What we check, how we check it, and what we don’t claim.
Most Lomivu articles are comparisons built from the makers’ own pages, and they say so. When we run a test ourselves, we publish the method and every result. This page explains the difference, so you always know what a verdict rests on.
- 27
- Articles published
- 380
- Sources cited
- 3 Nov
- Last updated · 2026
Three kinds of article
Comparisons
Claude vs ChatGPT, SSD vs HDD. Built from pricing pages, plan tables, spec sheets, model cards and privacy settings, checked on a stated date. They end with our verdict and the conditions that would change it.
Guides
How to fix weak Wi-Fi, how to back up a phone. Steps come from the makers’ official help pages, in the order we would try them: free fixes first, paid ones last.
Tests
Where we can measure something ourselves, we do, and we publish the set-up, the data and the result, including the parts that make a product look bad.
What we check in every comparison
- Prices on the maker’s own page.Not a reseller, not another review. If the page showed us a regional price, such as euros from Ireland, the article says which region.
- What each plan actually includes.Usage limits, context windows, file sizes, which model you really get on the free tier and on the $20 tier.
- The exact version.GPT-6 Astra, Claude Opus 5.5, Gemini 3.8 Flash, a named SSD model. “ChatGPT” or “the new iPhone” is not specific enough.
- Privacy and training defaults.Whether your chats are used for training unless you turn it off, and where to find the setting. We take this from the company’s own privacy pages.
- Vendor benchmarks, labelled as vendor benchmarks.When a company publishes test scores, we quote them with the source and say they are the company’s figures, not ours.
Our own tests so far
120 texts run through ZeroGPT
We built a set of 120 texts: 40 written by people (20 by native English speakers, 20 by non-native writers), 40 written by Claude Fable 5.1, and 40 mixed or edited texts. We ran all of them through ZeroGPT’s public detector and published every score.
Read the full test- 0 / 40
- Human texts wrongly flagged
- 7 / 40
- AI texts caught
- 3 / 28
- Edited AI texts caught
- 51 %
- Overall accuracy
The other detectors in that article need paid API keys, so we compared them on their published claims and pricing, and the article says exactly that. We’ll add more tests where we can run them fairly and repeat them over time.
What we don’t do
No invented numbers
If a maker doesn’t publish a price or a limit, we write that it isn’t published. We never estimate a figure and present it as fact.
No hands-on claims we didn’t earn
We don’t say “we tested” unless we ran the test. A comparison built from documentation says so in its sources.
No copied press releases
We read them, then check each claim against the product pages. A launch announcement is a lead, not a source.
No paid verdicts
Advertisers and affiliate partners don’t see articles before publication and can’t change a verdict.
Keeping articles current
AI models and prices change every few weeks. When something in an article changes, we update it and change the “Updated” date, and for comparisons the date of the verdict as well. Every article ends with a numbered list of the pages we used and what we took from each, so you can check any figure yourself.
Spotted something out of date? Tell us and we’ll fix it with a dated note.
Report a correction