How this was made

Methodology

Every claim in this dataset is binary, checkable, cited and graded. Here is exactly what each of those words means, and where the corpus is weak.

Claims, not ratings

There are no stars and no composite score. Each capability is a single yes/no proposition about a product, scored in one of five states:

• Yes — present, natively, with evidence.
• Partial — present but materially limited. The record must name the shortfall.
• No — positive evidence of absence.
• Unknown — assessed, unresolved.
• n/a — out of scope for this product’s peer group. Computed, never authored.

Evidence grades

Every claim carries a grade and, for A through E, a source URL and the date it was retrieved.

A yes on a differentiator-weight claim requires grade A or B. Nothing weaker. Vendor marketing establishes claims; documentation establishes facts. A feature listed on a product page tells you the vendor wants you to believe it exists; the admin guide describing how to configure it tells you it does.

Adversarial verification

Every dossier is re-examined by an independent pass instructed to refute it, defaulting to downgrade when uncertain and permitted to mark a value upheld only when it located the supporting documentation itself. That pass challenged 3,362 values and corrected 1,042 of them — 826 downgrades against 216 upgrades, a 31% correction rate on contested cells. The systematic error it caught was always the same direction: researchers reading marketing copy as capability.

2,110 cells currently carry adversarially-confirmed evidence. The full challenge record — what was contested, on what evidence, and which way it moved — is published on each product page rather than discarded after application.

Peer groups

Comparison happens within a peer group and never across one. Each group publishes objective inclusion criteria, contestable by pull request, and each claim is scoped to the business shapes it applies to. Drive-thru timers are n/a for a taproom POS, not a failure. Denominators count only in-scope claims, so a specialist is never penalised for a capability its buyers do not want.

Counting rules

The census count is generated by a published function, never authored on a record. Gross and net are reported side by side, because a defended number beats a contested one: 100 records, 93 counting toward the census. A record is excluded when it is dead or absorbed, or when it is a white-label, reseller or predecessor of another. Acquired products are never deleted — their status changes and the record persists, because the acquisition history is the point.

Sources we will not use

We do not scrape G2, Capterra, GetApp, Software Advice or Gartner Peer Insights. Their terms prohibit automated access regardless of public visibility, and importing their sentiment data would import their sampling bias with it. Capterra, GetApp and Software Advice are one Gartner Digital Markets database behind three pay-per-lead front-ends; G2 acquired Capterra in January 2026, so cross-checking one against another is no longer an independence check. They may be cited by hand for a specific proposition; they may never be evidence for a capability claim. This is enforced in code — the fetch layer throws on those hosts, and validation rejects them as evidence.

“Best restaurant POS 2026” listicles are discovery-only, never evidence. They are overwhelmingly published by competitors ranking their own categories.

Conflict of interest

Where this corpus is weak

• 15,010 of 27,472 in-scope cells are unknown. Many enterprise vendors publish almost nothing, and their records say so. A low score frequently measures disclosure rather than capability, and every product page states its own disclosure rate for exactly that reason.
• 4,168 cells are grade F — analyst inference with no source URL. They are marked as such rather than dressed up as documented.
• 61 records were scored before the six-grade rubric existed and carry rubric: legacy-v0. They have had no new research. The re-verification queue is empty, which means no individual cell of theirs breaks a rule — a weaker statement than the record having been verified. The marker records that provenance; it no longer buys those records a lower evidence bar, because the differentiator grade floor applies to every record alike.
• 57 records are scored on a headline subset rather than the full taxonomy. Coverage is stated on every record so a thin one is never mistaken for a thorough one.

Corrections

If a value here is wrong, the most useful thing anyone can do is say so with a source. Corrections arrive as pull requests against the data files and are gated by the same validation that produced them: a value change with byte-identical evidence is rejected — cite the source that shows the change, or bump the retrieval date if you re-verified. Changes requested by a vendor, what changed, and what was declined and why are all recorded publicly. That log is worth more than the corrections, because it is demonstrable independence.