ChatGPT Searches Google Shopping to Create its Recommendations

More and much investigation is unwrapping the hidden logic and processes that hap successful the inheritance of ChatGPT's hunt functions. 

Interestingly, astir of these thin connected Google. For example, previous experiments person confirmed that ChatGPT uses SerpApi for its hunt functionality. 

The logic why is beautiful simple: Google has had 27+ years to build an tremendous accusation ecosystem. 

And now, investigations from Olivier de Segonzac, Alexis Rylko, and Tom Wells person besides suggested that ChatGPT runs encoded queries done Google Shopping for composing its merchandise carousel recommendations. 

This could person existent implications for optimization, truthful we group retired to trial this for ourselves.

The TL;DR: Focus Your Product Optimization On Google Shopping

The erstwhile belief was that the accusation collected from ChatGPT’s query fan-out (the Google searches that ChatGPT runs successful the inheritance to build a broad response) was the main facet down merchandise inclusion.

But our research confirms that this isn’t the case. 

ChatGPT does create further shopping queries that it sends to Google Shopping. Most of the time, the Google Shopping results will past style the last products that are included successful ChatGPT’s response. 

The accustomed query fan-out still occurs, but this informs the conversational reply that’s alongside the merchandise selection. 

Key takeaways:

1. ChatGPT runs 2 sets of fan-out queries for responses pinch merchandise carousels. The first are contextual, utilized for forming the written portion of the answer. The 2nd are Google Shopping searches, wherever ChatGPT corrects for these results.
2. Brands successful ecommerce should attraction connected optimizing for Google Shopping first and foremost. Products that rank highly present are very apt to beryllium included successful immoderate ChatGPT Shopping recommendations.

How We Ran Our ChatGPT Shopping Experiment

We group retired to heavy dive into the fan-out queries that style the last answers, testing first whether ChatGPT is creating these shopping queries, and past whether the resulting products matched. 

Therefore the first measurement was to peek down the curtain astatine the fan-out queries taking place. 

Step 1: Asking ChatGPT for Product Recommendations 

We logged successful to ChatGPT and simply asked for merchandise suggestions pinch immoderate defined criteria. The purpose present is to prosecute pinch the level arsenic a personification would and springiness immoderate guidelines for the resulting products. 

In this example, we used: “best fund Android phones pinch awesome cameras”.

Step 2: Finding the Fan-Out Queries

You tin usage a solution for illustration Semrush Enterprise AIO to automatically place these hidden inheritance searches, but it’s besides imaginable utilizing Chrome Dev Tools. Here’s what to do: 

1. Open Chrome Dev Tools 
2. On the Network tab > Fetch/XHR, select utilizing the speech URL that ChatGPT creates (the last portion of the URL) starting pinch a number
3. Hit the refresh fastener to reload the speech and let the results to beryllium captured
4. Use CMD+ F to hunt the dev devices sheet for “search_model_queries” 
5. Here you’ll find the query instrumentality outs nether “queries”

In this lawsuit we have: "best fund Android phones pinch awesome cameras 2025" and "what defines a fund Android telephone and which fund phones person bully cameras 2024 2025".

Step 3: Finding the Shopping Fan-Out Queries

Shopping fan-out queries are hidden pinch an further furniture of encoding.

This is simply a spot trickier, but they tin beryllium recovered utilizing these steps:

1. In the aforesaid record arsenic before, hunt for “id_to_token_map”.
2. Find the portion of matter beside this that starts pinch “ey”. This is simply a portion of Base64 information that we’ll request to decode.
3. Copy the complete snippet (in this illustration it was astir 500 characters) without the opening and ending quote marks.
4. Paste the information into a free instrumentality for illustration Base64 Decode to make it readable and uncover the shopping fan-out queries.

Here’s the result:

The astir important constituent is aft “query” wherever we spot “cheap+android+smartphones+good+camera+2025”. This is the shopping fan-out query. Typically, you’ll spot only 1-2 unsocial queries here.

Step 4: Comparing the Results

Now we person each the accusation needed to comparison results. Just participate the shopping fan-out query into Google Shopping. Remember to cheque that the locale-location settings lucifer successful ChatGPT and Google Shopping for accuracy. 

Here’s what we saw for the Android telephone example:

Google Shopping

ChatGPT

The first 2 entries are recovered successful some Google Shopping and ChatGPT. In truth the retailer, title, and value accusation lucifer exactly. 

The Results: Shopping Query Fan-Out Is ChatGPT’s Additional Search Layer

After moving the research 100 times, we recovered that the apical ChatGPT merchandise was included successful Google Shopping’s first 3 results 75% of the time. There was besides important overlap pinch the 2nd and 3rd results.

Why Does ChatGPT Use Google Shopping Results? 

ChatGPT uses Google Shopping Results because they’re specified a rich | root of data. This isn’t conscionable a action of products, but a room that includes reviews and unrecorded pricing info. Products tin beryllium confidently recommended pinch the correct prices and retailers. 

While ChatGPT is making efforts to move distant from Google and Google Shopping (shown by the caller Etsy integration, arsenic good arsenic its caller Shopping Research features), it can’t yet lucifer this personification experience. Google’s ecosystem is conscionable excessively vast.

Accurate unrecorded pricing info is crucial, particularly erstwhile prices tin up and down astatine a moment’s announcement during ecommerce events for illustration Black Friday.

What Does This Mean for Ecommerce Brands?

It intends that ecommerce brands request to see their shopping feeds and guarantee that immoderate merchandise pages aliases listings that provender these are ever up to date. Google Shopping is the astir important of these, undoubtedly ChatGPT and different platforms are moving to create their own.

For merchandise queries successful ChatGPT, the logic seems to beryllium separated into: 1. Retrieving buyer-guide discourse and 2. Retrieving the products from Google Shopping.

Therefore correct now, optimizing your Google Shopping results is 1 of the astir important factors for appearing successful the ChatGPT merchandise carousel. 

Additionally, brands request to hole for the future. In-chat purchasing and agentic e-commerce are only going to proceed to grow. Developments for illustration Google’s Universal Commerce Protocol (UCP) will streamline these processes. 

We’re approaching a world wherever transactions hap straight wrong the chat, without ever visiting your website.

About the Experiment

This research was based connected the fan-out queries of 100 prompts. Each punctual was tally 5 times, pinch the astir communal apical products successful the carousel recorded to correct for the probabilistic quality of LLMs. 

While this is simply a mini sample size, it gives a clear starting constituent for AI shopping optimization. We’ll proceed to analyse shopping query fan-out pinch larger datasets to supply further insights.

You tin tally this trial for yourself. While we utilized a logged successful account, you whitethorn spot different results depending connected whether you’re logged in, person a free account, aliases person a paid account. This was seen successful similar ChatGPT and Google experiments we’ve conducted.