← The Map · Table of contents
Summary
A PDF is not just "a file". At any moment it is in a state: is it a scan or real text? Is it locked? Signed? Large? Does it have bookmarks? Which way up are its pages? Every Documentize app needs the document in a certain state and changes it to another. Once you think this way, ordering stops being a list of rules to memorise: you simply ask, "what state does this app need, and what state will it leave behind?"
Details
The idea in three lines
document in state A ──► [ app: needs A, changes to B ] ──► document in state B
- Apps never modify your original. Each run produces a new file, so you can always step back.
- The danger is not a run that fails loudly. It is a run that succeeds and quietly changes the state (kills the signature, drops the text layer, makes bookmarks point to the wrong pages). The state model exists to help you see those.
The state dimensions
| Dimension |
Values |
How to tell |
Why it matters |
| Health |
healthy / damaged |
It will not open, opens blank or partly |
Everything else assumes a healthy file. Recovery comes first |
| Lock |
open / password-protected |
A prompt asks for a password |
Most apps cannot read a locked file. Unlock it first (PDF on the web) |
| Content type |
scanned (image only) / has text layer / born-digital text |
Try selecting a word in a viewer |
Search, Extract Text, TOC AI and all the AI-understand apps read text, not pictures |
| Orientation |
upright / sideways or upside-down |
Look at it |
Recognition (OCR) is hurt "more than anything else" by wrong orientation |
| Skew |
straight / slightly tilted |
Look at it |
Tilt hurts recognition and looks sloppy; Deskew fixes it |
| Page size |
uniform / mixed; A4, Letter, custom |
Check the page info |
Matters for printing, booklets and tidy packs |
| Forms |
interactive fields / flattened |
Click a field |
Flattened forms cannot be read by Extract Form Data |
| Bookmarks / contents |
none / outline / contents page |
Open the bookmark panel |
Many apps leave them alone; some break their meaning (page numbers shift) |
| Size |
small / large |
File size |
Compress after all edits, not before |
| Signature |
unsigned / signed (intact) / signed (broken) |
Verify Signature |
Any rewrite after signing breaks it |
| Standard |
ordinary / PDF/A-compliant |
PDF/A Validator |
Any step that rewrites the file can change the result |
Three states that cause most trouble
- Scanned vs text. A scan is a stack of pictures. Until it has a text layer, Search (except its scanned-box path), Extract Text, TOC AI, Ask PDF, Checklist, Cheat Sheet, PDF Knowledge Base and Parser's text file see nothing. Searchable is the bridge.
- Signed vs edited. A signature certifies "unchanged since signing". Edit anything and the certification is gone. So signing is last.
- Locked vs everything else. A password stops most apps. Unlock is the first step of nearly every chain that starts with a protected file; Protector is the near-last step of chains that end with sharing.
State-transition table
How to read it: Needs = the state the input should be in. Changes / leaves = what the output looks like compared with the input. "Unchanged" means the chapter says so or the app is read-only. Entries marked expected, not tested are inferred from how the app works, not documented.
Organize
| App |
Needs |
Changes / leaves |
| Merger |
Unlocked inputs (encrypted files are refused); upright pages if scans |
New single PDF; page content carried over as is (PDF + PDF). Office inputs are converted first, so page breaks may differ. Optional OCR adds a text layer to scans. Any signature on an input: expected, not tested to be lost |
| Splitter |
Unlocked; known page numbers |
Many PDFs, lossless pages. Bookmarks and links pointing outside a piece cannot follow it. Combined size may exceed the original |
| Delete Page |
Unlocked; known page numbers |
Fewer pages, lossless, usually smaller. Existing contents page and bookmark page numbers may be wrong afterwards |
| Rotation |
A document with some pages on the side |
Orientation fixed, lossless for 90/180/270. Auto Skew resamples (text layer survival: expected, not tested) |
| Resize |
Mixed or wrong paper size |
Page size unified. Standard mode keeps text as text. Tile mode turns pages into images: file may grow, text only searchable if OCR is ticked |
| Booklet |
Final content at the half-sheet size |
Pages reordered for printing; text stays text. Result is a print layout, not a reading copy |
Add & Structure
| App |
Needs |
Changes / leaves |
| Add TOC |
Final page order; a PDF |
Adds a contents page and bookmarks. Pages not re-rendered. Whether entry numbers refer to pre- or post-insertion pages: verify on the live app |
| TOC AI |
Text layer; unlocked; final page order |
Adds a nested bookmark outline only; nothing visible changes. Varies slightly between runs |
| Add Table |
A PDF plus an .xlsx; final page order (insert position is a page number) |
New table page(s) inserted; later pages shift by that many pages |
| Flatten |
Unlocked; filled-in form |
Form fields become fixed page content (irreversible on that copy). Answers can no longer be read as form data |
| Collage |
Photos (not a PDF) |
A new image-based PDF; no text layer |
| AI Illustrator |
TXT or PDF with real text |
Original text pages plus cover, borders, up to 5 artworks. Output is PDF only |
Protect & Prove
| App |
Needs |
Changes / leaves |
| Protector |
Final, unlocked, unsigned content |
Lock = open password. Page content lossless. Result must not be edited afterwards. Already-encrypted input is not re-encrypted in the quick path |
| Unlock |
The password |
Lock removed, content identical. If the file was signed, unlocking can invalidate the signature |
| Sign |
Final content; a .pfx plus its password; normally no open-password |
Signed. Any later change breaks it. Position of the signature is not selectable |
| Verify Signature |
The file exactly as received |
Read-only. No file produced; verdict on screen |
| PDF/A Validator |
The final file (encrypted files fail) |
Read-only. JSON report; the PDF is never changed |
Shrink & Repair
| App |
Needs |
Changes / leaves |
| Compressor |
All edits finished; unlocked; healthy |
Smaller (images lossy, text stays text). Do not run twice. Breaks a signature that was already there |
| Optimize |
A normal, openable PDF |
Same look, tidier structure; lossless by the site's statement; savings vary |
| Deskew |
Upright, scanned pages |
Skew fixed, but pages are rebuilt as images. Likely drops the text layer; run Searchable afterwards (verify on the live app) |
| Recovery |
One damaged, unlocked PDF |
Health repaired; default policy "Full" also strips metadata, attachments, history and hidden content. Reconstructed pages may differ |
| Unwatermark |
Text or PDF with real text |
Removes invisible AI-text marks only. A PDF comes back as re-typeset plain text: layout, images, links gone; original metadata dropped |
| App |
Needs |
Changes / leaves |
| OCR Image |
Upright JPG/PNG images |
New .txt files; the images are untouched |
| Searchable |
Scanned PDF, upright, unlocked |
Adds an invisible text layer; looks the same; file can grow. Fallback: cleaned pages without the layer if the step fails |
| Extract Text |
Existing text layer |
.txt out; PDF unchanged |
| Extract Image |
Unlocked PDF |
PNG files out; PDF unchanged |
| Extract Metadata |
Any PDF (locked may fail) |
CSV out; read-only |
| Extract Form Data |
Interactive filled form (not flattened, not scanned) |
CSV out; read-only |
| Parser |
Unlocked PDF; text layer for the text file |
ZIP out with text, images, attachments, annotations, bookmarks, form data; read-only |
| Search |
Text layer, or tick "scanned" |
A marked copy (highlights, frames, a "Search results" bookmark tree). Only PDFs with matches are returned. How the new bookmark tree sits beside an existing outline: not documented |
| Global Search |
Nothing |
Not a document tool: finds site pages |
| Resume Analyzer |
A CV (PDF or DOC) |
PDF report; the CV is untouched |
Convert & Create
| App |
Needs |
Changes / leaves |
| Converter |
Unlocked input whose content matches the chosen type |
New file in the target format. Into PDF: whether the result has a text layer depends on the source (expected, not tested). PDF to images: pixels only, no text layer |
| Chat |
Optional PDF with real text |
New plain-text-style PDF (contents page for multi-page documents); edit mode gives _edited.pdf |
| SVG Maker |
An idea or reference image |
SVG plus a PDF holding it as a picture |
AI: understand
| App |
Needs |
Changes / leaves |
| Ask PDF |
Text PDFs |
Nothing in your PDF; answers and a text file on screen. Session kept about a day |
| Abstracts |
PDF (scans are read slowly, but the text is not saved into your PDF) |
Original pages unchanged plus one added Abstract page |
| Cheat Sheet |
Text PDF(s) or a topic |
A new sheet PDF; your PDF untouched |
| Checklist |
Text PDF and a specific topic |
A new checklist PDF and a summary; your PDF untouched |
| PDF Knowledge Base |
Text PDFs |
A wiki (Markdown ZIP); PDFs untouched. Kept about a day |
AI for developers
These do not process documents, so they do not move a document's state at all. They take a repository, package, assembly or URL, and give back reports, files or pull requests. See AI for developers.
Reading the table: a short worked example
Goal: take a crooked, sideways phone scan, make it a searchable, small PDF.
| Step |
App |
State before → after |
| 1 |
Rotation |
sideways → upright |
| 2 |
Deskew |
tilted → straight (text layer likely gone; it is a scan anyway) |
| 3 |
Searchable |
image only → image plus invisible text layer (file grows) |
| 4 |
Compressor |
large → small (text layer stays, since text is lossless) |
Swap steps 2 and 3 and the text layer Deskew is likely to discard takes your OCR work with it. That is the entire point of the model: state tells you the order. The full lab for this chain is Phone scan to 5 MB.