Optimizing eCommerce sites for AI agents: enabling actions and task completion

Modern websites have been A/B tested and optimized around a single assumption: a human is on the other end of the screen. We've layered on popups, banners, promotional overlays, and more, all designed to capture attention and move a human toward checkout.

That assumption is no longer safe. A growing share of visitors on commerce sites are AI agents and they interact with your site differently than a human does.

For an AI agent trying to add a specific product to a cart, every one of those elements seriously increases drop-off. It slows execution, burns tokens on screenshots, and introduces brittle failure modes that turn a ten-step task into a forty-step loop that may entirely fail.

Over two years ago, we built Agent Sites to solve this problem for publishers.

We’re now applying this to eCommerce and other interactive website workflows, where the agent needs to complete actions like adding a product to cart, booking a reservation, or filling out a form.

These workflows require a working UI, and the human-facing UI is where agents get stuck.

Here is why the human web breaks for agents, what Agent Sites does about it, and data showing the importance of optimizations for agents.

Why the web breaks for agents

The same patterns that lift human conversion rates often degrade agent performance.

• Overlays and promotional interruptions. Every popup, modal, or newsletter gate is an extra decision the agent has to detect, dismiss, and verify before moving on.
• Navigation instability. Dropdowns, hover menus, and dynamically injected elements create click paths that break under automation timing.
• Security and anti-bot layers. CAPTCHAs and bot-mitigation systems block legitimate agents alongside scrapers.
• Page clutter and visual noise. Ad blocks, chat widgets, and decorative elements expand the surface agents must scan and increase the number of screenshots needed to complete a task.
• Dynamic UI and client-side rendering. Pages re-render into different DOM structures (the underlying blueprint of a webpage) between visits, forcing agents to constantly re-parse the page.

The result is that agents operating on human-facing sites face slower task completion, higher compute cost, failed workflows, and inconsistent outcomes across repeated runs.

Creating a site agents can navigate

Agent Sites follow the same pattern as mobile-optimized sites: the underlying system stays the same, but the interaction surface adapts for a different visitor.

Your site is routed through a TollBit subdomain (e.g. tollbit.yoursite.com). Agent traffic is served through this dedicated surface, separate from your human-facing site. The products, pricing, business logic, and checkout flows are identical. Only the rendered page the agent sees at runtime is different.

At a technical level, Agent Sites sit between the agent and your site. It fetches the unoptimized HTML from your origin, runs it through a transformation pipeline, and returns an optimized HTML string the agent can actually work with. The pipeline removes ad scripts and trackers, hides promotional clutter, suppresses modals and overlays, disables CSS animations and transitions, and converts sticky and fixed-position elements to static layout.

Benchmark Testing & Results

We partnered with KERNEL to run a controlled benchmark analysis to determine for agents in using a human site versus an agent site. We ran this across five ecommerce sites (Pela Case, MM LaFleur, Kopari Beauty, Quip, and Cool3C). Each site ran 100 independent executions per variant for 1,000 runs total, with task, prompt, model, and agent configuration held constant so the only variable was the surface the agent interacted with.

For the full methodology, per-site distributions, and failure-mode breakdown, read the co-published report with KERNEL.

Across 1,000 runs, Agent Sites reduced mean time to completion by 24 to 35 percent and matched or exceeded default-site task completion on every site tested, reaching 100% completion on all five. Default sites ranged from 91% to 100% completion.

Site	Time to Completion Improvement	Completion Delta	Step Reduction
Cool3C	34.8% faster	+4-6 pts	37.9% fewer
Kopari Beauty	24.3% faster	No material change	18.6% fewer
MM LaFleur	29.9% faster	+6-10 pts	25.4% fewer
Pela	28.0% faster	+0-18 pts	29.5% fewer
Quip	26.1% faster	+0 pts	35.2% fewer

“In working with TollBit on this benchmark, what stood out was the consistency of the results. Running the same workflows across multiple sites, we saw that Agent Sites led to measurable improvements in time to completion and reliability compared to default site experiences.”

— Catherine Jue, Co-Founder & CEO, KERNEL

Key takeaways

The interaction surface matters. When we kept the task, prompt, and model constant, and only changed the surface from human-facing sites to agent sites, we observed material improvements on every site tested. The page itself was the bottleneck for agent.

Step reduction translates directly into cost reduction. Agent Sites reduced total steps required per workflow by 18 to 38 percent. After each step under the agent takes screenshots to verify the state after actions, and screenshots drive vision-token cost. Fewer steps means fewer screenshots, which results in lower per-task spend. That spend increasingly flows to the end user. In April 2026, Anthropic moved Claude Enterprise customers to usage-based billing, and most of the other major providers have been heading in the same direction. When your site forces an agent into extra steps, the bill for those steps lands on whoever is running the agent, which is increasingly your customer.

Less variance of outcomes. Default sites show widespread in time to completion and inconsistent completion outcomes on three of five workflows. Agent sites show tighter execution paths and zero failed runs across the full 500-run Agent Sites dataset. For anyone deploying agents at scale, predictability is often more valuable than peak speed.

Failures on default sites are rarely about the task. Of the 21 terminal failures on the default variant, replay review attributed 12 to bot detection (including silent same-page blocking that never surfaced to the agent as an error), 8 to agent decision loops triggered by popups and unstable navigation, and 1 to a stuck page-load state. In every case, the agent had correctly identified what to do but the page prevented execution. This is a different class of failure than model or prompt issues.

You can read the full methodology, per-site distributions, and the full breakdown in our co-published report with KERNEL.

What Agent Sites add beyond the optimization layer

The optimization layer is what makes agents succeed on your site. The rest of the AOS stack is what makes that success measurable and monetizable.

• Isolated agent traffic. Agent sessions route to the Agent Site subdomain rather than hammering your human-facing origin. Your CDN bill stays predictable. Your human users don't feel the impact.
• Measurement and analytics. TollBit Analytics gives you visibility into which agents are hitting your Agent Site, what workflows they run, and where they succeed or fail, broken down by AI platform, bot, and page. If you are not yet ready to create an Agent Site for eCommerce, you can still utilize TollBit to start monitoring AI bot activity on your site.

The bottom line

Speed, reliability, and cost are the three things that matter most when agents are running real workflows on your site. Agent Sites improved all three in the controlled benchmark: 24-35% faster completion, zero failed runs across 500 executions, and up to 38% fewer steps per task. Continuing to serve agents the same human-facing pages works against all three. By setting up an Agent Site, you can serve this growing class of visitors without changing anything about your current tech stack.

Get started with Agent Sites for eCommerce. We support sites hosted on Shopify and across various eCommerce platforms.