What does the Consumer Duty have to do with AI search?

The Consumer Duty’s consumer understanding outcome requires a firm’s communications to help retail customers make effective, timely and properly informed decisions. AI assistants now repeat those communications, often without their date or context. The Duty does not make a firm answerable for what ChatGPT says, but it does cover the pages those answers are built from.

The outcome sits in PRIN 2A.5 of the FCA Handbook (last updated 26 June 2026). In plain terms, firms must make sure their communications meet customers’ information needs, are likely to be understood, and equip customers to make good decisions at the right time. Rate pages, fee tables, eligibility rules and FAQs on a firm’s website are communications in exactly that sense.

The FCA’s own review of how firms deliver this outcome, published on 13 March 2026, surveyed 38 firms and found testing that was often superficial or one-off, and firms treating a lack of complaints as proof that customers understood (FCA, Consumer understanding: good practice and areas for improvement). It also drew on its Financial Lives research, which found 12% of adults had limited understanding of the products they held.

Customers increasingly check those products with an AI assistant first. In a Which? survey published in November 2025, about one in six people (17%) said they rely on AI for financial advice (Which? findings, reported by The Register, 18 November 2025). When the answer they get is built from an old page on your site, the gap between what you offer and what the customer believes starts on your own domain.

How do AI assistants end up quoting retired rates and wrong eligibility rules?

Usually because the wrong figure is still published somewhere they can read. Old rate pages left live, undated blog posts, PDFs from past product ranges and comparison listings that were never refreshed all stay crawlable. An assistant that finds a confident, undated statement has no reliable way to know it was withdrawn, so it may repeat it as current.

Consumer testing shows the problem clearly. In a Which? investigation published on 23 July 2026, Copilot and Google AI Overviews quoted capital gains tax rates that had not applied since October 2024; ChatGPT invented a savings product that does not exist; and Copilot quoted inaccurate savings rates (Which?, Can AI answer your money questions?, 23 July 2026). An earlier Which? test of six AI tools found repeated factual errors on money questions, including the ISA allowance (Which?, November 2025).

Some of those errors are the model’s own. Many are not: they are accurate statements from the past, still online, read without their date. The common sources inside a firm’s own estate are:

  • Pages for withdrawn products that were unlinked from navigation but never removed, redirected or marked as closed.
  • Launch announcements and blog posts that quote the rate or fee on the day they were written, with no note that it has since changed.
  • Old PDF versions of key facts documents, terms and eligibility guides, still reachable at their original addresses.
  • The same figure in several places, updated on the product page but not in the FAQ, the comparison table or the help centre.

Why do JavaScript rate calculators leave AI engines guessing?

Because most AI crawlers read only the HTML a server sends and do not run JavaScript. A rate calculator or eligibility checker that builds its figures in the browser shows those crawlers an empty shell. With nothing current to quote from your page, assistants fall back on older or third-party sources that state a figure plainly in text.

Vercel and MERJ’s analysis of AI crawler traffic, published in December 2024, found that none of the major AI crawlers it measured render JavaScript, including OpenAI’s and Anthropic’s. OpenAI’s GPTBot downloaded JavaScript files in about 11.5% of its requests but did not execute them (Vercel, The rise of the AI crawler, December 2024). Google’s own crawler does render JavaScript, so Google AI Overviews can see more of a calculator’s output, but ChatGPT, Claude and similar tools often cannot.

The answer is not to remove calculators. Customers find them useful. Keep the calculator, and next to it publish a short, server-rendered summary in plain HTML: the current headline rate or fee, the date it applies from, the main eligibility rules and one worked example. Any figure shown must still meet the promotion rules for that product, such as a representative example where one is required.

Which page problems cause stale AI answers, and what fixes them?

Most stale answers trace back to a handful of page problems: missing dates, withdrawn products left live, figures locked inside calculators or PDFs, and the same number published differently in several places. The table below sets out what a crawler sees in each case, what an assistant may then tell a customer, and the fix that removes the stale source.

Page problemWhat a crawler seesWhat an AI answer may sayFix
Undated rate or fee pageA figure with no “as at” dateQuotes it as current, months after it changedVisible “rates as at” date, a matching modified date, and a logged monthly check
Withdrawn product left liveA complete, indexable product pageRecommends a product that is no longer offeredRedirect it, or mark it clearly as closed to new customers and remove it from sitemaps
JavaScript calculatorAn empty container with no figuresFalls back on third-party or older figuresServer-rendered summary of current rates, eligibility and a worked example beside the calculator
Eligibility rules only in a PDFA dense document, sometimes an old versionWrong age, income or residency rulesEligibility listed in HTML with a date; one current PDF, older versions removed
Same figure in several placesTwo different numbers for one productPicks one at random, or blends themOne source of truth that feeds every page where the figure appears
Old announcements and blog postsA launch rate inside news copyRepeats the launch rate as today’sA dated note that the figure was correct at publication, linking to the current rates page

How does monthly dated verification reduce stale citations?

It gives every public rate, fee and eligibility rule an owner, a check date and a visible “as at” stamp. Each month, and whenever a product changes, someone confirms each figure against the system of record, updates the page and its date together, and logs the check. Engines then find a current, dated statement, and the firm has evidence of what was checked.

  1. Build a register of public figures. List every rate, APR, fee, limit and eligibility rule your firm publishes, and every URL where each one appears, including FAQs, help pages and PDFs.

  2. Feed pages from one source. Where possible, pull figures from the product system rather than typing them into each page, so one change updates every page at once.

  3. Show the date, and mean it. Add a clearly labelled “as at” or “last updated” date. Google’s guidance on page dates (updated 10 December 2025) asks for a visible date that matches the page’s structured data and reflects a real update.

  4. Check monthly and on every change. Run the full register once a month, and check the affected pages the same day a product, rate or rule changes. Record who checked, when and against what.

  5. Retire products properly. When a product closes, redirect or clearly mark its page, remove old PDFs, update sitemaps and add a dated note to announcements that quoted it.

  6. Ask the engines what they say. Each month, put your customers’ most common rate, fee and eligibility questions to ChatGPT, Perplexity and Google AI Overviews, record the answers, and trace any wrong figure back to its source.

Freshness also affects which pages get cited. Ahrefs’ analysis of around 17 million citations, published on 28 July 2025, found that AI assistants cited content 25.7% fresher than Google’s organic results, with Google AI Overviews the exception, citing slightly older pages (Ahrefs, 28 July 2025). A dated, recently verified page is more likely to be the version an assistant picks up, though no engine promises to prefer it.

What can’t a firm control, and what should it do about other sites?

A firm cannot control which sources an AI engine uses or how it words an answer. McKinsey’s analysis puts a brand’s own sites at 5 to 10% of the sources AI search references, so comparison sites, publishers and forums shape most answers. The practical response is to fix your own pages first, then ask the sites that publish your figures to correct or date them.

  • Comparison and rate-listing sites. Send corrected figures, with dates, whenever a product changes, and keep a record of what was sent and when.
  • Introducers and affiliates. Where they publish your rates as part of your distribution, your agreements can require them to update or remove figures within a set time.
  • Press coverage. Old articles stay online. You can’t edit them, but a current, dated page of your own gives engines a better source to cite instead. McKinsey’s figure comes from New front door to the internet, 16 October 2025.
  • AI tools themselves. Most assistants offer a feedback button on answers. It is worth using for serious errors, but it is not a reliable fix and should not replace correcting the source.

What evidence should a firm keep for the Consumer Duty?

Keep enough to show that public figures were checked, that problems were found and that they were fixed. That means the register of public figures, the monthly check log, a record of what AI engines said and where wrong answers came from, and the changes made as a result. Together these show monitoring that leads to action, not monitoring for its own sake.

The FCA’s March 2026 review made the same point in general terms: good practice included documenting what changed, why and with what effect, and firms were expected to act on what their monitoring revealed (FCA, 13 March 2026). Under PRIN 2A.8, a firm’s board must review at least once a year whether the firm is delivering good outcomes for retail customers. A short section on public figures and AI answers gives that report something concrete to point to.

Keep the record simple. A spreadsheet with the figure, the pages it appears on, the date checked, the person who checked it, the AI answers recorded that month and any fix made is enough for most firms. What matters is that it is kept every month, not only before a review.