A bank is compared before it is contacted. Someone asks an assistant which savings account pays most, or which card carries no annual fee, and reads one composed answer assembled from rate cards, aggregators, regulator material and money-desk explainers. That answer names a rate. If the rate is yours and it is current, you are in the comparison. If it is yours and it is stale, your branch or app inherits a conversation that opens with a correction. For banking, freshness is not a hygiene detail. It is the whole of the visibility problem.
How customers actually choose a bank now
The choice never runs on the headline rate alone. A prompt as simple as best savings account interest rates in India fans out inside the engine before it produces a sentence, and what comes back is a comparison on interest rate, minimum balance, fees and charges, branch and ATM access, what the app can actually do, and deposit protection. Real cost, in other words, not advertised rate. A bank that publishes a strong rate and hides its non-maintenance charge loses that comparison to a bank with a lower rate and a legible fee table.
The journey behind the prompt is longer than the prompt looks. A need becomes a question about who is worth considering, then about what this will actually cost, then a comparison, then a judgement about which account will not irritate the customer every month. Those stages used to produce site visits. Now most of them resolve inside an answer, and only the last one reliably reaches you.
Cards are where this bites hardest, because that is where volume is moving. Credit card transaction volumes in India reached 570 crore in calendar 2025, against 216 crore in 2021, on Reserve Bank of India payment system figures reported by Outlook Money in 2026, while debit card volumes fell over the same period. A prompt like compare credit cards with no annual fee is now a high-frequency, high-intent question, and it gets answered from the fee tables an engine could parse, not from the ones it could not.
Why is rate freshness the whole problem?
Because the number moves on a schedule your publishing calendar does not follow. ICRA recorded in December 2025 that the Monetary Policy Committee had unanimously cut the policy repo rate by 25 basis points to 5.25 per cent. Each policy move sets off bank-level revisions to deposit slabs, card charges and lending spreads, and those land whenever treasury decides.
The drift is larger than most marketing teams assume. CareEdge Ratings reported in February 2026 that rates on fresh deposits at scheduled commercial banks had fallen 90 basis points over the year, to 5.67 per cent in December 2025. A rate card written twelve months earlier and never re-dated was therefore wrong by close to a percentage point on the single row customers care most about. It was not wrong because anyone lied. It was wrong because nothing on the page said when it had been true.
In classic search a stale page was survivable, because the customer clicked, landed, and saw the current figure. That correction step is disappearing. The Reuters Institute's Digital News Report 2026 found that just 4 per cent of respondents globally say they always or often click through to the underlying sources from an AI chatbot, against 19 per cent from search. The answer is the destination. Whatever number it carries is the number the customer walks in with.
What does an engine do with an undated rate card?
It looks for a source that behaves like it knows the date. Engines assemble a banking answer from product pages and rate cards, RBI circulars and directions, comparison portals, ratings and annual reports, financial publishers' explainers, and reviews and complaint data. Your rate card is one input among those, and the only one you write.
An undated rate is a weak input. The engine cannot tell a current figure from an archived one, so it hedges, attributes the number to somebody else, or takes the aggregator's figure because the aggregator stamps its listings with a date. That is the mechanism behind the complaint most banks arrive with, which is that an answer quotes a rival's rate and not theirs. The rival was not preferred. The rival was legible.
So the cheapest visibility fix available to a bank is also the least glamorous. Publish the rate with its slab, its balance tier and the date it was set. An engine that can date your number can defend quoting it.
Where a banking answer goes wrong
These failures repeat across banks, and each one is specific enough to check by hand this afternoon.
A rate that moved. Rates change more often than crawls run, so the answer quotes yesterday's number and the customer arrives expecting it. The cost is a service conversation that opens with the bank appearing to have advertised something it does not offer.
Charges buried in a schedule. If the fee table sits inside a PDF or behind a tab the crawler never opens, the engine describes the account as cheaper or dearer than it is. One direction loses the customer at the branch, the other loses the comparison outright.
No way to open. The customer is convinced and the answer has no application path to hand them, so the assistant sends them to whoever published one.
Eligibility stated too loosely. The answer implies anyone qualifies, the branch has to refuse, and the complaint becomes a review, which becomes another source the engine reads next time.
The rate page an engine can read
The remedy is a page shape, not a campaign. Rates with the slabs and balance tiers they apply to, and the date each was set. Every charge in one table, covering minimum balance, non-maintenance, transaction and card fees, rather than scattered across a schedule. Eligibility and documents stated plainly, so an engine can say who qualifies instead of telling the reader to call. Access, meaning branches, ATMs and what the app genuinely does. And a way to open, with the online route separated from whatever still needs a branch visit.
None of that is written for engines at the expense of people. EY India's 2026 study of Indian banking customers found 55 per cent want improved digital support across the app, the web and the chatbot, which is the same request stated from the customer's side. Facts a person can find without calling are facts a model can quote without guessing.
The test to apply to any rate page is simple. Could a stranger reconstruct the real cost of this account, and the date those costs were true, from the page alone? If not, the engine cannot either.
You cannot audit this from your own site
Reading your own pages tells you what you published. It does not tell you what is being said. The only way to know is to run the questions customers actually type, across the assistants they use, and read what comes back: the position you hold, the domains cited, which rivals appear beside you, and which figure the answer quotes.
Run against a fixed prompt set and the result becomes a ladder rather than a score. Absent from the ranking. Named with no rate attached. Cited as the source for a rate the answer quotes. Present in the comparison the customer is reading. Recommended for this customer's balance and usage. Each rung is a different piece of work, and a stale rate card usually blocks the middle one.
Ansyra runs that prompt set across ChatGPT, Gemini, Perplexity and AI Overviews, records the cited sources beside each answer, and shows the quoted rate next to the page it came from, so a stale figure reads as a defect on a named page. One honest limit: engines disagree with each other and with themselves, so the trend across a fixed prompt set is the signal, not any single reply.