Is my website usable by AI agents?
Usable means an AI agent can complete a task on your site, not just read it: find the right product, add it to the cart, book the slot, submit the form. The measure is task completion, and behind it time, steps and tokens. In a controlled benchmark across five commerce sites and 1,000 runs, agents on the human-facing pages completed between 91% and 100% of tasks. On the agent-facing version of the same sites they completed 100%, 24% to 35% faster, with up to 38% fewer steps and zero failed runs.
Readable is the first bar. Usable is the second, and it is where computer-use agents, the research-and-task agents that act on a user's behalf, succeed or give up.
The four numbers
- Task completion. Did the agent finish? A failed run on a commerce site is an abandoned cart, except the customer never saw your site and will not try again.
- Time to completion. Agents run against a clock and a budget. Improvements in the benchmark ranged from 24.3% to 35.1% across the five sites.
- Steps. Every click, scroll and dismissal is a step. Agents verify each step with a screenshot. Step reductions in the benchmark ran from 18.6% to 37.9%.
- Tokens. Screenshots are vision tokens. Fewer steps, fewer screenshots, lower cost per task, and that cost lands on whoever runs the agent, increasingly your customer.
What breaks a task
The benchmark reviewed every failure on the human-facing sites. None was the agent misunderstanding the task. Of 21 terminal failures, 12 were bot detection (including silent same-page blocking that never surfaced as an error), 8 were decision loops triggered by popups and unstable navigation, and 1 was a stuck page load. In every case the agent knew what to do and the page stopped it.
The patterns that cause this are the same patterns that lift human conversion:
- Overlays. Newsletter modals, cookie walls, promotional interstitials. Each is a detour the agent has to detect, dismiss and verify.
- Hover menus and dropdowns. Navigation that exists only while a pointer rests on it is navigation an agent cannot see in a static snapshot.
- CAPTCHAs and bot-mitigation defaults. A challenge on a product page blocks the agent that was about to buy, alongside the scraper it was aimed at.
- Visual clutter. Chat widgets, ad slots, carousels and animations enlarge what the agent must scan and multiply screenshots.
- Client-side rendering. A DOM that reshapes itself between visits forces the agent to re-parse the page every time.
How to run your own five-task test
You do not need a lab. Pick five tasks a customer would ask an assistant to do on your site, for example: find the cheapest variant of a named product and add it to the cart, find the return policy for an opened item, book the earliest available appointment next week, download the spec sheet for a model, and find out whether a product ships to a given country.
Run each task ten times with a computer-use agent (the agent mode of a major assistant, or a developer framework) against your live site. Record four things per run: completed or not, seconds, steps, and where it stalled. Ten runs per task matters, because the clearest finding in the benchmark was variance. Human-facing sites completed the same task in widely different times from run to run, and agent-facing sites did not.
Then read the stalls. They will cluster on two or three page elements. Those are your fixes.
What an agent-facing rendering changes
An agent site keeps the human site exactly as it is and serves agents a transformed version: ad scripts and trackers removed, overlays and modals suppressed, CSS animations disabled, sticky and fixed elements made static, forms and buttons kept intact. Products, prices, inventory and checkout logic are the same, because the rendering is produced from your origin page at fetch time.
Where TollBit fits
TollBit Agent Sites are the agent-facing rendering tested in the benchmark, served to identified agents on your own domain through a subdomain and a CDN rule. Server-log analytics show which agents ran which workflows and where they succeeded or failed. Per-site results are in the whitepaper.
Frequently asked questions
Is usability only a commerce problem? No. Any task an assistant might do for a user applies: booking, quoting, downloading documentation, filling in a lead form, checking availability. Commerce is where the cost of a failed task is easiest to count.
Will fixing usability for agents hurt human conversion? It does not touch human pages. The agent-facing rendering is served to agents only, so the overlays and promotions that work on humans stay where they are.
Do agents really bear a cost per step? Yes. Major providers have moved toward usage-based billing, and each step an agent takes to verify state is a screenshot billed in tokens. A site that takes 40 steps instead of 25 costs the user more every time.