ChatGPT Searches Google Shopping to Create its Recommendations

Further research unlocks the hidden logic and processes behind ChatGPT’s search functions.

Interestingly, most of these depend on Google. For example, previous testing confirmed that ChatGPT uses SerpApi for its search functionality.

The reason why is so simple: Google has had 27+ years of building a great information ecosystem.

And now, an investigation from Oliver de Segonzac, Alexis Rylko, and Tom Wells also suggests that ChatGPT uses coded queries with Google Shopping to compile recommendations for its product carousel.

This can have a real improvement effect, so we are willing to test this for ourselves.

TL;DR: Focus on Optimizing Your Product for Google Shopping

The previous belief was that the information gathered from ChatGPT’s fan-out (Google searches run by ChatGPT in the background to generate a comprehensive response) was a key factor in product placement.

But our tests confirm that this is not the case.

ChatGPT does not create additional shopping inquiries that it sends to Google Shopping. Most of the time, Google Shopping results will shape the final products included in the ChatGPT response.

The usual fan-out question still comes up, but this informs the answer to the discussion related to product selection.

Important takeaways:

1. ChatGPT uses two sets of follower exit questions with answers through product carousels. The first is the context, used to form the written part of the answer. The second is Google Shopping searches, where ChatGPT adjusts these results.
2. Brands in ecommerce should focus on Google shopping optimization first and foremost. The top rated products here will likely be included in any of ChatGPT’s shopping recommendations.

How We Used Our ChatGPT Shopping Test

We set out to dive deep into the fan questions that shape the final answers, checking first if ChatGPT creates these buying questions, and then if the resulting products are compatible.

So the first step was to take a peek behind the curtain at the fan exit questions that are happening.

Step 1: Ask ChatGPT for product recommendations

We logged into ChatGPT and simply asked for product suggestions with some defined criteria. The goal here is to interact with the platform as the user would and provide specific guidelines for the resulting products.

In this example, we used: “best budget Android phones with good cameras”.

Step 2: Getting Follower Exit Questions

You can use a solution like Semrush Enterprise AIO to automatically detect these hidden background searches, but you can also use the Chrome Dev tools. Here’s what you need to do:

1. Open Chrome Dev Tools
2. On the Network tab > Download/XHR, filter using the chat URL that ChatGPT creates (the last part of the URL) starting with a number.
3. Click the refresh button to reload the conversation and allow the results to be captured
4. Use CMD+ F to search the dev tools panel “search_model_queries”
5. Here you will find the fan output under “questions”

In this case we have: “the best budget Android phones with great cameras 2025” and “what defines a budget Android phone and what are the budget phones with good cameras 2024 2025”.

Step 3: Getting questions from the shopping fans

Shopping fan queries are hidden with an extra layer of encryption.

This is complicated, but can be achieved by using these steps:

1. In the same file as before, search for “id_to_token_map”.
2. Find the piece of text next to this one that starts with “ey”. This is the piece of Base64 data we will need to decode.
3. Copy the entire summary (in this example it was about 500 characters) without the opening and closing quotation marks.
4. Paste the data into a free tool like Base64 Decode to make it readable and display questions for potential buyers.

Here is the result:

The most important thing is after the “question” where we see “cheap+android+smartphones+good+camera+2025”. This is a question of buying followers. Usually, you will only see 1-2 unique questions here.

Step 4: Comparing Results

Now we have all the information needed to compare the results. Just enter a shopping query in Google Shopping. Remember to check that the locale settings are the same for ChatGPT and Google Shopping for accuracy.

Here’s what we saw with an Android phone example:

Shopping on Google

ChatGPT

The first two entries are available for both Google Shopping and ChatGPT. In fact the seller, title, and price information are exactly the same.

Results: Fan-Out Query to Buy ChatGPT Advanced Search Layer

After testing 100 times, we found that the top ChatGPT product was included in the first 3 results of Google Shopping 75% of the time. There was also a big conflict with the second and third results.

Why Does ChatGPT Use Google Shopping Results?

ChatGPT uses Google Shopping Results because it is a rich source of data. This is not just a selection of products, but a library that includes reviews and live pricing information. Products can be recommended with confidence at the right prices and sellers.

Although ChatGPT is making efforts to move away from Google and Google Shopping (shown by the recent Etsy integration, and its new Shopping Research features), it cannot match this user experience yet. The Google ecosystem is very long.

Accurate live pricing information is important, especially when prices can fluctuate rapidly during ecommerce events like Black Friday.

What Does This Mean for Ecommerce Products?

It means that ecommerce brands need to consider their shopping feeds and make sure that any product pages or listings that feed this are always up to date. Google Shopping is the most important of these, no doubt ChatGPT and other platforms are working to create their own.

For product queries in ChatGPT, logic seems to be divided into: 1. Retrieving the context of the buyer’s guide and 2. Retrieving products from Google Shopping.

So right now, improving your Google Shopping results is one of the most important things to appear on the ChatGPT product carousel.

Additionally, brands need to prepare for the future. Chat shopping and agent e-commerce will continue to grow. Developments such as Google’s Universal Commerce Protocol (UCP) will facilitate these processes.

We are approaching a world where transactions happen directly in the conversation, without visiting your website.

Regarding Examination

This test was based on 100 command follower questions. Each prompt was run 5 times, with the most common top products in the carousel recorded for possible LLM adjustments.

Although this is a small sample size, it provides a clear starting point for AI shopping. We will continue to explore the shopping question with larger datasets to provide more insights.

You can do this test yourself. While using a signed in account, you may see different results depending on whether you are signed in, have a free account, or have a paid account. This was seen in the same ChatGPT and Google test we did.