A common problem when using AI for web research is that not all content from a website is extracted. Calendar information is one example. Earlier models often struggled with this, producing incomplete results or even hallucinating events. We therefore tested how well current models – Astra and GPT-5.6 Sol – can handle this type of content.
The aim was to extract information on more than 50 events from a website. The input consisted of the page URL and a prompt asking the model to list all events taking place in September and October 2026.
- GPT-5.6 Sol with medium reasoning initially responded that the calendar data was loaded via JavaScript and therefore not directly accessible. Only in a second step did the model recognise that it needed to use its internal browser functionality to access the content. We could then observe ChatGPT navigating the calendar through a small preview window and extracting the data, partly using computationally intensive image recognition. The entire process took around six minutes.
- Astra with light reasoning received exactly the same request. Without any follow-up questions, separate browser functionality or cumbersome intermediate steps, it returned all events, including the corresponding event links, correctly in under a minute. Astra appeared to be able to process the JavaScript content directly and required neither screenshot analysis nor separate browser control.
What explains such a large difference in performance? Websites cannot all be treated in the same way. Their technical structures vary, and AI models also use different methods behind the scenes. Content that is immediately visible to a human user may still be difficult for a model to access or may require additional steps to retrieve.
As our example shows, the difference came down to the fact that some information was only loaded and displayed after scripts such as JavaScript had been executed. A browser handles this automatically for us, but not every AI model does so in the same way. Choosing the right model for the task is therefore key to efficient and effective research.
By Marcel Ohrenschall