Methodology: how the library is built

We follow publishers directly. Each source is a place a publisher releases its work:

• Report, research and investor pages on publishers’ own websites
• RSS and Atom feeds of newsletters, analysts and trade press
• SEC filings of publicly listed game companies
• YouTube channels of industry podcasts, talks and interviews
• A small number of shared folders of industry reports

Some publishers hand out a report only after you fill in a form. We fill it in like any reader, with an address on our own domain, and import the PDF they send. Signed-in readers can also upload a PDF; an administrator reviews every upload before it is processed and again before it is published.

Only from the publisher. We do not take copies from re-upload sites such as Scribd, SlideShare or FlipHTML5, and pages that are not documents — homepages, social media posts, sign-up and about pages — are rejected on sight.

Screening. New items from sources we trust are imported directly. Items from other sources are screened first: an AI model judges whether each one is an original report or article, a second, stronger model confirms every approval, and anything neither can decide waits for a person.

No duplicates. Before a document is added it is checked against the library: the same address, the same file, or a near-identical title from the same publisher and year. A match is not imported again.

• For a PDF: a copy of the publisher’s file — the one you open in the viewer and download — and its text, extracted page by page so every passage keeps its page number.
• For an article: its text, with a link to the publisher’s own page.
• For a video: the YouTube video, played from YouTube, and its captions as a transcript, one paragraph per minute.
• Attribution: every document is credited to its publisher, with a publisher page listing all their documents in the library.

We use AI models (large language models, mostly Google Gemini models reached through OpenRouter) for the work below. All of it is generated from the document’s own text:

• Titles, topic tags, glossary terms and the match to the right publisher
• The summary of each document
• Key insights, each linked to the page it draws on
• Slides: the key insights as a short deck, each slide showing its source page or chart
• English translations of documents in other languages, page by page, labelled as machine translation and shown next to the original
• Values read from charts in a page image, marked as read by AI
• Answers to questions in Ask AI, written from passages found in the library

Search ranks results with Jina AI’s embedding and reranking models. AI does not change the documents themselves: the PDF you open is the publisher’s file, and search, AI answers and MCP read the original text, not the translations.

Every passage in the library keeps the page it came from — for a video, the minute. AI answers cite the document and page for each claim, and a citation opens the original at that page; a cited video plays from the second the quote is spoken. When an answer lists figures, each one is checked against the passage it cites, and a figure that cannot be found there is flagged. Where sources disagree, the answer shows both instead of picking one.

Every document page gives a ready-made citation — APA, MLA, Chicago, Harvard, IEEE or Vancouver, or an export to BibTeX and RIS — naming the publisher as author.

• AI summaries, insights, slides, translations and answers can contain errors: a misread number, a missing qualifier, the wrong emphasis. Check the original page before you quote a figure.
• Chart values read by AI are approximate; the chart in the PDF is the source.
• Video transcripts come from YouTube’s automatic captions, so names and numbers can be misheard.
• Text extraction can garble complex tables and multi-column layouts.
• A forecast or estimate is the publisher’s, not ours, and is only as good as their method.
• The library is not complete: it holds what the sources we follow publish, and some reports stay behind a paywall.

Found a wrong summary, a misattributed publisher, a broken file or a document that should not be here? Email [email protected] with the link to the document and, if you can, the page. Publishers who want a document corrected or removed can write to the same address.

Each source is checked daily, weekly or monthly, depending on how often it publishes, and the discovery run goes through the due sources several times a day. New documents are processed every half hour and published as soon as their text and metadata are in place. 189 documents in the library were published in the last 30 days.