This is a republished copy of our methodology page. The original, with the live score receipt and every category's script, lives at benchtested.app/how-we-test.
Every app in a category goes through the same written script on the same clean Windows machine. Same inputs, same steps, same measurements, so a score of 8 means the same thing for every app.
The four steps
- A clean machine. Every app is installed on a fresh Windows 11 virtual machine restored from the same snapshot. Nothing from the last test is left behind.
- The same script. All apps in a category run identical tasks, pasted word for word, with answers we collected before the test so results can be graded.
- Measured and logged. Each step is timestamped and screenshotted. We record time, cost, memory, prompts and anything the app leaves on your PC.
- A weighted score. Each axis is scored from 0 to 10 and combined with fixed, published weights. Apps that have not finished the script get no score.
Skip one step and the app stays "in testing".
Our test machine
| Host | VMware Workstation virtual machine |
| Operating system | Windows 11 Pro, English (US) interface |
| Hardware | 4 virtual CPU cores, 8 GB RAM, no GPU passthrough |
| Display | 1920 × 1080 at 100% scaling, light theme |
| Starting point | Restored from the same clean snapshot before every app |
| Accounts | Separate test accounts; free plans or trials unless stated |
We test on a virtual machine so every app starts from an identical, clean Windows install, and so nothing an app installs can reach a real computer. After each test the machine is rolled back. A virtual machine has no graphics card, so animation-heavy apps can look slower than on a typical laptop; when that affects a result we say so next to the number.
How a score is calculated
Each category has five or six axes, such as task success, speed, privacy and value. After the script has run, every axis gets a score from 0 to 10 based on what we measured. The axes are combined with fixed weights: what matters most for that kind of app counts three times, minor things once. The overall score is the weighted average, rounded to one decimal. An axis we could not measure is marked "not scored" and left out rather than guessed.
Example: Airtable, tested October 31, 2026
| Axis | Score | Weight |
|---|---|---|
| Ease of use | 7.0 | ×3 |
| Views & flexibility | 7.0 | ×2 |
| Speed | 7.5 | ×2 |
| Automation & AI | 7.0 | ×2 |
| Integrations | 7.5 | ×1 |
| Value for money | 5.0 | ×2 |
| Overall | 6.8 | Σ(score × weight) ÷ 12 |
What the verdicts mean
- Top pick: the best overall result in its category, usually 8.5 or higher.
- Recommended: a strong choice for most people who match its "best for".
- Situational: good at the job, with a drawback you should know about first.
- Avoid: failed core tasks or put your data or PC at risk.
- In testing: has not finished the script. No score, no ranking.
Test scripts by category
AI browsers (6 tested)
- T1 Install from the official site and reach the first agent task; count steps, prompts and minutes, and note anything installed outside the browser.
- T2 Run the same 8 scripted tasks on a free test account, pasted word for word: research from official sources, a form fill, a price comparison, a multi-tab extraction, a scheduled routine, a request to buy a book on a demo shop, a five-slide deck and a tricky multi-step form.
- T3 Check what the agent may do without asking: purchases, sending messages, entering passwords, and whether every action can be reviewed afterwards.
- T4 Measure cold start, idle and under-load RAM/CPU against Microsoft Edge on the same 10 tabs.
- T5 Check where chats and page data are processed and stored, training opt-outs, and what is left behind after uninstalling.
Weights: task success ×3, privacy & control ×3, speed & resources ×2, ease of use ×2, value for money ×2, everyday browsing ×1.
Project management (12 tested)
- T1 Build the same 40-task product launch project from a written brief; count steps and minutes.
- T2 Set up a recurring task, a dependency and a status automation.
- T3 Import a 200-row CSV and check what survives.
- T4 Measure page-load time for a 500-task board.
- T5 Invite a second test user and check permissions.
Weights: ease of use ×3, views & flexibility ×2, speed ×2, automation & AI ×2, value for money ×2, integrations ×1.
Evidence and screenshots
Every review includes a timestamped test log, a checklist of everything we checked, and screenshots taken straight from the test machine; nothing is mocked up or taken from vendor marketing. Screenshots carry a small badge and a faint BenchTested pattern. Personal data that appears on screen is blurred before publishing. Facts taken from a vendor's documentation rather than seen ourselves are labelled "Vendor-stated", and prices come from the vendor's pricing page with the date we checked it.
AI-assisted, human-checked
We test with the help of an AI agent and say so openly. A human editor creates accounts and handles passwords, payments and trials; the AI agent runs the script, times steps, measures and takes screenshots; the agent and then the human editor check every answer against the source; the agent drafts the write-up and scores; the human editor decides the final verdict and publishes.
Rules we never break
- No score without a test. Apps that have not finished the script are shown as "in testing", with no score and no rank.
- No paid rankings. Affiliate links never change a score, a rank or a verdict, and vendors cannot pay to be reviewed.
- No made-up results. If a test went wrong on our side, we say so and mark it "not measured" until we rerun it.
- Retests on major versions. When an app changes significantly, we run the script again and update the date on the review.
See the method in action: the AI browser ranking and the project management ranking.
BenchTested tests software on a clean Windows PC with the same script for every app. We may earn a commission through some links on the site; it never changes a score.
Top comments (0)