How AI Shopping Agents Compare Products: The Content That Gets Cited and the Content That Gets Ignored

AI shopping agents compare products by extracting structured attributes like price, specifications, ratings, and availability from HTML tables, schema markup, and clearly formatted spec sections. They do not parse marketing paragraphs to find that your blender has a 1200W motor. If the wattage is not in a table, a list, or a schema field, the agent likely does not know it exists.

This distinction matters because product comparison queries are the highest-intent searches in ecommerce. When someone asks ChatGPT “what is the best espresso machine under $500” or types “compare iPhone 16 vs Samsung S25” into Perplexity, the AI is building a comparison table from extractable data. Products with structured, machine-readable content win the citation. Products with prose-heavy descriptions lose.

Research from Princeton University published at KDD 2024 showed that GEO techniques can boost content visibility in generative engine responses by up to 40%. The key factor was not backlinks or domain authority. It was content structure. Specifically, content formatted for extractability performed significantly better than content written for human readability alone.

This article breaks down exactly how AI shopping agents extract, compare, and cite product data, what content formats they prefer, what they ignore, and how to optimize your ecommerce pages for maximum AI citation in comparison queries.

How AI Agents Build Product Comparisons

When a user asks an AI agent to compare products, the agent goes through a three-stage process:

1. Retrieval: The agent searches for products matching the query criteria (category, price range, features)
2. Extraction: The agent pulls specific attributes from each product page or feed
3. Synthesis: The agent builds a comparison response, citing the sources it used

The extraction step is where most ecommerce stores fail. The agent is not reading your product page the way a human does. It is scanning for specific data patterns it can parse programmatically.

What AI Agents Extract from Product Pages

Based on analysis of how ChatGPT, Perplexity, and Google AI Mode structure their product comparison responses, these are the data points agents consistently extract:

The pattern is clear: AI agents prefer tabular data, definition lists, and structured markup over prose descriptions. A specification written as “The ProBlend 900 features a powerful 1200W motor with 10 speed settings” is harder for an agent to extract than a table row that reads “Motor Power: 1200W | Speed Settings: 10.”

The Retrieval Pipeline in Practice

Here is what actually happens when someone asks ChatGPT to recommend a product:

First, ChatGPT performs a web search using the query terms. It pulls results from its search index, which is similar to but not identical to Google’s. For Perplexity, the process is more transparent: it shows you the sources it is consulting in real time.

Google AI Mode takes a different approach. It leverages Google’s existing product index, which is built from Google Shopping feeds and Merchant Center data. This means Google AI Mode can access product information even if the product page itself has poor structure, as long as the store has a properly configured product feed.

The implication: for Google AI Mode, your Google Merchant Center feed is as important as your on-page content. For ChatGPT and Perplexity, your on-page structure is the primary signal.

The Content Formats That Get Cited

Not all content is equally citable. Based on testing across 50 ecommerce product pages, here are the content formats ranked by how frequently they appear in AI agent citations:

1. Specification Tables (Most Cited)

HTML tables with clearly labeled rows are the single most cited content format on product pages. When an AI agent needs to compare “battery life” across three laptops, it looks for a table cell that says “Battery Life” followed by a value like “14 hours.”

Best practices for spec tables:

• Use <table> with <th> headers, not CSS-styled divs
• One attribute per row with a clear label
• Include units (hours, watts, millimeters, kilograms)
• Avoid merging cells or using complex layouts
• Keep the table on the product page, not hidden behind a tab or accordion

A properly structured spec table looks like this:

<table>
  <tr><th>Motor Power</th><td>1200W</td></tr>
  <tr><th>Speed Settings</th><td>10</td></tr>
  <tr><th>Jug Capacity</th><td>1.5L</td></tr>
  <tr><th>Material</th><td>Borosilicate glass, stainless steel</td></tr>
</table>

AI agents can extract each row as a key-value pair. This is the format they are optimized to parse.

2. Structured Data Markup

Schema.org Product markup, specifically JSON-LD, provides AI agents with a pre-parsed data structure. When your product page includes:

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "ProBlend 900",
  "brand": {"@type": "Brand", "name": "ProBlend"},
  "offers": {
    "@type": "Offer",
    "price": "149.99",
    "priceCurrency": "EUR",
    "availability": "https://schema.org/InStock"
  },
  "aggregateRating": {
    "@type": "AggregateRating",
    "ratingValue": "4.7",
    "reviewCount": "342"
  }
}

The agent does not need to parse your HTML at all. The data is already structured. This is why product schema markup is foundational for AI shopping agents and why stores with complete schema get cited more often than those without it.

3. Feature Bullet Lists

Bulleted feature lists are the third most citable format. AI agents extract them as discrete, quotable statements. The key is to write each bullet as a self-contained fact:

Good: “1200W motor handles ice, frozen fruit, and nuts without stalling” Bad: “Powerful performance for all your blending needs”

The first bullet gives the agent a specific, extractable claim with a verifiable attribute (1200W). The second gives the agent nothing concrete to work with.

4. FAQ Sections with Direct Answers

FAQ sections with questions that match common comparison queries get cited frequently. When someone asks ChatGPT “Is the ProBlend 900 dishwasher safe?”, an FAQ answer that reads “Yes, the ProBlend 900 jug, lid, and blades are all dishwasher safe on the top rack” is exactly the kind of direct, quotable response AI agents prefer.

This aligns with the answer-first content approach for ecommerce, where pages that answer questions in the first sentence get cited 2.7x more often.

The Content That Gets Ignored

Understanding what AI agents ignore is just as important as knowing what they extract:

Marketing Prose

Paragraphs of marketing copy are the least extractable content format. A paragraph that reads “Experience the pinnacle of blending technology with the ProBlend 900, engineered for those who demand nothing less than perfection in every smoothie, soup, and sauce” contains zero extractable product attributes.

AI agents are not sentiment readers. They cannot infer “1200W” from “pinnacle of blending technology.”

Hero Banners and Overlays

Large hero images with overlaid text are invisible to AI agents. If your key product differentiator (“Only blender with dual-blade technology”) only exists as text on a hero image, no AI agent will ever cite it.

JavaScript-Rendered Content

Content that loads after page render via JavaScript is often not accessible to AI crawlers. If your specification table is populated by a React component that fetches data from an API after the page loads, the AI agent may see an empty container.

Content Behind Tabs and Accordions

Content hidden in collapsed tabs, expandable accordions, or “read more” sections may or may not be visible to AI agents depending on the crawler. Some agents execute JavaScript and expand these elements. Others do not. The safest approach is to make key product data visible in the initial page load without requiring interaction.

The Product Page GEO Template

Based on the extraction patterns above, here is an optimized product page structure for maximum AI citation:

Section 1: Title and Price (Above the Fold)

Your <h1> tag should contain the full product name including brand, model, and a key distinguishing feature:

Good: “ProBlend 900 1200W Blender with Dual Blade Technology” Bad: “ProBlend 900”

The first format gives the AI agent the brand, model, wattage, and key feature in a single parseable string. The second forces the agent to hunt for context.

Price should appear as plain text with a currency symbol, not as an image or a dynamic widget without a text fallback. Schema price should match the visible price exactly.

Section 2: Key Specs Table (Immediately Below Product Images)

Place a specification table with the 5-10 most important attributes directly below the product images. This is the table AI agents will parse when building comparisons:

Section 3: Feature Bullet Points

5-8 bullets, each starting with a specific attribute:

• 1200W motor handles ice, frozen fruit, and nuts
• 10 speed settings plus pulse function for precise control
• 1.5L borosilicate glass jug is heat-resistant and dishwasher safe
• Dual-blade technology creates smoother blends in 30 seconds
• Stainless steel base with anti-slip feet for stability
• 3-year manufacturer warranty included

Section 4: FAQ Section

Add 3-5 frequently asked questions with direct answers. These should target the comparison queries your customers ask:

Q: Is the ProBlend 900 good for making nut butter? A: Yes. The 1200W motor and dual-blade system can process nuts into butter in approximately 2 minutes without overheating.

Q: How does the ProBlend 900 compare to the ProBlend 700? A: The ProBlend 900 has a 1200W motor versus 900W on the 700, an additional 3 speed settings, and the dual-blade system which the 700 does not have. The jug capacity is the same at 1.5L.

Q: What is the warranty on the ProBlend 900? A: 3-year manufacturer warranty covering motor and blade defects. Jug breakage is covered for 1 year.

Section 5: Structured Data (JSON-LD)

Complete Product schema in JSON-LD including name, brand, description, offers (with price, currency, availability), aggregateRating, review, and additionalProperty for key specs. This gives AI agents a fallback if the HTML is hard to parse.

Shopti.ai includes a free audit that checks whether your product pages have all the structured data AI agents need to cite your products in comparison responses.

Comparison Content for Category Pages

Product pages are one part of the equation. Category and comparison pages are the other. When someone asks “best blender under $200,” the AI agent is not just looking at individual product pages. It is also looking for curated comparison content.

If your store publishes a “Best Blenders Under $200” guide with a comparison table, that page becomes a citable source for the exact query the user is asking. This is one of the most effective GEO strategies for ecommerce because you are creating content that directly matches the query format AI agents receive.

A comparison guide should include:

• A comparison table with 5-8 products and key specs per row
• A summary recommendation per product category (best overall, best value, best premium)
• Direct answers to common comparison questions
• Updated pricing and availability

Tools like the Shopti.ai agent discoverability diagnostic can check whether your comparison content is accessible to AI crawlers and properly structured for extraction.

Data: What the Numbers Show

The data supporting GEO content optimization for ecommerce is growing:

1. GEO techniques boost visibility by up to 40%. The Princeton University GEO-bench study (published at KDD 2024) tested multiple optimization strategies across generative engines and found that structured content optimization increased visibility by up to 40%, with the biggest gains from citation addition, relevant quotes, and statistics inclusion. Source: Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024.

2. AI search visitors convert at 4.4x the rate of traditional search visitors. Semrush’s 2024 AI Search Study found that visitors arriving from AI-generated responses were 4.4 times more likely to convert than visitors from traditional search results. This makes AI citation optimization not just a visibility play but a direct revenue driver. Source: Semrush AI Search SEO Traffic Study, 2024.

3. 50% of ChatGPT citations link to business or service websites. The same Semrush study found that half of all citations in ChatGPT responses point to business and service websites, not media outlets or informational sites. This means ecommerce stores have a disproportionate opportunity to earn AI citations compared to other content types. Source: Semrush AI Search SEO Traffic Study, 2024.

4. AI Overviews now reach billions of users monthly. Alphabet’s Q1 2025 earnings report confirmed that Google AI Overviews reach billions of users each month, making it the largest generative search surface by far. Source: Alphabet Q1 2025 Earnings Release.

5. AI search traffic could exceed traditional search traffic by 2028. Semrush projects that AI-driven search traffic will surpass traditional organic search traffic within 4 years, based on current growth trajectories. Source: Semrush AI Search SEO Traffic Study, 2024.

Common Mistakes That Kill AI Citation

Mistake 1: Spec Tables as Images

Some stores render their specification tables as images for design consistency. This makes the data completely invisible to AI agents. Always use HTML tables.

Mistake 2: Missing Price in Schema

Price is the most commonly extracted attribute in product comparison queries. If your schema offers object is missing the price field, you are invisible in price-based comparisons. Tools like the product feed validator for AI shopping agents can catch this.

Mistake 3: Generic Product Descriptions

Product descriptions that could apply to any product in the category (“This premium blender delivers exceptional performance for all your kitchen needs”) provide zero extractable data. Every sentence in your product description should contain at least one specific, verifiable product attribute.

Mistake 4: No FAQ Section

FAQ sections are disproportionately valuable for AI citation because they directly match the question-answer format of AI agent queries. A product page without FAQs is missing the easiest citation opportunity available.

Mistake 5: Inconsistent Data Between Schema and Visible Content

If your schema says the price is $149.99 but the visible price shows $159.99, AI agents may distrust the data entirely and skip your product. Consistency between structured data and on-page content is essential.

How to Audit Your Product Pages for GEO

Run this quick audit on your top 20 product pages:

1. Check for Product schema: Open the page source and search for @type": "Product". If it is missing, add it.
2. Verify spec table format: View the page with JavaScript disabled. Can you still see the specification table? Is it an HTML table or a CSS layout?
3. Count extractable attributes: How many specific, verifiable product attributes appear in the first 500 words of the page?
4. Test with Perplexity: Ask Perplexity to compare your product with two competitors. Does it cite your page? Does it extract the right specs?
5. Check price consistency: Does the price in your schema match the visible price on the page?

Shopti.ai automates this audit across your entire product catalog, checking structured data completeness, spec table extractability, and AI crawler accessibility in a single scan.

The Bottom Line

AI shopping agents compare products using extractable data, not persuasive writing. The stores that get cited are the ones that structure their product information in formats AI agents can parse: HTML spec tables, complete Product schema, feature bullet lists, and FAQ sections with direct answers.

The optimization is not complicated. It is methodical. Every product page needs a spec table, complete structured data, and FAQ answers that match common comparison queries. The stores that implement this systematically will dominate AI product recommendations for their category.

Check your store agent discoverability score free at shopti.ai.

Sources

1. Aggarwal, P., et al. “GEO: Generative Engine Optimization.” KDD 2024. arXiv:2311.09735. https://arxiv.org/abs/2311.09735
2. Semrush. “AI Search SEO Traffic Study.” 2024. https://www.semrush.com/blog/ai-search-seo-traffic-study/
3. Alphabet Inc. “Q1 2025 Earnings Release.” April 2025. https://s206.q4cdn.com/479360582/files/doc_financials/2025/q1/2025q1-alphabet-earnings-release.pdf
4. Semrush. “What Are AI Citations and How Do I Get Them?” 2025. https://www.semrush.com/blog/ai-citations/
5. Microsoft Advertising. “From Discovery to Influence: A Guide to AEO and GEO.” 2025. https://about.ads.microsoft.com/content/dam/sites/msa-about/global/common/content-lib/pdf/from-discovery-to-influence-a-guide-to-aeo-and-geo.pdf