Mathis Lichtenberger wrote in a LinkedIn post[1] that this was the first major improvement after he began working on ChatPDF full time, with more to come. The improvement does not sound like the kind of thing an AI company would put in a headline: putting the PDF and the chat window side by side on the same screen. A ChatPDF official post on the same page added that this lets users look at the source text while asking questions, without switching back and forth between windows.
For a product built on a language model, the first thing it boasts about after going full time is not which version of the model it switched to, but how a screen is arranged. That ordering says more about what it is trying to solve than any feature list.
The trouble with reading PDFs is mostly right here. Suppose you are dealing with a forty-page lease and want to figure out 'under what conditions the deposit can be refunded.' The old way is to search for keywords with Ctrl+F and then read a few paragraphs up and down to judge the context; or paste the whole file into a general-purpose chat box. If it only gives you a passage that reads smoothly but does not point out which part of the source it comes from, you still have to judge for yourself whether it came from the file or from the model's own common sense.
ChatPDF takes over the middle part: drag the file in, type the question in, and the answer comes back together with where it lands in the source text. The official FAQ[2] says it builds a layer of mapping for a document's content and meaning, first finds the relevant passages when a question arrives, and then generates an answer with citations; specific requests are dynamically routed between GPT-4o and GPT-4o-mini. In the side-by-side view and built-in citations shown on the official site, links in answers can jump directly back to the corresponding location in the PDF. These are the company's own statements; the cited material contains no independent accuracy or performance testing.
But there is a defensible judgment in this design: when it comes to 'reading an unfamiliar document,' being able to check an answer matters more than an answer that looks polished. The side-by-side layout is the other half of that judgment. A citation gives you a page number, but for the page number to really save effort, your eyes must not be forced away from the answer.
Image: Source: ChatPDF publisher's product image on Product Hunt[3], retrieved 2026-11-11; original interface preserved.
There is no account barrier to using it. The official site[2] says you can start without registering; the benefit of registering is that history is saved and you can open multi-document conversations. Supported formats include PDF, Word, PowerPoint, Markdown, and plain text; the homepage hero also says file, video, website, but the FAQ does not list video and websites, so the hero cannot be treated as a complete input list.
Multi-document conversations deserve a separate mention. The company says you can create folders and put several PDFs into the same conversation, asking questions, comparing, and finding connections across files. When reading a whole set of contracts, this is probably where it differs most from single-document Q&A: instead of reading them one by one and then putting the pieces together yourself, you let them stay in one room first.
The line between free and paid is usage. The official FAQ[2] says the free plan can analyze 2 documents per day, while ChatPDF Plus has no quantity limit and more advanced features. This quota caps free use at 2 per day; only when you need to handle more than 2 contracts or judgments in a day would you hit the limit.
But a free quota is not a conversion rate. The cited official site does not give ChatPDF's revenue, profit, model and file-storage costs, nor Plus's actual price or billing cycle; none of these can be inferred backward from the free quota. On the same homepage, user scale is written two ways, '10M+ Researchers' and '10M+ users,' with no statistical definition given; this cannot be used as a verified user count.
On security, the company says documents are strictly private by default, transfers use SSL, storage remains encrypted (the storage provider has passed SOC2 Type II), and users can delete them at any time. When you actually need to hand a document to a colleague or client, you can also generate a secure link for a particular PDF; invitees can open it for free without an account, and sharing permissions can be managed and revoked. For this kind of scenario, 'can be revoked' feels more real than 'encrypted': encryption is a promise, revocation is a remote control.
If you are not a reader but want to connect this capability into your own system, the barrier immediately becomes concrete. The backend API[4] can connect this capability into a system, but the division of responsibility changes: the chat API is stateless and will not automatically carry over what was said earlier in the same source. The caller needs to send the relevant history itself in a single request; if page citations are wanted, they must also be explicitly enabled in the request. In other words, the model does not remember what you just asked; the caller's code has to remember.
In another delivery model, 'answers must be able to point back to the source text' is turned into an API delivery that deducts credits per page. PDF.ai's Extract API[5] extracts structured fields from PDFs according to a JSON schema, and the returned results come with citations that link each field to a passage in the source text. What is being sold here is not a chat interface but traceable structured output.
This comparison covers only one shared condition: both require answers to be traceable back to the source text. Beyond that, the ChatPDF side is mainly about the web experience of people conversing with documents, while the PDF.ai Extract API side is about a program extracting structured fields according to a schema. What it can show is that the two delivery paths are different; it cannot be said that the two companies serve only non-overlapping buyers.
Keeping answers from straying away from the source text is the most consistent move in this product. Page-numbered citations, the side-by-side window, revocable sharing links, and the citation switch in the API that must be explicitly turned on all add weight to the same thing: generating an answer is not hard; what is hard is getting people to trust it and to be able to check it themselves with one click. In the post, Mathis says he is fascinated by the possibilities for new UI/UX improvements in this field. The interface has to keep the answer next to the source text so that people are willing to trust it and can verify it with little effort.