Rankings
Ranking methodology
Where the ratings come from, how we calculate the result, how we protect the ranking from manipulation and what its limitations are.
1. Who creates the ranking and what interest they have
The ranking is created by the Zaksio team, maker of an ERP system. This is a conflict of interest, so we say it openly: Zaksio does not take part in the ranking and never has a position or rating in it. We do not accept payments, commissions or “packages” for a place. Answers to the author's question (whether anyone has heard of Zaksio) are kept separate and do not affect any result.
2. Where the reviews come from
Only from people who use or have used the given product. The author rates the criteria on a scale of 1-10 and describes their experience in three sections: pros, cons and summary. Before filling in, everyone consents to participation and publication of the review, and it appears only after clicking the link sent by e-mail. We do not store the address: only its irreversible hash stays in the database, which lets us recognise the same author. After submitting, the author gets a code with which they can delete the review.
3. How we calculate the result
The rating of a criterion is the weighted average of the ratings of that criterion. The rating of an area is the average of the ratings of its criteria, and the overall rating is the average of the ratings of all criteria. The answer “not applicable” does not enter the average. The calculations are deterministic: no AI model performs them.
Some rankings also contain “yes / no / don't know” questions or questions with a single choice. Only those affect the rating for which the Criteria page says so explicitly (“yes” counts as 10, “no” as 1, and “don't know” is skipped). The others only give the number of answers, shown in the results as percentages.
A product with a single rating of 10.0 cannot rank higher than a product with an average of 9.8 from thousands of users, so the area ratings and the overall rating are smoothed using the Bayesian method: (v · R + m · C) / (v + m), where R is the product's average, v its effective number of reviews, C the average of the whole ranking (within the same product type) and m the prior weight. In this ranking m = 8. A product with few reviews is therefore “pulled” towards the average, and the more reviews, the closer to its own average.
4. Review weight: time and author reputation
Products change, so an old review counts for less: after 1.5 years its weight halves (the half-life period of this ranking is 540 days), but never falls below 0.1.
Author reputation: an author whose accepted reviews readers found helpful has slightly more weight (+0.1 for each “This review was helpful” vote, at most +0.5). A new author has the base weight, so having no history penalises no one.
5. Rating distribution and margin of error
For each product we show a chart of the rating distribution from 1 to 10 (what percentage of reviews has a given average) and a 95% margin of error (±1.96 × the standard deviation of the authors' ratings ÷ the square root of the number of reviews). When the difference between two products falls within the sum of their margins, we mark it instead of pretending that one is better. Reviews can be sorted: most helpful, newest, highest and lowest.
6. Credibility thresholds
A product gets a position in the ranking only after collecting at least 5 reviews, including at least 2 confirmed by invoice. Until the thresholds are exceeded, we do not publish results, only the progress of collecting reviews.
7. Invoice verification
The author can confirm using the product with an XML invoice file from KSeF (issued within the last 180 days) and receives the “Verified user” badge. The file is read only in the browser. We only receive the name of the program from the invoice header, the issue date and a hash of the invoice number, which ensures that one invoice cannot confirm many reviews. Amounts and counterparty data are never sent. Anyone can switch the results to the “verified only” view.
8. Moderation and protection against manipulation
New reviews are moderated before publication. An author whose review has already been accepted is treated as trusted, but reviews with links, repeated text or filled in suspiciously fast still go to review. Readers can report reviews (after several reports a review is hidden until a moderator decides).
We also check, among others, identical answers to all questions, repeats from the same browser or e-mail address, submission limits, sudden bursts of reviews and series of extreme ratings (all ones or all tens) about one product. Suspicious reviews are held for review and do not count until approved. We do not publish the detailed thresholds so as not to make it easier to bypass the checks.
9. AI-generated summary
When a product has at least 10 reviews, an AI model summarises the most frequently repeated pros and cons. The summary is clearly marked, refreshed after new reviews and calculates nothing: it does not affect the ratings or the position.
10. Limitations
This is a collection of reviews from people who chose to write them, not a representative survey of the whole market. That is why we provide the number of reviews, the rating distribution, the margin of error and data for your own analysis.
11. Details of this ranking
This ranking covers invoicing, accounting and ERP systems. The author of an opinion first chooses the type of system (online or installed), then picks a system from the list of most frequently mentioned ones or enters their own. The opinion is confirmed with a link sent by email.
An opinion can additionally be confirmed with a KSeF invoice file (FA), which earns the "Verified user" label: the file is read only in the browser, and all that reaches us is the program name from the invoice header, the issue date and a hash of the invoice number. Details in the article on verification and privacy.