Track your products in ChatGPT

Why a screenshot is not enough

Ask ChatGPT the same question twice: you do not always get the same answer. The products cited change, so does their order, and sometimes their price. A single answer therefore says neither that you are visible nor that you are not.

Tracking your products means answering a different question: out of a hundred answers to this request, how many cite my product? It is a proportion, and it is measured like a poll.

Step 1: pick the right questions

Start from what a customer would ask before knowing you. A good shopping query contains a category, a use and often a constraint:

• "Which electric bike for commuting, under €2,500?"
• "Which sunscreen for sensitive skin, fragrance-free?"
• "Which firm mattress for a 90 kg person?"

Three rules for the panel:

1. Do not name your brand. A query that cites you measures what the assistant knows about you, not whether it thinks of you.
2. Stay stable. The same questions, word for word, from one week to the next. Otherwise nothing is comparable.
3. Cover your categories. Ten to twenty queries per range are enough to start.

Keep queries that name a product ("Volt E1 or Kolven Air?") apart: they are useful to check accuracy, but they do not measure the same thing.

Step 2: fix the conditions

A measurement only means something if the conditions do not move.

• Web search on. Without it the assistant answers from memory, with outdated prices and stock.
• Language and country of the market. A question asked in French from France does not get the answer of a question asked in English.
• No history. A fresh conversation every time, with no memory and no custom instructions: otherwise the assistant answers you, not your customer.

Step 3: repeat, and know how many times

The number of answers sets the precision. For an observed share of 30%, the 95% confidence interval is:

With ten answers, a 30% share could just as well be 5% or 55%. Going from 30% to 40% from one week to the next, on ten answers, is not progress: it is noise. The formula and its limits are in our methodology.

Step 4: record what matters

For each answer, note:

• the products cited, in order, yours and your competitors';
• the price stated and the availability, when given;
• the specifications claimed (range, weight, material, warranty);
• the sources cited, meaning the pages the information comes from.

You get three numbers per product: its shelf share, its average position in the answers that cite it, and the number of answers containing an error.

Step 5: check against the catalog

A product cited with a wrong price is a problem, not a win. Compare every stated price with your price today, in the same currency, and every specification with your product page. Sort what you find:

• accurate: the answer says what the catalog says;
• missing: the answer does not mention the attribute;
• distorted: the value is close, but inexact;
• wrong: the value contradicts the catalog.

For each error, go back to the cited source. That is almost always where the fix lies: an old comparison, a marketplace, or your own markup.

Step 6: link to orders

A click coming from ChatGPT lands on your store with a referrer (chatgpt.com) and, most of the time, a utm_source=chatgpt.com parameter in the URL. Your analytics tool or your store can therefore count the visits and orders it brings.

Two limits to know: mobile apps do not always send a referrer, and a customer who reads the answer and comes back later through Google is not counted. The resulting figure is a minimum.

What a tool automates

By hand, twenty queries replayed four times on three assistants mean 240 answers to read every week. It can be done once, not every Monday.

skoup replays your queries on the assistants you choose, extracts the products cited, compares price, stock and specifications with your catalog, shows every share with its confidence interval and attributes orders to assistants. The details are on the Product page. The e-commerce GEO guide explains what to fix once the measurement is in place.