PDF Analysis Showdown: BDM Ranks ChatGPT, Gemini, Claude, Perplexity, Copilot, DeepSeek, Le Chat

Just two years ago, generative AI tools were promoting unique features to stand out. Currently, their functionalities have become more uniform. But do they truly deliver the same level of performance? To find out, BDM has compared the features of the major tools available on the market.

In this article, we focus on PDF analysis. This useful feature allows users to upload a document and then draft a query to receive various responses: a structured and condensed summary, extraction of key information, a numerical analysis, or answers to specific questions related to its content. This is particularly valuable for professionals who deal with PDFs on a daily basis, provided they use the right tool for the job.

For this test, BDM used three PDFs with different structures, each accompanied by a query to test the capabilities of the analyzed models:

• Apple’s Q3 2025 financial report, which contains only numerical data, accompanied by the following prompt: “Analyze this PDF, extract the key financial figures and provide a reliable analysis of their evolution.”
• The 2025 SELL report on video gaming in France, mainly containing graphs to illustrate data, accompanied by the prompt: “From this PDF, tell me the percentage of players purchasing games in physical format and the top three reasons for their preference for this format.”
• The Digital News Report 2025 by Reuters, on news consumption in France, a dense document (171 pages) mixing text, graphs, and images, all in English, accompanied by the prompt: “Analyze this PDF, provide a comprehensive and numerical summary of the global situation regarding access to information. Then, perform a structured comparison between France and Spain on information access behaviors, including key figures and main differences.”

These three tests will help assess each AI tool’s ability to analyze diverse documents and to extract truly relevant data, whether presented in text or embedded in tables and graphics. We will also evaluate the accuracy and correctness of the information provided. Finally, we will analyze the tools’ ability to understand and organize content, even when the document is dense or segmented into multiple categories.

PDF Analysis in ChatGPT: Our Test

Overall, ChatGPT showed moderate effectiveness in analyzing PDFs, with the quality of its analysis depending on the type of data provided. For the Apple financial report, which contains only numerical data, it efficiently extracted key figures, presented them clearly, and even included percentage changes, enhancing the overall readability. The interpretation of the prompt was generally reliable, although the model sometimes took liberties with the elements highlighted.

The SELL report required more effort: processing time increased slightly, and it made an initial mistake in the results provided about the reasons for buying physical formats, mixing up the third and fourth reasons. Finally, for the longest PDF, ChatGPT understood the prompt well and structured its response correctly. However, the reliability of the information was further shaken, with data mixed and confused from one country to another.

In the end, ChatGPT delivers a fairly convincing performance in form, but still shows limitations on the rigor of the data and requires the user to be particularly vigilant, especially with voluminous PDFs.

PDF Analysis in Gemini: Our Test

Gemini stands out significantly in PDF analysis with its ability to structure information. For the Apple financial report, it immediately offered clear and usable tables, with the option to export data (to Google Sheets), while delivering a pertinent numerical analysis without vague interpretations. The test with the second prompt confirms this efficiency: Gemini responded quickly, provided accurate figures, and precisely indicated the page of the document each time, which is very handy for quickly verifying information.

For the third exercise, Gemini proposed a comprehensive and numerically rich synthesis, including a France/Spain comparison in table format. It’s a shame that the table lacked readability (see the table below) and that some information was omitted due to excessive caution, despite being available in the PDF. However, caution is always preferred over the invention of fanciful figures to provide an answer “at all costs”! Gemini also covered more than requested, at the risk of occasionally digressing.

PDF Analysis in Claude: Our Test

Claude is more surprising in its formatting, often resorting to lengthy bullet lists, sometimes overly enthusiastic and embellished with emojis, thus giving a less professional tone than its competitors and making some cases less convincing.

However, it understands the data to be extracted but takes more liberties with the initial prompt when it deems the practice pertinent. For example, it also cites the fourth reason for purchasing the physical format, considering it close to the third, which is not unwelcome.

In terms of data reliability, the overall content is good, but there are still a few errors to note. However, it is capable of correcting itself when indicated that the data presented is incorrect.

PDF Analysis in Perplexity: Our Test

For each prompt, Perplexity understands what is requested and is capable of extracting the expected information. It also makes some additions that were not originally asked for. Additionally, the reliability of the information is particularly fragile: some data, seemingly simple to retrieve, such as the reasons for purchasing physical formats in the SELL study, are not accurately reported: of the three criteria requested, only one really matches. It also makes mistakes in the comparison between Spain and France, requested in the third prompt.

Regarding the structuring of the response, Perplexity delivers well-formed results, with a correct hierarchy of information and a style that is pleasant to read: clearly its strength. The comparative table between France and Spain on the Reuters report is relevant, and it provides a comprehensive synthesis of the global situation, if one overlooks the errors in the data presented.

PDF Analysis in Copilot: Our Test

In terms of reliability, Copilot is without fault: it extracts and transmits data precisely, whether they come from numerical tables or graphs, on short PDFs. However, extracting data becomes problematic for longer documents: Copilot was simply unable to analyze the Reuters PDF, deeming it “too voluminous to be processed here.” This flaw is particularly disappointing for a tool that is part of a suite meant to assist professionals on a daily basis.

As for the presentation of data and the structuring of the response, Copilot sometimes adds emojis, which can be jarring, especially when they are not always well chosen. For the analysis of Apple’s figures, the task takes more time due to the six separate tables present in the document. Consolidating them into one table would likely have improved the clarity of the analysis.

PDF Analysis in DeepSeek: Our Test

DeepSeek adopts a very structured approach in analyzing PDF documents. It efficiently extracts key data, also indicating where it found these data, but sometimes lacks objectivity in the formulation of its analysis. For financial data, it automatically proposes tables to facilitate reading. It takes more time than its competitors when the document contains graphics. However, the real issue arises with the analysis of more voluminous PDFs. Indeed, the model requires access to an API to handle such a large amount of data and provide a complete response. Thus, it’s impossible to truly evaluate its performance in the same conditions as its competitors.

PDF Analysis in Le Chat: Our Test

Le Chat is less convincing than its competitors in this exercise. While it manages to understand what is expected by extracting the requested data, it takes a significant amount of time to do so. Le Chat also makes numerous mistakes with figures, particularly in voluminous PDFs. It can correct itself when it is pointed out that it has made a mistake. In terms of data presentation, the form also falters: it proposes excessive tables, but does not delve deeply enough into the analysis. Its clear and measured interpretation is appreciated, however, without speculative guesses like some other AI tools. This transparency is welcome, but the initial lack of reliability weighs heavily in the evaluation.

Bonus: NotebookLM, A Promising Tool for Your PDFs

Apart from this AI tool comparison, another platform worth exploring is NotebookLM. Developed by Google, this tool adopts a very different approach from the other tested AIs. Rather than acting as a generalist chatbot, it positions itself as a documentary assistant, designed to explore and query content in a targeted manner. Simply import your PDF and NotebookLM creates a knowledge base that can be navigated through questions posed in natural language.

One of its strengths lies in its ability to locate information exactly where it is in the document, citing corresponding excerpts and indicating their location. An advantage that allows you to verify any advanced figure. Unlike ChatGPT, Gemini, or Claude, NotebookLM does not seek to reformulate or interpret beyond the text: it focuses on reading, searching, and understanding the contextual understanding of documents.

A good practice would be to use a generative AI tool initially to obtain data and a synthesis based on your initial prompt, then verify the extracted information by querying the PDF via NotebookLM.