Benchmark: search restaurants on Lieferando with an AI agent
On Lieferando, Wield's recipe passed 9 of 9 restaurant searches, with a median of 1.8 seconds and no model cost. A browser agent (Browser Use driving Claude Sonnet 5.5) passed 0 of 9. Lieferando's bot protection blocked it in all nine runs. Its median run took 55 seconds and cost $0.05. Both sides used three address and cuisine queries on 7 November 2026.
Lieferando Search restaurants that deliver to an address
Read-only, no loginNo reports yet
Parameters (8)
| Metric | Wield recipe | Browser Use + Claude Sonnet 5.5 |
|---|---|---|
| Success | 100% | 0% |
| Median time | 1.8 s | 55 s |
| Cost per run | $0 | $0.052 |
9 runs each, 7 November 2026. How we measure
The task
Each side ran the same three searches. Each search gives a delivery address and a cuisine.
| Query | Parameters |
|---|---|
| q1 | address = Alexanderplatz 1, Berlin, cuisine = sushi |
| q2 | address = Marienplatz 1, München, cuisine = pizza |
| q3 | address = Jungfernstieg 1, Hamburg, cuisine = burger |
Pass check: at least five restaurants, each with a name and a cuisine.
Results (7 November 2026)
| Wield recipe | Browser agent | |
|---|---|---|
| Runs passing the check | 9 of 9 | 0 of 9 |
| Median time per run | 1.8 s (1,796 ms) | 54.8 s (54,783 ms) |
| Model cost per run | $0 | $0.05 |
| Requests per run (median) | 3 | 5 (to the site: 2) |
| Blocked by the site | 0 runs | 9 of 9 runs |
Wield's fastest run took 1.0 s and its slowest took 7.6 s. The recipe had no failed runs.
Why the results differ
The recipe makes three requests per run. It runs from a Lieferando tab in your own browser, at human pace, so the search uses the session you already have. No model sits in the loop.
The browser agent ran in a fresh headless Chromium, the setup that cloud browser agents usually use. Lieferando's bot protection refused its requests. The log records nine HTTP 403 responses from www.lieferando.de, and the agent never reached a restaurant list. Its runs were cheap at a median of 6 steps, 16,150 input tokens and 948 output tokens, but they did not return the data.
Wield does not try to get past a check. If Lieferando shows one to a Wield user, Wield stops with site_challenge and the user solves it in their own tab.
Caveats
- Small sample. Nine runs per side, on three queries.
- Preliminary. The benchmark was still running on 7 November 2026.
- Paced. Runs are paced at about 30 site requests per hour, so counts are low on some sides.
- Model cost only. Wield's $0 covers model cost. It does not count your time.
Terms of Lieferando
Lieferando's terms prohibit scraping and automated systematic retrieval. The recipe runs one search at a time that you start, from your own browser, with three requests. The address lookup is billed to Lieferando by Google, so do not use it for bulk geocoding. Ordering stays on lieferando.de. Check the site's terms before any automated use.
FAQ
Does Lieferando block browser agents?
In these runs, yes. Lieferando's bot protection blocked the browser agent in all nine runs, and the log records nine HTTP 403 responses. Wield's recipe passed all nine runs in the user's own browser.
How fast is the Wield recipe for search restaurants on Lieferando?
The median run took 1.8 s. The fastest took 1.0 s and the slowest took 7.6 s. The browser agent median was 54.8 s, and it passed none of its runs.
Can I rerun this benchmark?
The method, the task file and the raw logs are described on the benchmark page. The logs record destinations, timing, tokens and checks. They hold no page contents.