Agent vs. the Web

E-commerce

Can AI Agents buy a laptop online?

We Tested AI Agents on the World's Biggest E-Commerce Sites. Here are the results.

FlowTester Research · February 2026

There's a multi-trillion dollar bet being placed on AI agents that can shop for you. McKinsey projects AI agents could mediate $3 trillion to $5 trillion of global consumer commerce by 2030. Google, Anthropic, OpenAI, are all racing to build the agent that becomes your default checkout button.

There's just one inconvenient problem: agents are struggling to use most websites.

We know this because we tested it. We gave 9 AI models, from lightweight to frontier-class, one simple task across 6 of the world's biggest e-commerce sites: search for a laptop, filter by brand, sort by price, add the cheapest to cart, and verify.

A human does this in 90 seconds. Most agents couldn't do it at all.

This is Part 1 of our Agent vs. The Web series. Real sites. Real models. Real failures, recorded step by step. And every issue comes with a concrete recommendation for how to fix it.

The Hype vs. The Reality

The narrative around agentic shopping has exploded. AI-driven traffic grew by more than 200% in 2025 according to Akamai's analysis. OpenClaw , the open-source AI agent formerly known as ClawdBot that went viral this year, has collected over 193,000 GitHub stars and 33,000+ forks, making it the fastest repo to reach 100K stars in GitHub history. People are using it to autonomously browse, compare, and purchase products on the web.

So how often do these agents actually succeed?

A new paper from the Center for AI Safety and Scale AI put hard numbers to the question. Their Remote Labor Index (RLI) tested frontier AI agents including Manus, Grok 4, Claude Sonnet 4.5, GPT-5, ChatGPT agent, and Gemini 2.5 Pro on 240 real-world freelance projects worth over $140,000 in aggregate human labor. The best-performing agent (Manus) completed just 2.5% of tasks at a level a client would accept. The worst scored 0.8%.

Even on WebArena , the academic benchmark specifically designed to test web agents on simulated e-commerce, forums, and CMS sites, the best agents top out around 60% success in a controlled, static environment. When Microsoft researchers introduced real-world noise like pop-ups, slow-loading pages, and session timeouts (WABER study), accuracy for top agents dropped to near-zero in some categories.

So we stopped reading about benchmarks and ran our own tests on the real thing.

The Test

We used FlowTester to run one standardized test across 6 major e-commerce sites and 9 AI models. The test is deliberately simple, the kind of thing any shopper does daily:

steps:
  1: "Go to [website]"
  2: "Search for 'laptop'"
  3: "Select 'Brand' filter and choose 'HP'"
  4: "Sort the results by Price: Low to High"
  5: "Select the cheapest result and add it to the cart"
  6: "Navigate to the cart page and validate the laptop you chose is in the cart"

Six steps. One product. No login required. No checkout. Just: search, filter, sort, add to cart, verify.

The Models

We tested across three tiers to understand how model size and capability affect real-world web performance:

The Sites

Amazon.com · Best Buy · Target.com · Walmart.com · AliExpress · Temu

Every test was recorded with synchronized video and action logs. You can replay every click, every failure, every moment the agent got lost.

A note on WebMCP: Throughout this article we reference WebMCP, a new specification that allows websites to declare machine-readable tool endpoints for AI agents. WebMCP is very new, and none of the sites we tested have had time to implement it yet. We still check for it because it represents the direction the industry is heading, and early adopters will have a significant advantage as agent traffic scales.

The Results

🏆 Amazon.com: The Winner (With Caveats)

Amazon was the only site where all 9 models completed the full task. Sort of. The product search was clean. Filters were accessible. Sorting worked. Add-to-cart succeeded. However the smaller models, chose the cheapest result which was not a laptop, SOTA models made the assumption we're looking for a laptop.

But “winner” comes with a giant asterisk: Amazon is known to block 47 AI bots and has sued Perplexity over unauthorized agent access. Our tests ran in a clean browser environment, and the same conditions wouldn't hold for an unauthenticated agent hitting Amazon at scale. The site that works best for agents is also the one most aggressively blocking them.

Follow Gemini Pro 3 finding an affordable HP laptop on Amazon.com

Target.com: Well done but slower

Target scored fair on AI agent readiness, and the reasons paint a picture of a site that's halfway there. On the positive side, Target has a comprehensive llms.txt file, good ARIA labeling on key interactive elements like the search bar, buttons, and cart, and functional autocomplete on search. The foundations are solid.

However Target's SPA architecture means skeleton loading screens appear after every action, forcing agents to implement wait strategies before they can interact with anything. The add-to-cart flow is multi-step — requiring more clicks than necessary, which compounds errors for lighter models. An intrusive sign-in popup blocks initial interaction on every fresh session, and robots.txt restricts access to search and cart pages, the exact pages agents need most.

There's no WebMCP implementation, meaning all AI agent interactions must happen through DOM manipulation, the slowest and most error-prone path. For agents, every skeleton loader is a moment of uncertainty, and every extra step in the add-to-cart flow is another chance to lose context. The site would benefit most from adding explicit loading state indicators, simplifying the add-to-cart flow, and implementing WebMCP tool declarations for core commerce actions.

Watch GPT 5 shopping for a low-cost HP laptop on Target.com

AliExpress: Shadow DOM and Pop-Ups

A small detour first: AliExpress doesn't carry HP laptops as a filterable brand, so we swapped the filter step to Lenovo for this site. Same test, different brand.

AliExpress presents significant challenges for AI agents. The SPA architecture serves empty initial HTML, severely limiting AI crawlability. There's essentially nothing for an agent to parse until JavaScript hydrates the page. Worse, critical interactive elements like filter controls and sort dropdowns live inside Shadow DOM with cryptic or missing aria-label attributes, making them nearly invisible to agents relying on accessibility-tree parsing.

Then there are the pop-ups. Multiple intrusive overlays fire on page load — promotional offers, app install prompts, and cookie banners, creating an immediate gauntlet before the agent can even reach the search bar. Some models wasted 10-15 actions trying to dismiss them before giving up or accidentally interacting with a promotion.

The silver lining: once past the pop-ups, the core flow (search, filter, sort, add to cart) is functionally completable. Navigation patterns are somewhat predictable, and pages load within 3-5 seconds. But there's no WebMCP implementation, so all interactions must happen through DOM manipulation. The site needs descriptive aria-label attributes on filter and sort controls, SSR for critical content, pop-up management for automated agents, and WebMCP tool declarations for core e-commerce actions.

See how Claude Opus 4.6 search for a budget-friendly Lenovo laptop on AliExpress.com

❌ Best Buy: The Site Broke Itself?

Best Buy was the most fascinating failure in our test because the agents did everything right and the site still broke. Basic navigation works: agents could reach the homepage, find the search input (buried inside Shadow DOM, but reachable), and execute a search. The larger models even located the filter panel and attempted to apply brand filters.

Then two critical barriers appeared. First, the SPA product grid fails to render after sort or filter interactions. The page's heavy reliance on client-side JavaScript rendering with Apollo GraphQL means that after applying a filter, agents see empty states, partial loads, or stale results from the previous unfiltered view. The agents can't browse sorted or filtered results because the results simply don't appear. This isn't an agent problem; the website itself can't reliably complete the render.

Second, product detail pages return HTTP/2 protocol errors for automated browsers, completely blocking access to product information and the add-to-cart button. Even if an agent somehow navigated past the broken grid, it couldn't reach the product page.

The details get worse from there. The search input lives in Shadow DOM and fails silently without proper focus management. Multiple filter inputs have incomplete ARIA labels — literally containing placeholder TODO text. There's no WebMCP implementation, and aggressive bot detection adds another layer of friction. Any workflow requiring sort, filter, or product detail access is effectively blocked. Best Buy is one of the most challenging major retail sites for AI agent automation.

Follow Gemini Pro 3 spending 25 minutes debugging BestBuy.com

🚫 Walmart.com & Temu: The Bot Wall

Both Walmart and Temu stopped agents dead on arrival. Before any product search could begin, bot detection systems triggered and blocked all 9 models. CAPTCHAs, challenge pages, and behavioral fingerprinting prevented any meaningful interaction.

This is an increasingly common pattern. Retailers have legitimate reasons to block automated traffic (scraping, price manipulation, inventory abuse). But the same defenses that stop bad actors also block the next generation of customers.

Even Gemini Pro 3 cannot Press & Hold on Walmart.com

The Model Tier Breakdown

The gap between model tiers was dramatic, and not always where you'd expect.

Light Tier: Gemini Flash Lite 2.5 · GPT-5 Nano · Haiku 4.5

The lightweight models struggled across the board. Haiku 4.5 in particular failed consistently, not because of the websites, but because it couldn't handle the context required for browser-based interaction. The DOM snapshots from modern e-commerce sites are enormous, and Haiku simply couldn't process them effectively, producing malformed responses or failing to generate valid actions.

GPT-5 Nano and Gemini Flash Lite performed better but still failed on any site that required more than 3-4 clean interactions. Filters and sorting were their breaking point.

Balanced Tier: Gemini Flash 3 · GPT-5 Mini · Sonnet 4.5

The balanced tier showed the best cost-to-performance ratio. These models could navigate successfully and made meaningful progress even on Best Buy before the rendering issue stopped them. They handled pop-ups better than the light tier but still struggled with more persistent issues.

Heavy Duty Tier: Gemini 3 Pro · GPT-5.1 · Opus 4.6

The frontier models demonstrated impressive resilience. Honestly all of them performed flawlessly and they only failed when the website itself failed to render results.

But heavy duty comes with heavy costs. These models are much more expensive per run than the light tier. For recurring agent-based shopping, the economics don't work yet unless the task completion rate is near-perfect.

So what should you do for you website?

Across all 54 test runs (9 models × 6 sites), we identified five consistent failure patterns. Every one of them is fixable.

Pattern #1: The Bot Wall

Bot detection systems blocked all 9 models before any interaction could begin. This is the nuclear option. It doesn't just prevent bad actors; it prevents all agent-based commerce.

The industry is actively building alternatives. Google's Universal Commerce Protocol (UCP), developed with Shopify, Target, Walmart, and others, aims to create authenticated agent access. Stripe and PayPal have launched agentic commerce protocols for secure transactions. Visa's Trusted Agent Protocol and Mastercard's Agent Pay are building the trust layer.

The fix: Implement agent-permissioned access: a verified pathway for trusted AI agents that sits between “block everything” and “open to all.” Start evaluating UCP, Stripe's ACP, and AP2. Sites that solve this first will capture the next wave of commerce.

Pattern #2: The Persistent Pop-Up

Promotional overlays that appear on every page load are an agent killer. The AliExpress “free gift” pop-up used custom DOM elements without semantic roles, making it invisible to agents relying on accessibility-tree parsing. Microsoft's WABER study confirms this is a top failure pattern: agents faced a 34.3% increase in cost when handling pop-ups, and shopping task accuracy dropped to as low as 10%.

The fix: Ensure overlays use role="dialog" and aria-modal="true". Provide semantically labeled close buttons (aria-label="Close dialog"). Detect automated user agents and suppress non-critical overlays. If the pop-up isn't critical to the purchase flow, it shouldn't block agents from reaching the purchase flow.

Pattern #3: Context Poisoning

When a site's DOM is extremely dense, deeply nested, or slow to render, agents burn through their context window just navigating the page. Heavy page loads meant agents spent thousands of tokens on DOM state before reaching the filter step. By the time they got to “add to cart,” the earlier context had degraded, causing misidentification of elements, repeated actions, or complete loss of the task objective.

The fix: Optimize your page weight and DOM depth. Use semantic HTML landmarks (<main>, <nav>, <aside>, <section>) so agents can skip irrelevant content. Implement proper heading hierarchy. Reduce the number of nested <div> wrappers. Every unnecessary DOM node is a tax on agent performance.

Pattern #4: Rendering Failure

This was the most unexpected finding. Even when frontier models successfully navigated complex UIs, the website themselves sometimes failed to render proper results. The agents did everything right; the site did not. Client-side rendering failures, stale state after filter application, and incomplete hydration meant the agents were trying to interact with a page that hadn't finished building itself.

The fix: Ensure filtered/sorted results render reliably under all conditions including when the user agent doesn't execute JavaScript in the typical human pattern. Server-side render critical product listing state. Provide loading indicators with aria-live regions. Make sure your site works before you blame the agent.

Pattern #5: Model Capacity

Modern e-commerce pages produce enormous DOM snapshots, often exceeding the practical context window of lightweight models. Smaller models couldn't generate valid actions for any site in our test. This isn't the site's fault, but it is a reality that site owners need to plan for: the agent visiting your site might be running a small, cost-efficient model, not a frontier giant.

The fix: The more semantic and structured your HTML, the less context budget an agent needs to understand it. Clean markup isn't just good practice; it's the difference between being accessible to a $0.001/request agent and requiring a $0.10/request model to navigate your site.

Why This Matters Right Now

OpenClaw has shown what happens when agentic AI goes mainstream. With 193,000+ GitHub stars and 33,000+ forks and integrations spreading from Silicon Valley to China, people are already pointing AI agents at real websites and expecting them to work. The viral explosion of OpenClaw (which went from obscure hobby project to 100K stars in two days, faster than any repo in GitHub history) is a preview of what's coming. Users aren't waiting for the protocols to be finalized. They're sending agents to your site today.

And those agents are failing. Quietly, invisibly, with no error logged on your server and no support ticket filed. The agent just… moves on to a competitor whose site it can navigate.

If a checkout flow can't accept an authenticated agent, the agent will move the transaction somewhere else. It won't debate it. It won't try again. It will simply pick a merchant it can complete the cycle with.
— Previsible, “Agentic Shopping in 2026”

Amazon's Andy Jassy acknowledged the gap publicly, telling analysts that most AI shopping agents fail to provide a satisfactory customer experience, lacking personalization and providing inaccurate pricing and delivery information. When even the CEO of the world's largest e-commerce company admits the technology isn't working, the question shifts from “should we care?” to “how fast can we fix it?”

What's Next

This is Part 1 of our Agent vs. The Web series. E-commerce is just the beginning.

What should we tackle next? Every installment will include real test data, real failure analysis, and actionable recommendations you can implement immediately.

The trillion-dollar agentic commerce future isn't going to build itself. The websites have to be ready for it. Right now, they're not. Let's fix that.

Methodology note: All tests were conducted in February 2026 using FlowTester's browser-based testing framework. Each test was run with a consistent, standardized task across all 9 models and 6 sites. Results reflect the state of the tested websites at the time of testing; site updates may have changed behavior since publication.