Publishing Financial Reports for AI

Challenge: reports may not be read

Analysts and investors increasingly interact with financial reports through AI tools — large language models that ingest, summarise, and answer questions on their behalf. They may never read the original document at all.

Fact is, someone might be in their car, asking their AI assistant on their Apple CarPlay or Android Auto-connected phone about your latest financial report, and simply listen.

This means we need to publish reports in a way that influences AI to generate answers aligning with our intent – the right messaging, tone, and sets of facts.

Solution: dual-audience publishing

We may continue publishing "printed reports" – typically carefully designed PDF documents with imagery, charts, and brand-coherent styling and typography – intended for the human wanting to consume the original document exactly as intended by the company behind it.

While AI tools can consume a PDF – and generally produce accurate answers based on its contents – a PDF gives us little leverage over how AI interprets it.

If we want to keep the legacy-style report, we have to produce two separate versions:

1. visual for humans through HTML (or equivalent) and/or PDF
2. structured text for machines through Markdown

Practical considerations for AEO version

What to avoid

• images and charts – AI interpretation of graphics is unpredictable, use text and tables
• typography – AI doesn't see design, it sees structure

What to prioritise

• concise to-the-point summaries – think 4 pages (roughly 2000 words) of plain language describing every fact, claim, and commitment in the report
• structured text – Markdown allows specifying layout – not design – of text, making it easier for AI to ingest the content as intended

How to produce

1. create the original report as you normally would
2. generate draft of the Markdown version based on the original report (perfect use-case for AI with a custom skill)
3. verify numbers in the reports against XBRL data, if available (also great for AI)

How to publish

1. official report (PDF/XHTML/iXBRL) can be published as-is, e.g. with a link on the website
2. AI report (e.g. Markdown) should be published and linked along the official report for transparency
3. llms.txt file should be published in the site root with links to both report versions, clarifying that while the files contain the same factual content, the Markdown version has been structured for machine readability

Things to consider

• system prompts – AI tool may have built-in instructions to ignore corporate bulls**t, distill content to essentials
• internet context – AI may analyse, and base its answers on, any available content, including web searches and competitor reports
• lost in the middle content – LLMs tend to focus, recall, and reason more on details at the beginning and at the end of the content
• write for the question – AI is more likely to find the "correct" answer if the content is, token-wise, closer to the question asked
• lean on headings – LLMs use headings as structural anchors to locate relevant content, so make them descriptive and specific

How we tested it

We created 4-page Markdown versions for multiple quarterly and annual reports from Fortune 500 and FTSE 100 companies.

We loaded the Markdown versions and the original reports into separate AI context windows, then asked the same questions and compared the answers.

Findings

Answers based on the Markdown versions were consistently more precise, with fewer instances of the AI inferring implicit facts or reasoning on its own.

Two factors likely contribute:

1. the original reports were significantly larger – driven by graphics, extra content, and considerably more text – producing larger context windows with longer token distances between our questions and the most relevant report sections
2. the Markdown versions were editorially condensed: the act of summarising isolates the facts and commitments that matter most, which makes them easier for AI to surface

Our test does not fully separate these two effects. But whichever is the main factor, the practical implication is the same – for AI consumption, a shorter, structured, purpose-written summary outperforms the full report.

Final notes

• regulatory compliance – summaries should reflect the official report (narrative content may be condensed, but statutory disclosures should appear complete and unaltered)

• disclaimer standard – consider including a machine-readable disclaimer at the top of the Markdown file, explicitly pointing to the legally binding original report

• this is an emerging approach – different AIs behave differently, llms.txt is a convention – not a standard, user behaviors and technologies evolve, so expect a shift in way of working rather than a one-time fix