Five Things Every AI Tool Test Keeps Proving

I test AI tools the same way every time. Hands-on, on the free tiers, on real tasks, and I write down what actually happened rather than what the launch post promised. Do that across enough categories, assistants, coding tools, search engines, and the same handful of patterns keep surfacing. I only realised how consistent they were when I fed three of my own comparisons into Google NotebookLM and asked what they had in common. It drew the through-lines for me, and they were hard to argue with. Here are the five that show up every single time.

1. Correctness has converged, so fit beats power

In my assistant test, all three, ChatGPT, Gemini and Claude, got every task right; the difference was voice and interface, not accuracy. In the coding test, basic completion was a solved problem across the board. In the search test, all three tools tied on getting the facts right. The lesson repeats. The top tools in a category are now close enough on raw capability that the real question is which one fits your workflow, not which is cleverest.

2. The free tier is usually enough

I run these tests on the free tiers on purpose, because that is where most people start, and the free tiers keep passing. Every assistant answered every question free. The search tools handled daily queries free. Even in coding, the free tiers cover light use, and you only pay when volume or a specific feature blocks you. Paying is worth it when a limit actually gets in your way, not by default, which is the whole argument behind the cheapest AI stack.

3. Trust, but verify, every time

This is the most consistent warning across everything I test. AI output looks confident and polished whether or not it is correct. Coding assistants produced code that compiled and was subtly wrong. Search tools can still err in fluent prose. Even the best of them needs a human check, a test suite, a clicked citation, a second read. Verification is not an optional extra, it is the job. The tools that earned my trust were the ones that made checking easy, by citing sources or showing their reasoning, not the ones that sounded most sure.

4. The friction is often the deciding factor

When tools are this close on capability, the small things decide it. Whether an assistant makes you log in or lets you start cold. Whether a coding tool lives in your existing editor or forces you into a new one. Whether a search shows its sources as clickable cards or buries them. Again and again, the right pick for a given person was simply the one that slotted into their existing setup with the least friction. Capability gets a tool onto the shortlist; friction decides the winner.

5. Marketing is noise, real tasks are signal

The single most useful thing I do is ignore the spec sheets and hand every tool the same real job. An awkward email. A buggy function. A study I made up, to see which search engine would invent a summary rather than admit it could not find it. Benchmarks and launch threads rarely predict how a tool feels in the work you actually do. The only reliable test is your own, run on something you genuinely need done.

So how should you actually choose?

If you take one thing from all of this, it is that you should not agonise over which AI tool is best in the abstract, because there is rarely a single answer. Pick the one that fits your task and your setup, start on the free tier, keep a human in the loop, and run one real job through it before you commit a penny. That advice has survived every test I have thrown at it, and it will outlast whichever model happens to be on top this month.

This piece grew out of feeding my own reviews into NotebookLM and asking what they agreed on, see how that test worked. To match a tool to a specific task, use the recommender, and to see what everything really costs, the pricing comparison.