages into a reference sheet, with the date: price and pack size of your top sellers, delivery countries, costs and cutoffs, the returns window and its conditions, and any live offer. Every answer is marked against this sheet, not against memory.

Run each question 3 times on each assistant

• The assistants: ChatGPT, Perplexity, and Google’s AI Overviews and AI Mode where they show. Add Gemini and Claude if your customers use them.
• A new chat for every run, the same wording each time, in a private window and signed out where the assistant allows it. Note which you did.
• The country you ran it from. If you sell abroad, ask someone there to run the same sheet.
• The plan, free or paid. In Product.ai’s data the same assistant differed between its free and paid versions.
• Every link the answer cites or shows. They become your fix list.

15 questions, 3 runs and 3 assistants make 135 answers. For each, log the date, country, assistant and plan, the question and run number, whether you were named, each claim with right, wrong or can’t check, the links, and a screenshot.

Score appearance and accuracy, not rank

Appearance rate is the answers that name your store divided by all answers to your discovery questions, for each assistant. Accuracy is the claims that match your reference sheet divided by the claims you can check. Count your rivals the same way: the brands named most often in your discovery answers are your competition inside that assistant. Leave rank out. In SparkToro’s data, how often a brand appeared was steadier than its position.

45 answers per assistant show direction, not a decimal

A preprint by Ronald Sielinski (arXiv, 2026) repeated queries on 3 consumer product topics and found single run visibility numbers give a “misleadingly precise” picture. To get an interval 5 points wide on how often a domain was cited, it needed roughly 30 to 50 distinct queries per topic on Gemini, 90 to 100 on Perplexity and 150 or more on OpenAI’s SearchGPT. It measured citations of domains, not brand mentions, and it has one author.

Your 15 questions sit below that on every engine. At 45 answers, a 50% rate carries a margin close to 15 points either side, by our arithmetic, and that flatters you, because 3 runs of 1 question are not independent. So act on 3 results: a gap too wide for that margin, such as named in 5 of 45 answers on 1 assistant and 30 of 45 on another; a rival named far more often than you; and any wrong claim that comes back in 2 or 3 of 3 runs. A move from 12 of 45 to 15 of 45 is noise. A wrong price is worth fixing at any sample size, because checking 1 page costs minutes.

Fix the page the answer came from

For each wrong claim or missing mention, open the links the assistant cited. A wrong answer is built from 1 of 3 places: your product data, your own pages, or someone else’s pages.

Your product data

On Shopify, agentic storefronts are active by default for eligible stores, and your products reach AI channels through Shopify Catalog. ChatGPT is discovery focused: buyers check out on your own store. Google AI Mode and Gemini can also receive products through the Google & YouTube channel or Google Merchant Center. Shopify’s developer docs say agents act on search results that include product titles, images and variant details. So the price, title and variant names in your admin come first. You choose which AI channels you sell on under Sales channels, then Agentic.

Your policy and delivery pages

Put the returns window, delivery countries, costs and cutoff dates in plain text on 1 page each, with the date of the last update. Then check the assistants can read them: OpenAI says to allow OAI-SearchBot in robots.txt to appear in ChatGPT search, and Perplexity says the same of PerplexityBot. Each says a change can take up to 24 hours to apply. For Google’s AI features, there are no extra requirements or special files: a page has to be indexed and eligible for a snippet, and “still SEO” is Google’s own summary.

Pages that aren’t yours

If the cited page belongs to a retailer, marketplace, review roundup or comparison site, send its owner the correct detail with a link to your own page, and note the date you asked. Search your old price in quotes with the product name: any page that still shows it is a suspect.

Being named at all

If you are missing from discovery answers, the fix is slower, because the pages that mention you are other people’s: reviews, press, retailers that stock you, comparison pages. Google’s guide says chasing inauthentic mentions isn’t as helpful as it seems. Earn real ones where your customers read, and keep your own pages clear on what you sell, for whom and at what price.

Then ask again. Run the same sheet after every change and on a fixed day each month, and close a claim when the answer is right in 3 of 3 runs.

A check by hand stops at 1 person, 1 country and 1 day

The sheet can’t see what an assistant tells someone in Dublin on a Tuesday when you ran it from Leeds on a Friday. It holds only the history you keep, and a screenshot is weak proof of what an answer said on a date. AI visibility tools exist for that gap, and 4 describe themselves this way on their own pages (checked 23 October 2026; confirm prices there).

• Ahrefs Brand Radar covers AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini and Copilot. Custom Prompts (your own questions) start at $50 a month, free in paid Ahrefs plans from Lite up, and are checked daily. Its AI Visibility Index, a prebuilt database of prompts modelled on real searches, starts at $199 a month.
• Otterly.AI covers ChatGPT, Google AI Overviews, Perplexity and Copilot on every plan, with Claude, AI Mode and Gemini as paid extras. Plans run from $29 for 15 search prompts to $489 for 400, with daily tracking and unlimited countries. Its page lists 2 unlabelled sets of prices; these are the higher.
• Peec AI tracks ChatGPT, Perplexity and Gemini, with a separate AI Mode tracker, across countries, and offers CSV export, a Looker Studio connector and an API. Its pricing page lists Starter, Pro and Advanced monthly plans and a custom Enterprise plan, with no prices shown. Its docs say it collects answers by UI scraping, from a logged out user’s view.
• Profound has a free trial of 50 prompts run daily for 7 days on ChatGPT, Gemini and Google AI Overviews, for 1 region and 1 language. Its custom priced Enterprise plan tracks up to 9 answer engines, ChatGPT Shopping included, and a self serve Agency Growth plan gives 400 credits a month per client workspace, with no price shown. Its pages don’t say how it collects answers.

Two limits apply to every tool, because they come from the assistants. Answers move from run to run and with the person asking, as the studies above show. And Google’s guide says no third party tool has access to its internal ranking or AI systems, so a tool reports what assistants answered, not why. Rand Fishkin’s test for any provider is that it “shows their math”.

We make Obsession, which has a recipe in this area. It isn’t one of the 4 above, and its own page says what it does.

You don’t need an AI visibility tool while the sheet still shows fixes you haven’t made

A daily tracker shows you the same wrong price every day. If your first run finds no wrong claim and your store is named in most discovery answers, repeat the sheet monthly and after any change to a price, a policy or your product data. If it finds wrong claims, fix those first.

For agencies: report the repeat, not the screenshot

AgencyAnalytics surveyed 494 agency professionals between February and April 2026: 66% reported increased demand for AEO and SEO for AI driven search, the top new service clients requested in 2026 (a vendor survey). Clients will ask what changed, and January is when they decide who to keep. A report that survives the question names the questions, runs, dates and country; gives appearance rate and accuracy for each assistant with the number of answers behind them; lists each wrong claim with the page it came from, who fixed it and when; and shows the second run.

Google’s guide tells site owners to be wary of tools that promise ranking success. Promise the fix and a second run, and let the next month’s sheet show what moved.

Questions