We checked 64 of our own claims against the pages we cite. 28 held up.

Every comparison site asks you to take its word for something. The useful question is how often that word is good, and the only honest way to answer it is to draw claims at random and go and look. So we did: 64 claims, eight from each of the eight classes a trader actually decides on, opened one at a time against the page our own record cites.

28 of 64 were confirmed. That is 44%, and at this sample size the true figure is somewhere in 32%–56%. It is not a good number and we are not going to dress it up.

Two ways to be wrong

The failures split into two kinds that deserve to be counted separately. 13 were contradicted: the page we cite says something different from what we publish. Those are plain errors and every one of them is listed below.

23 were not covered: the page is readable, and simply does not address the claim. Some of those values are probably right. The point is that a reader following our citation would not find out — we would be asking them to trust us, on a site whose entire pitch is that they should not have to trust anyone. A number with no evidence behind it is a number we have not earned.

Nothing in this sample was unreadable. Where a plain fetch was refused or came back as an empty shell — FundingPips answers a script with a 429, TickTickTrader and half a dozen others render their prices only in a browser — the page was opened in a real browser and read there. Grading a claim as unsupported because our tooling is not a browser would measure the tooling and call it our accuracy.

By class

The intervals here are wide. Eight claims per class buys a point estimate with roughly thirty points of slack on either side, so treat the ordering as a queue for the next pass and not as a finding. Two things survive that caveat: activation and reset fees, where nothing at all was confirmed, and the fact that no class came out clean.

The worst class, and why it is worst

Every activation-fee claim in the draw was a zero, and a zero in that field publishes as the sentence "carries no activation fee". That is a positive claim about money — it is one of the first things a trader compares between two firms — and on eight out of eight, no page we cite says it. Not one firm was caught charging a fee we had recorded as free. The failure is upstream of that: we were printing an absence of data as a fact about price.

This site has made that mistake before, and has written about it. A missing consistency percentage was once rendered as "no consistency rule", which put firms on a recommendation list whose own help pages documented one. The remedy there was to require positive evidence before publishing the negative claim. The same remedy is owed here, and the fee field never got it.

A unit error, found three times in eight draws

The drawdown class turned up the same defect at three different firms. Top One Trader publishes an 8% max drawdown on 2-Step Plus and a 5% trailing drawdown on Instant Prime; Finotive Funding publishes a 9% maximum overall drawdown. Those percentages are stored in a dollar field, so the site printed "$8", "$5" and "$9" maximum loss. The number is right and the unit is wrong, which makes the published figure absurd by three or four orders of magnitude.

One sample of eight is not enough to size that fault, but it is a fault that is either present or absent per plan and it was present in three of the eight rows we happened to draw. We are treating it as systemic until a full pass says otherwise.

Every claim a firm's own page contradicted

13 of the 64. Each is a page we opened, and each is a correction we owe. They are listed here before they are fixed, because a site that only publishes its error rate once it is flattering is publishing marketing.

How the sample was drawn

The site publishes 3,301 claims that assert something and can therefore be graded. A field we have left empty asserts nothing — it renders as "Not stated" — so empty fields are never drawn. Including them would have padded the denominator with rows that say nothing and made the rate look better for no reason.

The draw is stratified: eight claims from each class, rather than eight hundred taken at random from the whole pool. A proportional draw would have been four fifths prices and plan rules and would have said nothing about the smaller classes. The cost is that the headline figure is a mean across classes weighted by our choice of eight, not a site-wide rate, and it should be read that way.

The draw is also reproducible. It uses seed firmfax-accuracy-2026-09-06 against a fixed generator, so anyone with the repository can re-run node scripts/accuracy-sample.mjs with that seed and get these exact 64 rows. A sample nobody can reproduce is not evidence of anything.

What each verdict means

• Confirmed. A page we cite states the claim.
• Contradicted. A page we cite states something different.
• Not on the page we cite. The pages we cite are readable and do not address the claim either way.
• Could not read the page. No cited page could be read — removed, blocked, or rendered only by script.
• The page changed and our value did not follow.

A claim is only ever graded against the pages our record cites for it. Where a value is plainly true but published somewhere else on the firm's site, that is still a failure of our sourcing, and it is counted as one.

Limits

• Eight per class is thin. Every per-class interval in the table above is roughly fifty points wide. The class ordering is a work queue, not a ranking, and no single class figure here should be quoted on its own.
• It measures sourcing, not truth. A claim graded not-covered may be perfectly correct. What we have measured is how often a reader could check us by following our own citation.
• Pages move. Firms change prices and rules constantly, and a claim that was right when it was recorded can fail this check because the page moved on. That still counts as a failure — a stale record is a wrong record — but it means the rate measures freshness as well as care.
• One grader, one day. Every verdict was reached by one person reading pages on 2026-09-06. Where a judgement was close, the reasoning is in the note on the row so it can be argued with.
• Some of our sources are not the firm. The draw surfaced citations pointing at third-party directories and comparison sites. Those cannot settle a claim about what a firm publishes, and rows resting on them were graded not-covered.

What happens next

Three fixes come out of this run rather than one number. The fee field needs the same positive-evidence rule the consistency field already has, so a zero cannot publish as "no activation fee" without a page that says so. The percentage-in-a-dollar-field fault needs a pass across every plan. And the citations that point at directories instead of firms need replacing.

Then this measurement runs again, on a fresh seed, and the two numbers go next to each other. A single accuracy figure is a claim about today; a series is the only thing that shows whether the work is landing. The runbook lives in the repository at docs/routines/accuracy-measurement.md.

All 64 graded claims

Grouped by class, failures first. "Correct means" states the bar each class was held to before any of it was checked.

Download the full run, all 64 graded rows (JSON)