Four of Our Card Pages Told You We Could Not Verify Data We Already Had

Every product page on this site carries an editorial section: our reading of what the verified figures mean, kept deliberately separate from the figures themselves. Those sections were written on 21 August 2026 for all 41 credit cards we track.

On 22 and 23 August we ingested a new field for the same 41 cards — the per-category earning rates, each with the issuer page it came from and the date we read it. That is the work that made the rewards calculator possible. What it did not do was change a single word of the analysis written two days earlier.

On four cards, the analysis had said the rewards were unverified. That had been true when it was written. It stopped being true almost immediately, and stayed published until 6 September.

A stale figure misinforms you. This did something different and, we think, worse: it told you this site could not help you with the single thing you came to the page for, and sent you off to the issuer to find it — while the answer sat a few hundred pixels further down the same page, sourced and dated.

The whole proposition of Bank Compass is that every rate, fee and cap links to where it came from and when we read it. A page that understates its own data is not a small blemish on that promise. It is the promise failing in the direction nobody checks, because a site claiming to know less than it does does not look like it is cutting corners.

This is the part worth being precise about, because it is the part that generalises. Nothing was broken. Type checking passed. The linter passed. All 457 unit tests passed. The production build passed. The accessibility sweep across all 155 pages passed.

None of them could have caught it. Every one of those checks asks whether the code is correct. None of them asks whether an English sentence about the data is still true of the data. Between those two questions is where this class of error lives, and it is not a small gap — we had already been bitten through it once, in September, when thirty figures typed into our bank analyses went stale after a data refresh and nothing complained.

That earlier incident produced a test that checks every number asserted in our prose still exists in the catalog. It did not occur to us at the time that a claim about the *absence* of data could rot exactly the same way.

We could have corrected the four pages and resolved to be more careful. We have some evidence about how well that works: this is the second time the same structural gap has produced a published error in a month.

So the correction ships with a test. For every card where the catalog holds verified earning rates, it reads the editorial prose and fails the build if that prose claims the rates are unverified. The same check covers caps. It runs on every commit, and it cannot be satisfied by remembering to look.

Writing it took some care in one respect worth mentioning, because it goes to the heart of what this site is for. A page should say when a figure is missing — that rule is why the two Capital One student cards in our lineup carry no APR today and say so plainly rather than quietly filling in a plausible number. The test had to catch the false claim without making the honest one impossible. Its first version failed that: it flagged a Bank of America page for correctly noting that relationship-tier bonuses are not something we verify, and then flagged those same two student cards for correctly saying we have no APR for them. Both patterns were narrowed until only the wrong sentences failed.

Each analysis was rewritten around the figures we hold. Chase Sapphire Preferred now sets out the 5x, 3x, 2x and 1x rates and says which part remains unverified — the issuer's own summary line, not the rates. PNC Cash Rewards leads with the 4% on fuel, the highest fuel rate in our lineup, and the shared annual cap that decides what it is worth. TD Cash explains the self-select structure. U.S. Bank Altitude Connect states the $1,000 quarterly cap and what it puts a ceiling on.

In the course of that rewrite we caught three sentences of our own, written the same afternoon, converting points into dollars. Point redemption value is not a field this catalog verifies, so those sums were resting on a number we do not have. They now stop at the points, and say why.

Nobody wrote in. No reader was harmed in a way we can identify, and the four pages have been quietly correct since 6 September. We could have said nothing and the error would have left no trace outside our own commit history.

Our editorial policy commits to correcting errors at the source of the data so the change propagates everywhere at once, to moving the verification date so the page shows it was re-checked, and to recomputing any ranking a correction disturbs. A commitment like that is worth exactly as much as the first time it costs something to keep.

If you find a figure on this site you think is wrong, the contact page reaches a person. That is the other half of the same policy, and it works better than our automated checks did here.