Audit our credit card data yourself

This site is months old, not decades. It has no masthead to borrow from and no regulator standing behind it, so “take our word for it” is not an argument we are entitled to make. What we can do instead is show our working: for each card below, the issuer's own page, the date we retrieved it, and whether the figure we publish actually appears in what that page returned.

The check is mechanical, re-runnable, and it publishes its own failures. Of the 127 figures we claim on cards whose source we could actually read, 88 were backed up on the last pass (69%) and 39 were not. A further 36 sit behind 6 issuer sites we cannot read at all — those are not evidence in either direction, so they are counted separately rather than quietly averaged in.

Where the gaps are, by field

The failures are not spread evenly, and the shape of them is the most useful thing on this page. Fees are on the card page and back up almost perfectly. The monthly finance charge is not on the card page at all — it lives in the fee schedule, which this pass does not yet read. That is our next piece of work, and saying so is more useful than a summary that averages it away.

The rate below counts only cards whose source we could actually read. Sites we cannot machine-read are held in their own column rather than folded in as failures: not being able to check something is not the same as checking it and coming up short, and a table that blurs the two would be doing what this page exists to stop.

What each verdict means

The sample

The value in each row is read live from the catalogue as you load this page — the audit record stores field paths and verdicts, never numbers, so this table cannot quietly drift away from the data it is auditing.

How old our point valuations are

Everything above is a card field with an issuer page behind it. The other number this site publishes is what a point is worth, and that one has no issuer page — it is our estimate, not a fact anyone can corroborate. What it does have is a date. So the honest audit of a valuation is not “is it right?” but “when did someone last look?”, and until now that was not answerable: every valuation on the site rendered under a single graph-wide review date.

Only the airline and hotel programme values are looked up by a person. What a bank's own points are worth — the 21 currencies priced on /valuations, and the number most readers actually care about — is derived from the routes out of each one, so it carries no check date of its own and inherits whatever the dates below say. That is why these 37 dates matter more than their count suggests.

Those 37 valuations used to render under one hand-typed header date reading 2026-07-02. It was not the newest of them, nor the oldest, nor an average — it is a separate constant, so it happened to match 8 of 37: 2 valuations were checked before it and 27 after. A date that is wrong in both directions at once is not a freshness claim about anything; it is a header. It no longer reaches a reader: /valuations, the public API and the frozen dataset releases now publish the window these dates actually span, oldest as verifiedFrom and newest as asOf, and every valuation carries its own date beside it. The same constant had also been the asOf on all 156 transfer routes, of which it described 5.

0 of 37 valuations carry an expiry date. That is the finding, and it is not a good one. A valuation with no expiry is not fresher than one that has passed its date — it means nobody has found a dated change to schedule it against. Unscheduled is not the same as current, and a page that showed only one date for all of them could not tell you which you were looking at.

There is no re-verification schedule published here, deliberately. A service level — “every valuation re-read within ninety days” — is a promise about future work, and nobody owns a standing cadence for re-reading 37 programme pages. Publishing an interval and quietly missing it would be a worse artefact than the dates themselves, which you can read above and judge for yourself. If a cadence is ever owned, it belongs here; until then the dates are the whole claim.

How this is built

1. The sample is fixed, not chosen per run. Thirty cards across fifteen issuers, weighted towards the pages people actually read. A sample re-picked whenever the numbers would look better is a selection, not a sample.
2. The source is the issuer's own page — the same URL our apply button points at, on the bank.in domains Indian banks are required to use. If the URL we hold stops resolving, that shows up here as a failure rather than as a quietly dead link.
3. Matching allows for how Indians write numbers. ₹1,00,000 and ₹100000 and “1 lakh” are the same figure; a zero fee is usually written “lifetime free”, not “₹0”. The verdict records the source's own wording that matched.
4. We do not force an archive capture. Where the live page cannot be read, the check falls back to an existing third-party snapshot, and only if it is under six months old. A snapshot we triggered ourselves at audit time would be a weaker claim wearing the same clothes.
5. Corroboration is not proof. A figure appearing on a card page does not prove which charge it is. The link and the date are published so the reading can be checked by anyone who disagrees with it.
6. Editorial judgement is out of scope. Whether a card is worth its fee has no primary source, so it is not audited here. Only figures with an issuer-published answer are.
7. There is no accuracy figure for our calculators, and there will not be one until we can earn it. Accuracy would mean comparing what we estimated against what a card actually paid you, and the only source of that is a real statement. We take statement uploads to read your spend and discard the file in the same request — we keep nothing, by choice, so we hold no ground truth to measure against. That is the trade: a number we cannot support, or a file we do not keep. We would rather publish the refusal than an accuracy claim with nothing behind it. ⚠ The corroboration rate above is not that figure and must never be quoted as one: it says an issuer's own page agrees with a field we hold, which is a different question from whether a calculator predicted your year correctly.

Found something wrong? That is the point of publishing it. Tell us and the correction is dated in the public change record.

The underlying data is free to download and cite from the open dataset, with immutable dated releases at /data/versions. Our ranking method is on /methodology.