top of page

What an AI shopping agent reads on your product page, and what it ignores

Writer: Daniel Pisano
Daniel Pisano
12 minutes ago
6 min read

By Daniel Pisano, Partner & CGO at Vurdere

An AI shopping agent reads what arrives in the page's initial HTML, its structured data and the product feed: name, price, stock, attributes, average rating and review count. It ignores anything that only appears after JavaScript runs, text baked into images and content hidden behind clicks. A beautiful page with no readable data drops out of the answer.


Why is this suddenly urgent for European retailers?

I run a simple test with retailers: I open one of their product pages with JavaScript switched off. Quite often the 4.8-star rating that looks great in the browser disappears. Sometimes the price goes too. Nothing changed for the human shopper. For an AI agent reading the page, that product has no reviews at all.

The data side of retail is not ready for this reader. An audit by the agency Summit of 200 UK retailers in beauty, fashion, home and garden and pet care, published in October 2026, found that 41% lack structured data and 20% lack descriptive image alt text. Structured data was the most common issue in every category.

Feeds look no better. Channable ran an AI readiness audit across 9,689 companies on its platform: 77% scored below 60 out of 100, and fewer than 1% reached the 80 points it treats as fully optimised for AI agents. Missing, incomplete or inconsistent key product information was the main problem.


How does an AI agent "see" a product page?

There are three ways in, and it helps to keep them apart.

  1. The crawler that reads the page. That covers OpenAI's GPTBot and OAI-SearchBot, Anthropic's ClaudeBot and PerplexityBot. Vercel analysed traffic on its own network and found that these crawlers fetch JavaScript files but do not execute them. In Vercel's words, "none of the major AI crawlers currently render JavaScript". The exceptions it named were Gemini, which uses Googlebot's infrastructure, and AppleBot.

  2. The agent that browses like a person. Some agents open the page in a real browser and see what the shopper sees, JavaScript included. Vercel points out that optimising for the crawler that does not run JavaScript covers both cases, while the reverse does not.

  3. The feed. Instead of being crawled, the store pushes its catalogue to the shopping surface, such as ChatGPT or Google Merchant Center. OpenAI's product specification for ChatGPT has dedicated fields for this.


I explained how these routes connect to the protocols in the protocols that decide whether AI buys from your store. This piece is more practical: which parts of the page actually reach the machine.


What does the agent read on a product page?

Information

Where it needs to be

Why the agent cares

Name, price and availability

Initial HTML, schema.org Product markup and the feed

The basic filter: does it exist, is it in stock, does it fit the budget?

Attributes (size, material, voltage, compatibility)

Page copy and feed

Lets the agent compare like for like without guessing

Average rating and review count

aggregateRating in schema.org; star_rating and review_count in OpenAI's feed

A machine-readable summary of social proof

Returns and shipping

Page plus the accepts_returns, return_deadline_in_days and shipping_price feed fields

Decisive when the shopper said "only if I can return it" or "delivery included"

Review text and answered questions

Initial HTML of the page

What the AI draws on when asked to summarise reviews


Two details in OpenAI's specification stand out. First, the rating field is the average on a 0 to 5 scale with two decimals, and it should come with a review count above zero: a rating without volume tells the agent nothing. Second, the specification states that individual reviews and Q&A lists are not part of the discovery feed. The number goes into the feed; the substance of your reviews only reaches the agent if it sits on the page as readable text.

That is why I don't treat reviews as decoration. They are data in two formats: the aggregate, which fits in a field, and the text, which lives on the page for when someone asks "what do customers say about the battery?". I covered the second format in Product Q&A as data for AI shopping agents.


What does the agent ignore?

  • Content assembled by JavaScript in the browser. If the reviews widget, the sale price or the size guide only appear once a script runs, a crawler that does not execute JavaScript reads the page without them. Vercel's agentic commerce readiness guide puts it simply: server rendering and static generation are readable; client-side-only rendering is not.

  • Anything behind a click. Tabs, accordions and "show more" buttons that load text on click can hide exactly the specifications the agent needs.

  • Text inside images. A "free delivery" banner or a size chart saved as a JPG is not information to a machine. The Summit audit's 20% figure on missing alt text shows how common this is.

  • Marketing adjectives. A Lengow guide to the ChatGPT feed notes that ChatGPT favours clear, comparable details over promotional language. "Revolutionary" does not help anyone choose between two kettles.

  • Data that contradicts itself. Vercel recommends that price matches across markup, landing page, checkout and feed. When each place says something different, the agent cannot tell which is right.


How can you test your page in 15 minutes?

  1. Open an important product page and use "View page source" in the browser (not the inspector). Search for the price, the rating and the review count. If they are not there, the crawler does not see them either.

  2. Run the same URL through Google's Rich Results Test or the schema.org validator and check for Product with offers and aggregateRating.

  3. Compare the price on the page with the price in the feed you send to Google Merchant Center or ChatGPT.

  4. Check robots.txt and your firewall or CDN rules. Vercel warns that firewall rules blocking AI bots can shut them out even when robots.txt allows them.

  5. Ask an AI assistant to summarise the reviews of your product, giving it the URL. If it says it found none, you know where to start.


Where does Vurdere fit in?

Our part is social proof. In the Vurdere app for Shopify, the aggregate rating and review count are written to Shopify's native reviews.rating and reviews.rating_count metafields, not just drawn on screen by a widget. I compared what each platform provides in Shopify, VTEX and agentic commerce.

The Social SEO module exists so that customer reviews, questions and interactions sit on the page as content search engines and AI systems can read. In A/B tests with current clients, Vurdere recorded +20% organic traffic. And because an agent needs volume before it trusts an average, collection matters too: in the same tests, Vurdere reached up to 10x more reviews collected.


My view

For years I heard that a product page has to "delight". It still does, for people. But it now has a second reader that cannot see the photo, does not care about the colour palette and never clicks anything. That reader wants the specifications, the price, the returns policy and what other customers said, in text.

The good news is that almost everything that helps the agent also helps the shopper and Google. I am not talking about a parallel page "for robots". I am talking about no longer hiding information in images, closed tabs and scripts.

What worries me is reviews. Plenty of retailers have paid to gather thousands of them and then serve them in a format machines cannot read. If I had to pick one task for December, it would be opening the source code of the ten best-selling product pages and checking that the rating is actually there.


Frequently asked questions


Does ChatGPT read reviews that load with JavaScript?

OpenAI's search crawler does not execute JavaScript, according to Vercel's analysis. If reviews only appear after a script runs, it does not see them when reading the page.


Is it enough to put the rating in the product feed?

It helps, but no. OpenAI's specification accepts an average rating and review count in the feed but leaves review text and Q&A out. That content has to be on the page.


Do I need an llms.txt file?

Vercel's guide ranks llms.txt below schema.org markup, Open Graph tags and a product feed. Start with those three.


What is aggregateRating?

It is the schema.org property that states a product's average rating and number of reviews in structured form, inside the page's code.

 
 
 

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page