Section 3: The evidence

AI-findability: testing which anyone can do

At one level, demonstrating the AI-findability problem and the potential solution is straightforward. Simply ask an LLM of your choice about any information that you would expect to be found in a particular company’s report published on the web in PDF format, then see what AI delivers and check its source for the answer.

Next, try the same questions for a company that has already published their annual report in a digital-first format - See examples: Oakley Capital Investments and Airtel Africa.

When we do this, we see a clear difference in AI’s accuracy. Reports produced using a digital-first approach generally give accurate answers sourced directly from the company report. This is far from the case for reports available in PDF-only format, where confidently asserted answers are provided by LLMs. Delve deeper into AI’s responses and these are frequently factually incorrect, out of date and/or from third-party sources the company has no control over.

This of course is not a truly scientific test, but it is a real-world experience all of us have faced. The results from identical requests can be very different depending on which LLM is used and results will often be different if the same questions are asked a week or month later. So it seems reasonable to view evidence from individual testing experiences as interesting and persuasive, but perhaps not enough to be conclusive.

Academic rigour proves that digital-first reporting delivers better AI results

Wouldn’t it be useful if someone had done such testing more scientifically, say over 20,000 times, with tests and methods intended to satisfy academic scrutiny, then published the outcomes?

The good news is that in late 2025 such a report was published. Titled “GenAI as a Reader: How ChatGPT and Co. use Annual Reports”, the authors were drawn from academic institutions in Germany and Austria alongside a European software provider. There were 23 academic testers involved and more than 24,000 results were analysed and considered.

A key finding was that annual reports published online in HTML (the web format) were cited as sources by ChatGPT three times more frequently than PDF annual reports. External third party sources were referenced 2.7 times more frequently where reporting was in PDF-only format, which in the words of the study, “implies a loss of narrative and factual control "  when PDF is the only available format.

So the finding is straightforward. If you want an LLM to find your company information first-hand, it’s best to make the entire report available online in HTML format, i.e. go digital-first.

The study also revealed some interesting findings about who, or rather what, was accessing the digital information and so let’s turn to the word of investment professionals in the next section.