Benchmarks
One test, published with its limits. Run 22 October 2026.
Look-alike targeting
A page with ten traps that real sites have: three product cards with the same "Add to cart" button, table rows with identical "Edit" links, two forms that both ask for an email, a shadow-DOM banner, a modal over a page with its own "Confirm", disabled and hidden copies, and icon-only buttons. Each tool was driven by two simulated agents:
- Careful: reads the context around each element before choosing.
- Careless: picks the first element with the right name.
| Tool | Careful agent | Careless agent |
|---|---|---|
| BrowserZap | 10 correct, 0 wrong | 10 correct, 0 wrong |
| Playwright MCP | 10 correct, 0 wrong | 3 correct, 6 wrong, 1 timed out |
| Chrome DevTools MCP | 10 correct, 0 wrong | 3 correct, 6 wrong, 1 reported no error but clicked nothing |
With a careful agent all three tools target correctly. The difference is what happens after a slip. Ref-based tools click whatever the agent picks, silently. BrowserZap refuses an ambiguous target and returns the candidates with their context, so even the careless agent got all ten right.
Cost per decision: about 1 KiB from BrowserZap, against a 55–106 KiB page snapshot from the other two. BrowserZap was the fastest in 9 of the 10 cases.
What this does not show
- The agents are simulated. Real models behave somewhere between the two.
- The timings come from one machine, headless, so compare within a row rather than across machines.
- It is one page with ten cases written by us. It tests targeting only, not navigation, speed of real sites or long tasks.
Questions about the method: [email protected].