When an AInvasion page says something was checked, this page tells you exactly what that means, who did the work, and what we do not claim.
Maintained by Tom Morgan, Founder & Editor · Last reviewed: October 2026 · Companion to the Methodology
Who does the checking
AInvasion uses Claude, an AI assistant made by Anthropic, for most of the legwork: finding primary sources, reading filings and pricing pages, cross-checking figures, and recalculating numbers. The editor sets the standards, reviews the results, decides what is published, and is accountable for corrections.
That split has a consequence we state up front: AInvasion does not currently run hands-on tests of commercial AI products. Nobody here subscribes to a tool and uses it for weeks. What you get instead is documented verification, and a label that tells you which kind.
AI output is never evidence by itself. A figure without a link to a source does not get published as fact, no matter how confident the draft sounds.
The four labels
Every verification note on the site uses one of these. They are ordered by strength of evidence, and a page may carry different labels for different claims.
| Label | What it means | What it does not mean |
|---|---|---|
| Run | Code or a calculator was executed on defined inputs, and the outputs were recorded. | That the tool is accurate on inputs we did not try. |
| Recalculated | A number was re-derived from stated inputs and matched (or did not match) the published figure. | That the inputs themselves are right. They are sourced separately. |
| Verified against source | The claim was traced to a primary source: a ruling, filing, official documentation or a live pricing page, with the date it was read. | That the source is correct; only that we report it accurately. |
| Not verified | The claim comes from the company or person making it. We attribute it and do not state it as fact. | That the claim is false. |
What we check, by type of page
| Page type | What gets checked | Where it comes from |
|---|---|---|
| Postmortems | Dates, amounts, who decided what, what a ruling actually held. Quotes are checked against the original text. | Court and tribunal rulings, company filings, official statements, first-hand reporting |
| Pricing and cost pages | Every price against the vendor’s live page, plan by plan. Totals recalculated from the stated inputs. | Official pricing and documentation pages, read on the date shown |
| Calculators and tools | Formulas re-derived and the tool run on sample inputs, including edge cases such as zero or very large values. | Our own code, vendor rate cards |
| Regulation | The legal text and its dates, separated from commentary about what it might mean. | Official journals, regulator pages |
| Benchmarks and vendor claims | Test conditions, version, dataset and who ran the test. A vendor figure stays a vendor figure. | Papers, benchmark documentation, independent reproductions where they exist |
The Test Card
Pages about tools and pricing carry a short card like this one. The fields below are the format; real cards contain real values only.
Subject: [tool or claim] Label: [one of the four labels] Plan / version: [exact tier or version checked] Checked on: [date] What was done: [plain description] What was not checked: [stated limits] Sources: [links]
The line “what was not checked” is the one we care about most. A verdict without its limits is a marketing claim.
What we do not do
- We do not sign up for, subscribe to or use the commercial tools we cover. “Tested” on this site never means “we used it for a month.”
- We do not run lab-grade benchmarks of model quality. When we cite one, it is someone else’s, with their conditions attached.
- We do not rank tools for payment. Affiliate links, where used, are disclosed and do not change a verdict (see Methodology).
- We do not treat a vendor’s own number as an independent finding.
- We do not promise a fixed re-check schedule. Prices and features move faster than any schedule, so every page shows when it was last checked, and you should weigh that date.
Where this can still go wrong
Claude, like any AI model, can misread a table, mix up plans, or state a number with more confidence than the source supports. The safeguards are the link requirement, the recalculation step, and an editor reading the result. They reduce errors; they do not remove them.
If you find one, tell us. Material corrections are noted on the page with what changed and why.
Questions
Do you hands-on test the tools you cover?
No. We do not currently run hands-on tests of commercial AI products. Entries are verified against documentation and primary sources, recalculated where numbers are involved, and our own calculators are run on defined inputs. Each page says which of these was done.
Is AI involved in the testing?
Yes. Claude, an AI assistant made by Anthropic, does much of the research, source-checking, and recalculation. The editor sets the standards, reviews the output, and is accountable for what is published.
How often are pages re-checked?
There is no fixed schedule. Pages show their last-checked date, are updated when facts change or a reader reports a problem, and note material corrections.
What does “Not verified” mean?
The claim comes from the company or person making it, and we could not confirm it independently, so we attribute it to them.
See also: Methodology · Postmortems · AI tools · AI Cost Calculator · Author