How Generative AI is Changing Product Discovery

The Numbers That Changed the Conversation

Let me put the Adobe Analytics data in context, because percentages can be misleading when the base is small.

In July 2024, generative AI traffic to retail sites was tiny — barely a rounding error in most analytics dashboards. But by July 2025, it had grown 4,700% year-over-year. On Prime Day 2025 (July 8-11), AI-driven traffic surged 3,300% year-over-year. During the holiday season, it jumped 693%.

Adobe's caveat is important: AI-driven traffic still represented a small share of total ecommerce visits. Paid search accounted for 28.5% of ecommerce sales during Prime Day. Email, direct traffic, and social still dominate. But the trajectory is unmistakable, and the quality metrics are what should grab every ecommerce manager's attention.

During the 2025 holiday season, shoppers arriving from AI assistants converted 31% more than visitors from other traffic sources. That conversion advantage doubled compared to the 2024 holiday season. AI-referred visitors had 33% lower bounce rates, spent 45% more time on site, and viewed 13% more pages per visit. Revenue-per-visit from AI sources grew 84% between January and July 2025.

And the conversion gap is closing fast. In July 2024, AI traffic was 43% less likely to convert than non-AI traffic. By February 2025, that gap had closed to 9%. By October 2025, AI-referred visitors were actually 16% more likely to convert.

This isn't a niche experiment. This is a new high-intent discovery channel.

How AI Shopping Actually Works

To understand what this means for your product data, you need to understand how people actually shop through AI assistants.

A traditional Google search: "womens black ankle boots" A ChatGPT shopping query: "I need comfortable black ankle boots for a teacher who's on her feet all day, waterproof would be great, budget around £80, and I'd prefer a brand that's sustainable"

The first query matches keywords. The second describes a person, a context, constraints, and values. The AI needs entirely different data to answer it well.

ChatGPT's shopping research feature (launched late 2025) turns this into a guided experience. It asks clarifying questions: "What style — heeled or flat? Do you prefer leather or vegan alternatives?" It pulls price, spec, and review data from the open web, generates a tailored buyer's guide, and refines based on feedback ("More like this" or "Not interested").

Perplexity takes a different approach. It blends real-time web search with AI synthesis, delivering product cards with source citations and — through its PayPal integration — enabling in-flow purchases. The merchant remains the merchant of record, handling returns and the customer relationship.

Google's Gemini benefits from direct access to Google Shopping data, product feeds, merchant information, and the entire Google ecosystem of reviews and pricing.

Each platform has its own quirks, but they all share one requirement: they need structured, rich, machine-readable product data to recommend your products effectively.

What This Means for a Retailer Like New Look

New Look is a high-street fashion retailer with thousands of SKUs across womenswear, menswear, and teens. Their typical customer is price-conscious, trend-aware, and increasingly shopping on mobile. Let's think about what happens when that customer starts asking AI for recommendations instead of browsing the New Look website directly.

The AI shopping query: "What should I wear to a casual Friday office in London, I like the oversized blazer trend, budget under £50, and I need it by Thursday"

For ChatGPT or Perplexity to recommend a New Look oversized blazer, they need:

The product exists in a feed the AI can access (Google Shopping feed, open web crawlable product pages, or a direct integration like Shopify's Agentic Storefronts). The product description mentions "oversized blazer" as a style descriptor, not just "blazer." The product data includes occasion information: "casual Friday", "office-to-evening", "smart casual." The price is clearly structured and current. Delivery information is machine-readable ("next-day delivery available" or "estimated delivery 2-3 days"). Reviews mention comfort, fit, and styling context.

Most traditional product feeds don't include this information. They have: product name, price, category, colour, size. That's enough for Google Shopping ads. It's not enough for a conversational AI that needs to match a complex, contextual query.

Here's what a GEO-optimised product listing looks like versus a traditional one:

Traditional product feed entry: Name: Oversized Single-Breasted Blazer Price: £39.99 Category: Womens > Jackets > Blazers Colour: Black Sizes: 6-22

GEO-optimised product feed entry: Name: Oversized Single-Breasted Blazer Price: £39.99 Category: Womens > Jackets > Blazers Colour: Black Sizes: 6-22 Style: Oversized, relaxed fit, dropped shoulders Occasions: Casual Friday, smart casual, office-to-evening, weekend brunch Styling notes: Layer over a T-shirt or rollneck. Pairs with wide-leg trousers for a relaxed office look or jeans for the weekend. Material: Recycled polyester blend, lightweight, unlined Fit context: Runs true to size in the body, deliberately oversized in the shoulders. Size down for a more structured fit. Sustainability: Made from 65% recycled materials. Part of the Kind to Earth collection. Delivery: Next-day delivery available. Order by 10pm for delivery tomorrow. Review summary: 4.2 stars, 847 reviews. Customers highlight comfort and versatility. Common feedback: "perfect for work", "great with jeans", "runs slightly large".

That second listing gives an AI assistant everything it needs to match the "casual Friday oversized blazer under £50 by Thursday" query. The first one doesn't.

What This Means for Selfridges

Selfridges operates in a completely different segment — luxury, experiential, editorial. Their customer is asking different questions:

"I want to buy my wife a special birthday gift, she loves understated luxury, thinks Bottega Veneta is too obvious now, what's the brand everyone in fashion is talking about but hasn't gone mainstream yet?"

This is a fundamentally harder query for AI. It requires understanding of fashion positioning, cultural cachet, and the subjective concept of "understated luxury." But Selfridges actually has the content to answer it — through their editorial features, buyer notes, and brand curation stories.

The problem: most of this content lives in editorial pages that aren't structured for machine consumption. The brand story about a emerging designer is in a blog post. The buyer's notes about why they selected a particular collection are in a newsletter. The styling context is in a lookbook.

For Selfridges, GEO optimisation means making editorial intelligence machine-readable:

Brand narrative structured data: Designer background, positioning ("emerging luxury", "quiet luxury", "avant-garde"), price tier, sustainability credentials. Editorial associations: "Featured in Selfridges Bright New Things", "Selected by our accessories buyer for AW26", "Featured in Vogue's Next Wave". Gifting context: Occasion tags, relationship tags ("for her", "for a fashion-forward friend"), price-point positioning within luxury ("entry-level luxury", "investment piece"). Craftsmanship details: Materials sourcing, production methods, artisan details.

The AI then has the context to say: "Based on current fashion conversations, you might consider [Brand X] — they're known for understated craftsmanship, have been featured in Selfridges' emerging designer programme, and offer pieces in the £200-500 range that feel special without the logo fatigue."

The Technical Implementation: Structured Data for AI

Whether you're New Look or Selfridges, the technical foundation is the same: your product data needs to be rich, structured, and machine-accessible.

module ProductFeed
  class GEOOptimiser
    def generate_ai_ready_listing(product)
      base = standard_feed_attributes(product)

      # Add conversational context
      base.merge(
        style_descriptors: extract_style_descriptors(product),
        occasion_tags: generate_occasion_tags(product),
        styling_context: generate_styling_notes(product),
        fit_context: extract_fit_information(product),
        sustainability: extract_sustainability_data(product),
        review_synthesis: synthesise_reviews(product),
        delivery_context: delivery_information(product),
        gifting_tags: generate_gifting_context(product),
        trending_signals: trending_data(product),
        brand_narrative: brand_context(product)
      )
    end

    private

    def generate_occasion_tags(product)
      # Derive occasions from category, style, and review analysis
      tags = []
      tags += category_occasion_map(product.category)
      tags += review_occasion_extraction(product.reviews)
      tags += editorial_occasion_tags(product)
      tags.uniq
    end

    def synthesise_reviews(product)
      return nil if product.reviews.count < 5

      {
        average_rating: product.reviews.average(:rating).round(1),
        review_count: product.reviews.count,
        common_praise: extract_common_themes(product.reviews.positive),
        common_concerns: extract_common_themes(product.reviews.negative),
        fit_feedback: extract_fit_sentiment(product.reviews),
        recommended_for: extract_recommended_contexts(product.reviews)
      }
    end
  end
end

The schema markup matters too. JSON-LD Product schema with extended attributes:

module SEO
  class ProductSchemaGenerator
    def generate(product)
      listing = ProductFeed::GEOOptimiser.new.generate_ai_ready_listing(product)

      {
        '@context': 'https://schema.org',
        '@type': 'Product',
        name: product.name,
        description: product.description,
        brand: { '@type': 'Brand', name: product.brand_name },
        offers: {
          '@type': 'Offer',
          price: product.price.to_s,
          priceCurrency: product.currency,
          availability: product.in_stock? ? 'InStock' : 'OutOfStock',
          deliveryLeadTime: {
            '@type': 'QuantitativeValue',
            value: product.delivery_days,
            unitCode: 'DAY'
          }
        },
        aggregateRating: listing[:review_synthesis] ? {
          '@type': 'AggregateRating',
          ratingValue: listing[:review_synthesis][:average_rating],
          reviewCount: listing[:review_synthesis][:review_count]
        } : nil,
        additionalProperty: [
          { '@type': 'PropertyValue', name: 'occasions',
            value: listing[:occasion_tags].join(', ') },
          { '@type': 'PropertyValue', name: 'style',
            value: listing[:style_descriptors].join(', ') },
          { '@type': 'PropertyValue', name: 'sustainability',
            value: listing.dig(:sustainability, :summary) }
        ].compact
      }
    end
  end
end

The GEO Checklist for Ecommerce

1. Enrich product descriptions beyond keywords. Include context: who is this product for, when would you wear/use it, what does it pair with, what problem does it solve? Write as if you're explaining the product to a knowledgeable personal shopper.

2. Structure your data for machines. JSON-LD Product schema on every product page. Price, availability, delivery, reviews — all machine-readable. Extended attributes for style, occasion, fit, sustainability.

3. Make reviews work harder. AI assistants synthesise review themes. If your reviews mention "perfect for petites" or "runs large", that becomes recommendation context. Encourage detailed reviews and make them accessible.

4. Keep product data fresh. AI models prioritise recent content — the average age of URLs cited by AI assistants is 1,064 days, which is 25.7% fresher than organic SERP results. Update product pages regularly. Seasonal refreshes. New styling suggestions. Current availability.

5. Build for conversational queries. Create FAQ content that mirrors how people ask AI: "What boots are good for standing all day?" not "comfortable women's boots." Page titles and headers should reflect natural-language questions.

6. Optimise for multiple AI platforms. ChatGPT uses web crawling and partnerships. Perplexity uses real-time search. Gemini uses Google Shopping feeds. Your product data needs to be accessible through all three channels. Google Merchant Centre feeds, crawlable product pages, and open structured data.

7. Track AI referral traffic. Set up UTM parameters or referrer analysis to identify traffic from ChatGPT, Perplexity, Claude, and Gemini. Measure conversion rates separately. You'll likely see higher intent.

8. Think beyond your own site. AI assistants synthesise information from across the web. Your product mentions in editorial features, comparison articles, and review sites all feed into AI recommendations. PR and editorial strategy now directly influence AI product discovery.

What Auto-Prammer.at is Doing

For our automotive marketplace, AI product discovery is already happening. Someone asks ChatGPT: "I need a family car for Austrian mountain roads, budget around €20,000, used is fine, ideally with good fuel economy and space for two car seats."

Our vehicle listings need to answer that query. That means structured data including: terrain suitability (mountain, motorway, city), family features (ISOFIX points, boot capacity, rear seat space), running cost estimates (fuel consumption, insurance group, service costs), and real owner feedback.

We're enriching every listing with these contextual attributes so when AI assistants search for "family car for mountains", Auto-Prammer.at listings surface with the right context.

The Bigger Picture

This shift connects directly to everything else we've built. The privacy-first data architecture means the data we use for GEO optimisation is consented and compliant. The DACH compliance framework ensures our product feeds meet Austrian, German, and Swiss requirements. The EU AI Act transparency provisions will eventually apply to how AI shopping assistants make recommendations — and the retailers whose data is most transparent and structured will benefit.

The old model: optimise for Google's algorithm, pay for ads, compete on keywords.

The new model: optimise for AI assistants that are becoming personal shoppers for millions of people. Give them the rich, structured, contextual product data they need to recommend your products.

The retailers who adapt their product feeds now — whether you're New Look making fashion accessible or Selfridges curating luxury — will be the ones AI assistants recommend first. And those recommendations convert 31% better than traditional search.

That's not a trend to watch. That's a channel to build for.