How editorial decisions
get made.

Editorial independence — the firewall

The scoring code is editorial; the affiliate-link code is commerce. They live in separate source-control branches and merge through a hard boundary in BankReviewLayout.astro: the scorecard renders before the affiliate ledger loads. A reviewer cannot change a score to lift a partner's rank; an affiliate manager cannot move a button to win a higher commission. The line is mechanical, not aspirational.

Editorial decisions — what to publish, what to retract, what to score, what to refuse — are made by the named editor (Stephan Kulik). Affiliate operations — which partners to accept, which commissions to negotiate, which programs to retire — happen on a separate calendar, recorded in the public ledger at /disclosure/. The two calendars never converge in the same person on the same day.

The fact-check workflow

Every claim that decides whether a neobank is safe to use — licence status, deposit guarantee scheme, regulator name, current pricing tiers, regulatory sanctions — runs through a three-stage check before publication.

Stage one: primary source. The factual claim must trace to a regulator's register or the bank's own legal-entity disclosure file. The URL of that source gets captured in the fact database ( /glossary/) alongside the claim itself. Secondary sources (Reuters, FT, Bloomberg) are cited but do not stand alone — at least one primary source per claim is the minimum.

Stage two: retrieval date. Every fact is stamped with the date it was last verified against its primary source. The stamp is what powers the "Last verified YYYY-MM-DD" line on individual data points (currently rolling out across the EU 5 cohort; full deployment tracked in /corrections/'s changelog).

Stage three: pre-publish cross-check. Before any score changes ship to production, the previous edition's claim and the new edition's claim are diffed. Any cell that moved by ≥5 score points triggers a second look at the underlying facts. Cells that moved by less than 5 points still get the retrieval-date refresh but do not require the extra review.

Correction procedure — the SLA

When a reader, regulator, or bank flags a factual error, the SLA is:

• Acknowledgement within 24 hours — a human editor reads every [email protected] message; auto-responders or AI-only handling are not used here.
• Verification within 7 days — we check the claim against its primary source. If we agree the claim is wrong, the cell is corrected immediately and a dated note is added to /corrections/.
• Score restoration within 24 hours of verification — when a verified correction would change a score (lift or lower), the change ships in the next deploy after verification, not the next edition. The changelog records the size of the adjustment and credits the reporter (if they're willing to be credited).
• Public log — every correction lives at /corrections/ with the original error, the corrected text, the source that triggered the fix, and the date of each step. We do not silently edit and walk away.

The same SLA applies whether the error is a brand-favourable miscalculation (we put a partner too high) or a brand-unfavourable one (we put a partner too low). The published log makes that symmetry verifiable.

AI-assisted content disclosure

Editorial uses generative AI for two specific tasks; everything else is written by a named human.

Used: (1) initial-draft i18n translation — when a body of EN editorial is translated to de / fr / es / pt / ru / ar, machine translation produces the first pass. A native-speaker human reviews the result before it ships, and translation choices that carry legal or regulatory weight (regulator names, regulatory short-codes, ceiling amounts, deposit-scheme acronyms) are checked against the same primary sources the EN version cited. (2) routine reformatting — moving a table from CSV to schema.org-shaped JSON, lifting a list from prose to a structured component — is sometimes generated by AI and then code-reviewed by the editor before merge.

Not used: the editorial verdict, the per-pillar score derivation, the regulator-citation lookup, the hands-on testing observations, and the corrections-log entries are not AI-generated. The byline on a review page corresponds to the human who wrote that review's verdict; if a section was AI-drafted in any meaningful way, it's flagged at the section level.

Training data: editorial work is opted out of AI training via the Content-Signal: ai-train=no header in /robots.txt. AI citation (real-time inclusion in ChatGPT / Perplexity / Google AIO answers) stays on — it's how a meaningful share of readers find the site.

Regional contributors — the vetting

Coverage outside the EU (United States, United Kingdom, GCC, APAC, LATAM, Africa) is contributed by regional subject-matter experts who pass the same fact-check workflow as EU coverage plus a regional-credential check. The check requires:

• Identity verification — a public LinkedIn profile or equivalent, plus a reference from one editor at a recognised industry publication or a senior employee at a regulator.
• Conflict declaration — every contributor publicly lists their current and prior equity positions in any bank they might cover. The list updates each edition and is signed.
• Sample-edit pass — a 1,500-word review of a regional neobank that survives the same fact-check workflow without any cells being overturned.

Contributors get a byline on the pages they author and appear in the per-edition disclosure ledger with their cumulative payout and conflict history. The editorial workflow does not accept anonymous contributions for any page that carries a YMYL score.

Regional contributor recruitment is in progress as of Edition №08. The disclosure ledger records the current contributor roster and is updated each edition.

Conflicts of interest — the recusal rule

A contributor (including the editor) who holds a personal-investment equity stake in a bank we score is recused from that bank's review and score. Recusal is documented at the review-page level (a footnote on the page itself naming the recused contributor and the replacement) and at the edition level (the changelog records every recusal that affected a score that edition).

The same rule applies to family-member equity stakes if the contributor is aware of the position. We do not investigate beyond the contributor's reasonable knowledge — but if a previously undisclosed conflict is brought to our attention, the affected scores are recomputed under the same correction-procedure SLA above.

A separate recusal applies when a bank's product team approaches a contributor about a paid-advisory or paid-speaking engagement. The contributor must either decline the engagement or be recused from that bank's editorial for two editions following the engagement, whichever is longer. The list of past advisory engagements ships in the contributor's profile in the disclosure ledger.

What we don't do

A short list of editorial choices the site has explicitly turned down:

• Paid placement. No bank can pay to appear in the index, to appear higher in the ranking, to get a "Recommended" badge, or to have a negative cell softened. The disclosure ledger publishes every commission rate; if a paid-placement transaction ever happened, it would show up there and would constitute a violation of this policy.
• Sponsored review pages. Every brand review is editorial. A review page existing on the site does NOT imply an affiliate relationship — about 20% of the index has no affiliate relationship with the site at all. Reviews are written because the bank is in the analytical scope, not because the bank pays.
• Reciprocity. We do not run "you write about us, we write about you" arrangements with banks, regulators, or other publications. Quotes are attributed to their source; we do not credit a bank's PR team for a quote a researcher in that team never said.
• Score retroactivity for legal pressure. A bank's legal team requesting a score change is treated as a correction request — it runs through the same workflow above. The change happens iff the underlying facts changed; "we'd like a higher score" is not a fact.