Quality
How we test the results
What we check before a result is shown and after every run, as our code does it today. The figures on this page are read from our test data when the page loads.
Sample results, checked twice
Many tool pages offer a “Sample result”: the real output of the tool’s own prompt for an example input, generated once and stored. Before one is published, it passes two checks.
- An automatic check rejects outputs that are too short, refuse or ask for more input, are in the wrong language, look cut off, leak the prompt or leave template placeholders.
- An AI reviewer then scores the rest for usefulness, language, completeness and safety, and low scores are dropped.
4169 sample results are published in 18 languages. In the latest checks, 3795 new outputs were reviewed and 568 of them were rejected.
Checked after every run
- After a writing tool answers, a fixed check with no second AI call removes assistant filler such as “Certainly! Here is your post:” and sign-off lines, and records any of 15 listed AI clichés left in English text.
- Writers with the “Fact-check with sources” switch search the web once for your topic, cite what they use as [1], [2] and list the sources at the end. A claim without a source is marked “(unverified)”, and no link may be invented.
The right model for each tool
Conductor sends each tool to one of 6 model families (Claude, GPT, Gemini, Cerebras, Grok, Mistral) and switches to another when a provider fails. Every result names the family that wrote it.
Every change is tested before it ships
A change reaches the site only through a pull request that passes an automated gate: code checks, contract tests, translation coverage in all 18 languages and a secret scan. After each release, browser test suites run against the live site.
What testing cannot promise
AI can still be wrong. Check facts, figures and legal, medical or financial statements before you rely on them.