© 2026 ShoppingPartnerLab · All rights reserved
© 2026 ShoppingPartnerLab · All rights reserved
Google's Universal Cart is live. AI agents now make purchases autonomously. We scanned 6,551 European webshops across 7 categories and 40+ signals. The result: virtually no one is ready.
On May 19, 2026, Google introduced Universal Cart at I/O — an AI-powered shopping cart that works across Search, Gemini, YouTube and Gmail. Via the Agent Payments Protocol (AP2), AI agents can now make purchases autonomously on behalf of users. Nike, Sephora, Target, Walmart and Wayfair are already onboard. McKinsey estimates the market at $3–5 trillion by 2030.
The core: the Universal Commerce Protocol (UCP) is the new standard through which AI agents find, compare and purchase products. Shops that don't provide machine-readable product data and structured data are simply not included in the Universal Cart.
The impact on clicks is already measurable. Gartner predicted 25% less search volume by end of 2026 — that prediction is running ahead of schedule. 80%+ of all searches end without a click. In Google's AI Mode: 93%. But AI-referred traffic converts 42% better (Adobe, Q1 2026). Those who invest now build a structural advantage.
To understand why the data in this report is so alarming, you need to understand what concretely happens when a consumer says: "Find me a good running shoe under €120."
In the old world, that consumer opened Google, clicked 5 links, compared manually, and ordered. In the new world, an AI agent takes over the entire process. Here's how it works technically:
The AI parses the question: product type (running shoe), price limit (€120), implicit requirements (quality, availability, trustworthy shop).
The agent crawls product feeds, Google Shopping Graph (60 billion listings), and webshops. It looks for Product schema with price, brand, availability and images.
Shops without structured data, with JS shell, or without machine-readable prices are immediately skipped. They don't even enter the comparison. This is where 44% of European shops drop out.
The agent evaluates: does the shop have reviews (AggregateRating)? Contact details? Return policy? SSL? Is the business verifiable? This is where it consults trust registries like the Trust Registry.
From the remaining shops, the agent creates a ranking based on price, reviews, delivery time, and trust score. The consumer sees the top 2–3 options — without ever visiting a website.
Via the Universal Commerce Protocol (UCP) and Agent Payments Protocol (AP2), the agent completes the purchase — checkout, payment, confirmation — without the consumer ever seeing the webshop.
The crucial insight: in this process there is no click, no website visit, no Google ranking that matters. The only question is: can the AI agent read your product data (step 2), do you pass the filter (step 3), and are you verifiably trustworthy (step 4)? That is exactly what our scanner measures — and where 99.7% of European webshops fall short.
There is a technical reality that few people outside the AI industry understand, but that explains everything about why structured data is so decisive: LLMs are fundamentally lazy — and that's by design.
Every Large Language Model operates on tokens — chunks of text that the model processes. Each token costs compute, and therefore money. GPT-4o costs roughly $2.50 per million input tokens. Claude 3.5 Sonnet sits at similar rates. When an AI shopping agent needs to understand a product page, the number of tokens it requires directly determines the cost and speed of the transaction.
Here's where it gets technically interesting:
The difference is a factor of 100×. According to HTTP Archive data, a median web page contains ~9,000 tokens of HTML after preprocessing. The same product information sits in ~80 tokens of JSON-LD. That's 112× cheaper, 112× faster, and — crucially — with 0% interpretation error.
The reality is even more radical. Research by ProxyEmpire (2026) shows that production systems detect JSON-LD and bypass the LLM model entirely: "Structured data embedded in pages — such as JSON-LD — can often be extracted with near-zero cost and perfect accuracy." The LLM is only invoked when no structured data exists. Your competitor with JSON-LD costs the system literally zero tokens. You without JSON-LD cost 9,000.
This has three direct consequences for how AI shopping agents will restructure the market:
1. Agents optimize for cost. Every AI provider — Google, OpenAI, Anthropic, Perplexity — optimizes their agents for token efficiency. An agent that needs 100× more tokens to extract the same information from a shop is 100× more expensive to run. Agentic commerce operates on margins of fractions of cents per transaction. Shops serving unstructured data won't be "ranked lower" — they won't be processed at all because it's computationally unprofitable.
2. Structured data eliminates hallucination risk. When an LLM must extract a price from unstructured HTML, there's always a risk of misinterpretation — the model "hallucinates" €119.99 as €11,999 or grabs the wrong <span>. With JSON-LD, there is no interpretation: the field "price": 119.99 is unambiguous. For an agent spending money on behalf of a consumer, that risk differential is unacceptable. Agents will always choose the source with the lowest hallucination risk.
3. Eligibility-based dynamics. Shops with good data structures are eligible to be processed by agents at all → those eligible shops then have the opportunity to transact, accumulate reviews and authority → which compounds their selection probability over time. Shops without machine-readable data are not "ranked lower" — they're structurally skipped, regardless of how good the underlying offer is. Readability is the entry requirement; recommendation depends on additional factors such as reputation, reviews and authority.
Clickvoyant (2026) summarizes it: "If your competitor exposes 20 structured attributes and you expose 5, the agent doesn't think harder. It just picks them." AI agents aren't neutral — they're efficient. And efficiency means: the path of least resistance.
This is not a theoretical scenario. Google's Shopping Graph already indexes 60 billion product listings, primarily via structured data feeds. OpenAI's shopping integration filters on schema availability. Shopify's Universal Commerce Protocol requires standardized data. And even after the agent recommendation, data quality is decisive: Yext research (2026) shows that 62% of consumers verify AI recommendations by searching themselves, and 58% visit the website directly. If what they find doesn't match what the agent promised — wrong price, outdated stock, inconsistent reviews — you lose a customer you technically already won. The market is converging on one truth: only shops that speak the language of machines will survive.
Consumers no longer ask a search engine for a list of links — they ask an AI agent to solve a problem. That agent consumes data structures: structured data, schema markup, machine-readable content. If your shop doesn't deliver those, you don't exist in the decision-making process.
For context: Shopify's own Commerce Readiness Tool measures an average of 42/100 across 1,000 Shopify stores. Digital Applied audited 5,000 sites and found only 22% fully pass the Rich Results Test. A Pattern report states 75% of e-commerce leaders admit they're not ready. Our research confirms and deepens this picture for the entire European market, cross-platform.
The ShoppingPartnerLab Feed Scanner analyzes each webshop across 7 categories with over 40 individual signals. Unique: for each shop, we run three independent AI models — Google Gemini, Anthropic Claude and Perplexity — to write an independent assessment. This goes beyond Shopify's Agentic Commerce Audit (31 checks, Shopify only) and broader than tools like Semrush or Ahrefs. It works on Shopify, WooCommerce, WordPress, Lightspeed, Magento and custom platforms.
Scope and limitations. This report measures what AI agents actually use, not what matters for human browsing experience. PageSpeed and Core Web Vitals (LCP, FCP, CLS) are deliberately excluded — AI agents don't open browsers and don't experience visual load times. The scanner analyzes homepage + product page sample, which mirrors how AI shopping agents evaluate shops: they don't crawl your entire site, they land on your homepage, look for structured data, and assess a handful of product pages. The dataset contains 98% shops from the discovery engine and 2% from manual scans. This means shops that are completely invisible online are not in our sample — the actual AI-readiness of European e-commerce is likely even lower than our data suggests.
2,191 out of 6,551 shops serve an empty JavaScript container. No text, no product information, no structured data. AI crawlers from Google, OpenAI and Perplexity don't execute JavaScript — these shops don't exist.
This is how an AI model describes a typical JS shell shop: "Zero content, missing metadata, and no schema make the site invisible to AI crawlers and recommendation engines."
Schema is the strongest predictor. Shops with schema score 53.8 on average, without 35.8 — a difference of 18 points. Product schema is used by only 278 shops.
Unique: each shop is evaluated by Gemini, Claude and Perplexity. Three independent AI models, three scores, three assessments. The average AI visibility score is 24.
Surprisingly: Next.js scores highest (48.3) with only 2% JS shell. The framework isn't the problem — the configuration is. The JS shell problem concentrates in the 4,236 shops without a detected framework (51% JS shell).
| 1bikeinn.com | 82 |
| 2tradeinn.com | 78 |
| 3canyon.com | 67 |
| 1apple.com | 77 |
| 2oneplus.com | 72 |
| 3ldlc.com | 72 |
| 1na-kd.com | 73 |
| 2hessnatur.com | 66 |
| 3goertz.de | 66 |
| 1vitaminstore.nl | 76 |
| 2farmaline.be | 68 |
| 3myprotein.com | 66 |
| 1hornbach.nl | 67 |
| 2hornbach.at | 65 |
| 3jysk.dk | 64 |
| 1rossmann.pl | 70 |
| 2cultbeauty.co.uk | 67 |
| 3parfumado.com | 66 |
| Feature | Top 50 | Bottom 50 |
|---|---|---|
| Schema.org | 100% | 0% |
| Canonical tag | 94% | 0% |
| Open Graph tags | 90% | 0% |
| Sitemap | 94% | 0% |
| Hreflang | 60% | 0% |
| llms.txt | 16% | 0% |
| JS Shell | 0% | 96% |
To understand why the Trust Registry is needed, you need to understand how LLM models think about trust — and where they fall short.
A Large Language Model can read text, interpret structured data and recognize patterns. But it cannot independently verify whether a webshop is trustworthy. It cannot check if a business actually exists, if the registration number is correct, if the price in the schema matches the real price, or if the return policy is fair. An LLM reads what a shop says about itself — but it cannot assess whether that's true.
This is the same problem the internet had before SSL certificates. Any website could claim to be secure, but your browser had no way to verify that. The solution was a certificate authority (CA) — an independent third party that verifies a website is who it claims to be. Your browser trusts the CA, the CA verifies the website, and thus your browser trusts the website.
Has structured data, product schema, reviews, business details — but these are unverified claims
Independently verifies: 40+ checks, 3 AI models, continuous monitoring. Provides a verified trust score
Consults the registry as external verification source. Can now rely on substantiated trust instead of guessing
Without this intermediary layer, an LLM model must decide on its own whether a shop is trustworthy — based on the same unverified data the shop itself provides. That's like asking a job applicant if they're trustworthy: the answer is always yes. The Trust Registry solves this by providing an independent, machine-readable verification layer that LLMs can consult.
This is not theoretical. Google's Universal Commerce Protocol already requires standardized product data and trust signals. OpenAI's Instant Checkout relies on partner verification. As AI agents gain more autonomy to actually spend money on behalf of consumers — via AP2, via ACP — the question "can I trust this shop?" becomes the most critical decision point in the entire transaction chain.
The analogy summarized: SSL certificates tell your browser "this connection is secure." The Trust Registry tells an AI agent "this shop is verified trustworthy for autonomous transactions." In a world where 80%+ of searches end without a click and AI agents take over purchasing, this is the difference between existing and not existing in the digital economy.
Check if SSR is enabled. The difference is 14.5 vs 46.0 points. This is the single biggest fix.
Minimum: Product, Organization, BreadcrumbList. Optimal: FAQPage, AggregateRating, OnlineStore. 68% are missing this entirely.
Add llms.txt, allow AI crawlers in robots.txt, implement complete OG tags, fix your H1 and meta descriptions.
Achieve a score of 80+ and be included in the Trust Registry. In a world where AI agents buy autonomously, verified trustworthiness is the new currency.
Want to know how your webshop scores? The ShoppingPartnerLab Feed Scanner checks your shop across 7 categories and 40+ signals — including the assessment from Gemini, Claude and Perplexity. Within 60 seconds you'll see your results and whether you qualify for the Trust Registry.
Full disclosure: this case study concerns a European wellness webshop on Shopify that used our scanner to systematically improve AI-readiness. We deliberately show the areas where the shop still falls short.
On May 1, 2026, the shop ran its first scan with the Feed Scanner. The site was a Vue-based SPA on Shopify — technically solid, but virtually unreadable for AI agents. The AI verdict was devastating:
28 days later, after systematically implementing every step in the roadmap, a new scan was run:
34 out of 50 checks changed. The key fixes:
From 2 schema types (Organization, WebSite) to 14+: Product with Offer, aggregateRating, Review, MerchantReturnPolicy, OfferShippingDetails, FAQPage, HowTo, VideoObject, BreadcrumbList, DigitalDocument. GTIN, MPN, and all Google Merchant Listings fields populated.
From a Vue SPA serving an empty HTML shell, to static prerender snapshots with embedded JSON-LD. Sitemap grew from a manual list to a sitemap index of ~70,000 URLs.
7 llms.txt files (~5,000 lines) with product facts, dosage guidelines, regional regulations, symptom matrices. Plus MCP discovery, ChatGPT plugin manifest, and OpenAPI spec.
5 GitHub Actions + 12 validation scripts: SEO validation, JSON-LD checks, broken links, product schema, and prerender gates. Regressions are caught before merge.
After the fixes, three AI models re-evaluated the shop:
What's still not good enough. Operational Maturity scores 64/100 — shipping policy and return window are incomplete. The Perplexity scan didn't detect llms.txt (the scan ran before the llms.txt deploy). And Google Search Console still shows 12 items with invalid price format and missing images in Merchant Listings. We're not done — but the architecture is in place.
The lesson: a shop can go from 35 to 86 in 28 days. It doesn't require a new platform, a redesign, or months of work. It requires systematically implementing structured data, solving the JS shell problem, and building a machine-readable content layer. The roadmap in this report isn't theoretical — it's exactly what we did ourselves.
ShoppingPartnerLab · State of AI-Readiness in European E-commerce · Q2 2026
6,551 webshops · 24 countries · 16 sectors · 7 categorieën · 40+ checks · 3 AI-modellen
Sources: Google I/O 2026, Adobe Digital Economy Index Q1 2026, Gartner Search Forecast, Shopify Commerce Readiness Tool, Digital Applied Schema Audit 2026, Pattern AI Report 2026, BCG Consumer AI Trust Survey
© 2026 ShoppingPartnerLab B.V. · KVK 88375692 · Maastricht (Nederland)
This report and its contents are proprietary. Reproduction, distribution, or citation without attribution is prohibited.
Contact: info@shoppingpartnerlab.com · shoppingpartnerlab.com