Skip to content
ProvenAIStack

How we test

Most reviews of AI tools are written from the launch post and the first hour of a free trial, in a category where the marketing changes weekly and the hype never sits down. That is not where these products succeed or fail. A writing tool reveals itself on the fortieth piece of production content, an automation platform in the run history of a workflow that has been live for a month, a meeting assistant in the summary you almost forwarded to a client without reading it. This page is the protocol our reports follow, the rubric that turns observations into scores, and the list of things we refuse to do.

One stance to know up front: we review tools at the workflow level, not the model level. The models behind these tools change too fast to review honestly, and most of what an operator pays for is not the model anyway. It is the workflow around it. So our reports ask what a tool does inside a working business, which changes slowly, rather than what the current model scored on someone's benchmark, which will be stale before you finish reading.

A second thing to know: the full test cycles are being stood up now, and reports published before a tool completes one say so plainly. Until then, their judgments are framework judgments built from tool documentation and the documented experience of operators, not measurements. We will never present a figure we did not observe in our own work or cite to its source.

The protocol

Every tool in a test cycle gets the same treatment:

  • Production workflows, not demo prompts. Each tool is run inside the operator's actual working systems: publishing pipelines, client deliverables, real meetings, and live automations. What a tool does with a curated demo task tells you nothing about what it does at week six.
  • The same tasks, like for like. Competing tools are given the same real tasks, repeated over the test window, so friction and failure are compared on identical work rather than on each vendor's favorite example.
  • Human review with failure modes logged. Every output is reviewed by the named human author, and the ways a tool fails, quietly or loudly, are recorded. Those logs are where verdicts come from.
  • Costs tracked at real usage. We track what each tool costs at the usage the real work generates, characterize the pricing at the model level (per seat, usage credits, flat tiers), and defer current figures to vendor pricing pages, because prices change faster than reviews.
  • An oversight level assigned from observed behavior. Each tool is placed on our fixed vocabulary: Drafting aid, Review every output, Spot-check ready, or Runs unattended. This assignment is the site's core judgment, because the honest question about any AI tool is whether you can trust it without watching it.

The scoring rubric

Scores run 0 to 10 and weigh six dimensions:

  • Output quality on real tasks. How good the work actually is on the tasks the business repeats, judged by the human who has to ship it, not by a benchmark.
  • Workflow fit. How well the tool sits inside an existing production process: integrations, handoffs, and how much of your process you must bend to fit the tool.
  • Oversight burden. How much human attention the tool consumes per unit of output, and whether that burden falls as you tune it or stays flat forever.
  • Cost behavior at realistic usage. How the pricing model behaves at the volume real work generates: per seat, usage credits, or flat tiers, and where each one bites.
  • Data control and exit paths. What you can take with you, what the tool retains, and what a migration actually costs when you leave.
  • Trajectory. Whether the vendor is investing in the parts operators depend on, judged from its shipping record rather than its roadmap promises.

What we will never do

  • Invent numbers. No fabricated benchmark scores, accuracy percentages, or time-saved statistics. Every judgment in a report comes from logged use on our own work or is cited to its source, and if neither is possible the report says the data does not exist yet.
  • Let commissions set a score. Affiliate relationships are disclosed on every monetized page, and tools that pay us nothing are reviewed under the same rubric.
  • Hide who we are. This site is operated by the maker of RankFlywheel, ProvenSEOTools, ProvenWPHosting, and ProvenCreatorTools, and pages that promote those properties label them as ours.
  • Pretend the recursion away. This site uses AI assistance to help produce reviews of AI tools. Every report is reviewed, fact-checked, and edited by the named human author before publication, and that experience of supervising AI output in production daily is part of the qualification to review these tools at all. A reviewer who has never had to catch a model's confident mistake in real work is reviewing the marketing.

Updates and corrections

AI tools change pricing, features, and underlying models without much notice, so a report is a dated snapshot, not a permanent truth. Reports carry their test window, re-tests replace stale judgments, and if we got something wrong the correction is noted in the report rather than silently edited.