What AI Agents Read on a Financial Website: llms.txt, Schema and Answer-First Pages

By Yamm Labs · Updated September 2026 · 9 min read

The head of digital at a mid-size private bank types a customer’s question into an AI assistant: what does her bank pay on a one-year fixed deposit? The answer comes back fast and confident. It quotes a rate from a rate-card PDF that was replaced months ago, and it leaves out the line that said the higher rate applies only to senior citizens. Nobody on her team wrote that answer. The assistant assembled it from what it could read on the bank’s own site. That is the problem this post is about: AI agents do not see your website the way your customer does. They read text, structure and labels, and they quote whatever sits closest to the number.

What does an AI agent actually read on a bank’s website?

An agent that visits a financial website is doing something closer to reading a document than looking at a page. It pulls the text, notes which heading each paragraph sits under, reads tables row by row, and checks any structured data the page declares. It does not admire the hero image. It does not click through a carousel. It rarely opens a rate card that exists only as a scanned PDF or an image, and when it does, it has no way to tell whether that file is current.

Three habits of financial websites work against agents. The first is rates published as images or inside downloadable PDFs, because the design team wanted them to look exactly like the print rate card. The second is conditions hidden in accordions, tooltips and footnotes several screens below the headline number. The third is product pages written as brand copy, with the actual answer to the customer’s question buried in the fourth paragraph.

Each habit made sense when the only reader was a person scrolling on a phone. None of them survives a reader that pulls out the first clear sentence containing a number and moves on. Google’s own guidance for site owners makes the underlying point in plain terms: important content should be available in textual form. That rule predates AI answers. It simply matters more now, because the reader who ignores your images is also the one writing the summary your customer sees first.

If you want to know what an agent will say about your fixed deposit, home loan or term plan, read the page with the images, scripts and styling stripped out. Whatever is left, in the order it appears, is what the agent has to work with.

An agent quotes whatever sits closest to the number, so put the condition there.

Is llms.txt worth adding to a financial website?

Yes, but only as a signpost. llms.txt is a proposal, first published in September 2024, for a plain markdown file at the root of a site that tells language models what the site is and which pages matter. The format is simple: a title, a one-paragraph summary, and lists of links with short descriptions. Its own specification describes it as a proposal to standardise, not an adopted standard.

Google is also explicit that you do not need new machine-readable files, AI text files or special markup to appear in its AI Overviews or AI Mode. So treat llms.txt as a low-cost courtesy to the agents that choose to read it, not as a ranking lever.

For a bank, a sensible llms.txt is short. It names the institution and its regulated entity type, points to the product pages that carry current rates and charges, points to the page that lists grievance and customer care routes, and points to the pages where answers to common customer questions live. It should never point to an old PDF rate card. If the file sends an agent to an outdated document, you have made the misquote easier, not harder.

The effort is an afternoon of work and one compliance read. The risk is pointing agents at the wrong pages. That trade is worth making, as long as nobody on the team believes the file does the heavy lifting.

SIGNALNewsletter by Yamm Labs

The next piece, straight to your inbox.

New writing on BFSI brands, B2B and influencer marketing, and AI search.

Email *

Subscribe

Subscription form is not available at the moment

One click to unsubscribe. Privacy policy

Does schema markup help AI agents understand financial products?

Schema is a label, and labels help only when they match the contents. Schema.org has a FinancialProduct type with more specific subtypes for bank accounts, loans and credit, payment cards, and investments or deposits. It includes properties for an annual percentage rate, an interest rate, and a specification of fees and commissions. On paper, that is a clean way to tell an agent that a page describes a home loan, what it costs and who offers it.

In practice, the rule that governs schema is the same one that governs everything else here. Google’s structured data guidelines say the markup must be a true representation of the page content, and that you should not mark up content readers cannot see. For a financial brand, this turns into a hard design rule: never put a rate, a fee or an APR in schema unless the same figure appears as visible text on the page. A number that lives only in the markup is a number your compliance team never reviewed in context.

The useful schema on a bank’s site is usually modest. Organisation markup that states the legal name and the regulated entity, matching the registration details the site already shows as text. Product markup on each product page that matches the visible rate table. FAQ-style question and answer structure on pages that answer real customer questions. Breadcrumbs that tell an agent whether a page is about savings, cards or loans. None of this is exotic. The discipline is keeping it in step with the page every time a rate changes, which means the schema update belongs in the same change ticket as the rate update.

Where should the disclosure sit so an agent quotes it with the number?

This is the regulator-driven decision most competitors skip, and it is the one that decides whether an AI answer about your product is accurate. Indian regulators already require disclosures to travel with the product. RBI requires banks and NBFCs to give borrowers a Key Facts Statement in a standard format for retail and MSME term loans sanctioned on or after 1 October 2024, including the annual percentage rate, and says fees not shown in the KFS cannot be charged without the borrower’s explicit consent.

SEBI requires mutual fund risk to be shown on a riskometer, from low to very high, in fact sheets and promotional material. IRDAI’s current advertising rules require mandatory disclosures to be clear, conspicuous and legible, and bar advertisements that leave out or understate a policy’s important exclusions, limitations and conditions.

SEBI’s own history gives the clearest reason to design for readability. In 2008 it set a minimum on-screen and spoken duration for the standard warning in audio-visual mutual fund advertisements, after finding that rushed delivery made the warning unintelligible. A disclosure that exists but cannot be understood was treated as a failure. An agent is the newest reader that cannot understand a disclosure placed in an image, a collapsed accordion or a footnote three screens away.

So the design decision we make on financial product pages is a fixed sequence, in one block of text, in this order:

1. The answer. The rate, premium or return basis, written as a sentence, not only as a graphic.
2. The condition. Tenure, age band, minimum amount in rupees, or eligibility, in the same paragraph or the next row of the same table.
3. The disclosure. The regulator’s required line, the riskometer level written as words beside the dial, or the APR beside the headline rate.
4. The date and source. When the figure was last updated, and a link to the full terms page.

None of this changes what the regulator asks you to say. It changes where you say it, so that the reader who pulls one paragraph gets the whole truth.

Is the answer-first page a compliance risk?

This is the strongest objection, and it deserves a straight answer. Compliance teams worry that a short answer at the top of a page invites an agent to quote a rate or a benefit without the conditions the regulator requires. They are right that agents clip. A one-sentence answer is exactly what gets lifted into a chat reply.

Our response is not to drop the answer. It is to make the condition part of the answer sentence. Compare a line that says a fixed deposit pays a stated rate with one that says the rate applies to a specific tenure and customer category, from a specific date. The second sentence is longer. It is also the only one that stays true when it is quoted alone.

We concede the cost. Every answer block becomes an approval item, and a product page with eight customer questions carries eight small statements that compliance must sign. The first round is slow. The way through is a template: agree the sentence pattern for each product family once, then approve the variables rather than the prose each time rates change. Brands that do this find the second month faster than the old process, because reviewers stop rereading brand copy that never needed their attention.

There is a quieter risk in doing nothing. If your site does not give a clear, conditioned answer, the agent will still answer the customer’s question. It will just use an aggregator, a forum thread or your own outdated PDF to do it.

What should a financial marketing team do on Monday?

Start with the questions customers already ask, not with the file formats. A useful first week has six steps:

• List ten real questions. Pull them from call centre logs, branch staff and search console queries. What is the FD rate for seniors, what does this card charge abroad, what is the risk level of this fund.
• Test what the page says in text. For each question, open the relevant page with styling off and check whether the answer and its condition exist as readable text.
• Move numbers out of images and PDFs. Rebuild rate cards and charge schedules as HTML tables, and keep the PDF as a download, not the only source.
• Write conditioned answer blocks. One to three sentences per question, with the condition and the disclosure inside the answer, reviewed by compliance as a template.
• Add matching schema. Product and question markup only for figures visible on the page, updated in the same ticket as the rate.
• Publish a short llms.txt last. Point it at the pages you just fixed, never at old documents.

We cover the wider programme on our AI visibility page for BFSI brands. If your team is still sorting out the vocabulary, the GEO and AEO FAQ for financial brands is a good primer, and the 100 AI search optimisation questions go deeper on individual tactics.

People also ask

Do AI assistants read PDFs on bank websites?

Some do, some do not, and none can reliably tell whether a PDF is current. Treat any rate or charge that exists only in a PDF as invisible, and publish it as text on the product page as well.

Does llms.txt replace robots.txt or a sitemap?

No. llms.txt is a proposed plain-text guide for language models, while robots.txt controls crawler access and a sitemap lists your URLs. Keep all three, and make sure llms.txt points only to pages with current figures.

Which schema type fits a bank or insurance product page?

Schema.org’s FinancialProduct type and its subtypes, such as BankAccount, LoanOrCredit, PaymentCard and InvestmentOrDeposit, cover most bank products. Use them only for details that also appear as visible text, and update the markup whenever the rate changes.

How can a bank check what AI agents say about its products?

Write down the questions customers ask most, from call centre logs, branch staff and Search Console, and ask them in the main assistants in logged-out sessions. Compare each answer with the product team’s current rates, charges and conditions, and note which page the assistant cites. Then open that page with styling off to see what text the agent actually had to work with.

Should a financial website block AI agents in robots.txt?

Blocking stops some agents reading your pages, but it does not stop them answering questions about your products. They will draw on aggregators, forums or old copies instead. For most banks and insurers the safer course is to let agents read pages that carry current, conditioned answers, and to decide any blocking with your security, legal and compliance teams.

When should the schema on a rate page be updated?

Every time the visible figure changes, in the same change ticket. Google’s guidelines say structured data must be a true representation of the page content, so markup that still shows last month’s rate after the table has changed is exactly the mismatch to avoid. Tie the schema update to the rate update and one review covers both.

Do answer blocks on product pages need compliance approval?

Yes, because they state rates, conditions and disclosures that AI agents may quote on their own. The way to keep this manageable is a template: agree the sentence pattern for each product family once with compliance, then approve only the variables when rates change. Confirm the current regulatory text with your compliance team when you set the pattern.