How we test personal AI apps

Review pages are only useful if you can tell which claims were checked and which were copied. So every review states its evidence level. This page is the standard we hold ourselves to.

Three evidence levels

• Verified by us — we created an account, uploaded documents, asked questions and recorded what happened. Pricing was read from the vendor’s own billing page on the date stated.
• Vendor-published — taken from official documentation, pricing pages or privacy policies. We quote or paraphrase and link to the source, but we did not test it.
• Not tested — anything requiring enterprise contracts, hardware we do not own, or a paid plan we did not buy. We say so instead of guessing.

What we actually measure

• Where data is stored and in what form — cloud vs device, encrypted at rest, end-to-end or readable by the vendor.
• Training policy — whether the vendor states your content is excluded from training, and whether opting out is required.
• Ingestion friction — which formats go in directly and which need conversion first.
• Answer grounding — does the product cite the passage, or just sound confident?
• Free tier reality — the actual monthly allowance, not the marketing word “free”.

What we do not do

We do not run automated benchmark suites and present them as user experience. We do not score apps out of ten to imply precision we do not have. We do not accept payment for placement, and a vendor cannot review a draft before publication.

Refresh cadence

Pricing and plan limits are re-checked quarterly; the last check for the comparison table was October 30, 2026. When a vendor changes something material, the article’s “updated” date moves and the changed line is called out.

Affiliate disclosure

Some outbound links to other products are affiliate links. When a link is affiliate, the sentence around it says so. Commission never affects ranking: our own product is listed in comparisons, and it is not ranked first where another app is genuinely better for a use case.