So let’s open one up.

An EPUB is a website wearing a trench coat

I downloaded a copy of Pride and Prejudice from Project Gutenberg and unzipped it. That’s not a metaphor — an EPUB file is literally a ZIP archive. The very first thing inside is a file called `mimetype` whose entire contents are the string `application/epub+zip`. After that: 24 files, including 17 HTML chapters, one CSS stylesheet, a cover image, and a small XML manifest that says which file is which and what order to read them in.

The numbers tell the rest. Pride and Prejudice is 130,207 words. The whole book — with its styles and cover — is 609 KB. Smaller than one photo from your phone. That’s what you get when a format stores text as text and lets the reading device decide everything about presentation.

That single design decision is the entire personality of EPUB. The file doesn’t know what a “page” is. Your reader paginates on the fly: bigger font, fewer words per page; night mode, different colors; dyslexia-friendly typeface, re-laid-out instantly. When people say an ebook “reflows,” this is the machinery underneath — the same HTML and CSS your browser handles, repaginated for whatever screen and eyes are in front of it.

PDF is a photograph of a book

A PDF is the opposite bet: pages exist, permanently. Every glyph is pinned to coordinates on a fixed-size plate. Nothing reflows because nothing is supposed to — the layout you approved is the layout everyone gets.

For the things PDF was designed for, this is exactly right. Textbooks with two-column layouts and marginal notes. Sheet music. Anything where “page 47” must mean the same thing to every reader — citing, annotating side-by-side, printing. A cookbook PDF on a tablet stand in the kitchen is ergonomically perfect.

For novels, it’s the wrong tool. A 130,000-word novel as a PDF means the type size was baked in by whoever made the file. Read it on a phone and you’re pinch-zooming across a frozen page, or squinting at a full page shrunk to six centimeters. Every accommodation EPUB gives you for free — font size, line spacing, contrast — becomes a fight.

MOBI and AZW3: Kindle’s two dialects

MOBI is the old Palm-era format Amazon bought and rode to dominance. It did the job, with the typography of its era: limited font handling, no great support for fancy layout. AZW3 (also called KF8) is Amazon’s replacement — same lineage, modern typography, better formatting. Kindle devices read both natively.

Here’s the folklore-free update: since 2022, Amazon’s own Send to Kindle service **accepts EPUB directly** and stopped accepting MOBI for new sends. The practical translation: Amazon spent years making EPUB the interchange format so you don’t have to think about their internal formats at all. You hand the ecosystem an EPUB; the Kindle side happens by itself.

MOBI files you already own still open on Kindles. But if you’re re-organizing a library today, storing your masters as EPUB (and converting to PDF on the occasions you need fixed pages) is the arrangement with no dead ends in it.

So which one, when?

- **Reading novels and prose on any device** → EPUB. It’s the format every non-Kindle reader speaks natively, and Kindle accepts it now too. — **Keeping a personal archive** → EPUB masters. Open, reflowable, tiny, outlives any single vendor. — **Textbooks, sheet music, board-game rules, anything layout-dependent** → PDF, and make sure it was born PDF, not converted from prose. — **Kindle ecosystem, new purchases** → whatever Amazon ships you (AZW3/KFX internally); don’t fight it. — **Sharing a document where you need page-stable context** — a chapter for a reading group, a fixed snapshot of a report → convert EPUB to PDF at the moment of sharing, and say “page 4” with confidence. This is the one conversion I do regularly: a browser tab like iLoveConvert’s EPUB to PDF converter, upload, download, no sign-up, and uploads auto-delete within 24 hours — which is the right default for anything in your library.

One honest boundary: DRM

None of the above works on books with DRM — the encryption layer publishers wrap around store purchases. Legitimate converters, online or offline, won’t touch a DRM-locked file, and that’s not a technical shortfall to engineer around; it’s the legal shape of the thing. The practical move is upstream: when a title is sold DRM-free (most indie presses, standard fiction from several major retailers, and essentially the entire public domain via Project Gutenberg and Standard Ebooks), buy that version. Then every format in this article remains a choice you can make later, rather than a door that closed at purchase.

The one-line version

EPUB for reading, PDF for pages that must not move, Kindle formats are Amazon’s private business that now speaks EPUB at the door. Keep masters open, convert at the moment of need — and the format war quietly stops being your problem.

I write practical guides about file formats and the boring decisions around them — previously: where your uploaded files actually go, and what the AI era does to them.

I thought that was the whole story. It isn’t anymore. The AI era didn’t just add new tools — it chained them together, and every link in the chain is another place your file goes.

One upload, four destinations

Watch what a normal person does with a sensitive document in 2026.

They drag a PDF into an AI chat and say “summarize this.” Then they paste a spreadsheet into a different assistant to “find the anomalies.” Then they hand a contract to an AI agent that’s supposed to extract the deadlines — and the agent, being an agent, calls three external tools to get the job done: a parser, a converter, maybe an OCR service.

From the user’s chair, that’s one seamless experience. From the file’s chair, it’s a world tour:

1. The chat history. Most AI assistants keep your conversation — file included or its extracted text — for months by default. Your “quick question about this NDA” now outlives the NDA’s relevance by a year. Some vendors let you opt out of retention; the opt-out is usually a settings page three clicks deep, off by default.
2. The training corpus — maybe. “We may use your interactions to improve our services” is the default posture for consumer tiers. Whether your contract actually tunes a model is unknowable from outside, and that’s precisely the problem: the honest answer varies by vendor, plan, and a checkbox you probably never saw.
3. The vector store. When a tool offers “chat with your documents,” it typically chunks them into embeddings and stores those. Embeddings are weird artifacts — technically derived data, practically a compressed fingerprint of your file, sitting in a database optimized for never forgetting.
4. The tool chain. Agents and “extended” workflows call other companies’s services mid-task. Each hop is another terms-of-service you never read, another retention policy, another bucket.”We take privacy seriously” is not a retention policy

The marketing has escalated faster than the disclosures. Every AI product page now has a privacy badge, a “your data stays yours” callout, a trust-center link. What almost none of them state in plain language is the only thing that matters:

After I upload a file, how long does it exist on your infrastructure, in all copies, in all backups, and who can read it?

The answers I’ve found in months of prodding tools fall into three tiers. Some vendors publish a number — “files are deleted within 30 days,” or if you’re lucky, 24 hours. Some vendors publish a maze: a support article referencing a policy referencing a help-center page that contradicts the FAQ. And some vendors publish nothing at all, which is an answer too. Silence plus free tier equals assume forever.

The uncomfortable part: the tiers don’t track company size. I’ve seen two-person startups with crisp, verifiable deletion claims and household-name platforms whose actual behavior required a support ticket to discover.The proxy that actually works

You can’t audit a vendor’s storage from outside. So I use a proxy, and after a year of testing tools it hasn’t failed me:

Look at what the tool says about deletion before it asks you to pay.

A service that says “files are auto-deleted within X hours” on the page where you upload — not in a footnote, not behind “Learn more” — has made a falsifiable claim in front of witnesses. That’s a company that thought about the lifecycle and decided to be accountable for it. A service whose upload page is all speed and fonts and drag-drop affordances, with retention mentioned nowhere, has also made a decision. It just made it quietly.

This is the same test I apply to plain file converters, and it’s why my shortlist barely changes even as the tools around it got “AI-powered.” The AI layer is presentation. The file’s journey is still: upload, process, store-or-delete. The AI era added destinations, not a new physics.

A five-minute habit for the AI age

You don’t need to abandon AI tools — I use them daily, including for grinding through document chores. You need one habit: before feeding a file to any new tool, converter, chat, or agent, spend five minutes on three checks.

Check what it keeps. Search the site for “delete,” “retention,” “24 hours.” One minute. If nothing surfaces, treat everything you upload as permanent and public-to-the-vendor, and act accordingly — which for most files means “fine,” and for contracts, IDs, and financials means “not here.”

Check the plan you’re on. Consumer free tiers carry the widest data rights for the vendor; enterprise seats usually carry the narrowest. The same product with your same upload can have a different lifecycle depending on a dropdown you clicked in a hurry.

Prefer tools that state their lifecycle where you upload. Vote with attention. When a converter tells you on its own homepage that files are deleted within a day, that transparency is worth something independent of the feature list — it’s the culture that everything else grows from. One example I keep handy is iLoveConvert, whose converter pages state the deletion window up front, which is rarer than it should be.

None of this is paranoia. Files are the most durable thing on the internet — outliving accounts, companies, and at least two of the services I tested last year. The tools will keep getting smarter about reading your documents. The least they can do is be equally smart about forgetting them.

It’s 11 p.m. You have a PDF of a signed contract that your client needs as a Word file, tonight. You search “pdf to word,” click the first result, and drop the file into the big dotted box on the page.

The conversion takes forty seconds. You download the result, attach it to an email, and close the tab.

You just handed a legally binding document — signatures, addresses, bank details and all — to a server you know nothing about. And if you’re like most people, the thought never crossed your mind.

I spend an unhealthy amount of time testing online file converters. Not because I love converters — because I kept noticing how differently they treat the same file once you hand it over. And the differences that matter most have surprisingly little to do with conversion quality.

The part of the pipeline nobody markets

Every online converter works the same way at the bottom: your file travels to their server, software on that server does the conversion, and the result travels back. That’s not a scandal — it’s just how the model works. Your file has to be somewhere else for a moment.

The real question is what happens after the download. And this is where converter sites quietly sort themselves into two very different species.

One species deletes your file within hours — often 24, sometimes less — because storage costs money and keeping files around is pure liability. The other species keeps them. Not necessarily out of malice; usually out of indifference. Files accumulate in a bucket, backups get made, the bucket gets copied to a new region, and three years later a “legacy storage asset” with your 2023 tax return in it is part of an acquisition.

You can’t tell these two species apart from the homepage. Both have friendly drop zones. Both say “fast, free, secure.” The word “secure” on a landing page means exactly nothing — it’s a color scheme, not a policy.

The account wall is a tell

Here’s a pattern I’ve hit dozens of times: the converter works, the download is ready, and then — create a free account to download your file.

Think about what that actually is. Rotating an image or extracting audio doesn’t need an account. There is no technical reason your email address is required to hand you back a file you just uploaded. The account wall exists because an account turns an anonymous file into an attributed file — one that’s now linked to a durable identity, sitting in a durable account, possibly with a version history.

None of this is disclosed as “we want to be able to associate your uploads with you forever.” It’s disclosed as “join our community!”

I no longer use converters that demand registration for a basic conversion. Not because I assume they’re evil, but because the incentive structure is wrong: a product that gates downloads behind sign-up is a product whose business depends on retention — of users, and of files.

(Related tell: the converter that promises “unlimited free conversions” after you sign up. Unlimited and free and server-side don’t coexist without something else paying the bills.)

What “deleted after 24 hours” is actually worth

Some sites state a retention window: files are deleted automatically, typically within a day. When I first started paying attention to this, 24 hours sounded arbitrary. It isn’t. It’s roughly the sweet spot where:

• The window is long enough for slow downloads, retries, and time-zone-difference support tickets.
• The window is short enough that a breach of that storage exposes yesterday’s traffic, not the last three years of it.

You can’t verify deletion from the outside — that’s the honest caveat. When a site says files are removed after 24 hours, you’re trusting a sentence in a help article. But trusting a stated, specific policy is still categorically better than the default, which is no stated policy at all — because “no policy” in practice means “kept until someone decides otherwise,” and that someone is never you.

A retention promise you can’t audit is weak. A retention promise that doesn’t exist is a decision that’s already been made about your files.

A five-minute audit before your next upload

This is the checklist I run before dropping anything sensitive into a converter. It costs five minutes and has sent me walking from more sites than any conversion-speed benchmark ever has.

1. Does it demand an account for a one-off conversion? If yes, leave. Life is short.
2. Is there a stated deletion window, in hours or days? “We take privacy seriously” doesn’t count. A number counts.
3. Is the entire flow HTTPS? Not just the homepage — the upload endpoint too. In 2026 this should be table stakes; occasionally it still isn’t.
4. Does the pricing page make sense? Someone pays for the servers. Ads, subscriptions, or a visible freemium limit — an understandable business model is a good sign. “Everything free forever, just register” is not.
5. Google the site’s name plus “breach” or “scam” before the first upload, not after.

None of this requires paranoia. It requires about the same level of diligence you’d apply before lending your car to a stranger — which is, electronically speaking, what you’re doing.

The boring conclusion

The best converter for your sensitive documents is the one that forgets you fastest. Fast conversion matters, output quality matters, but the file you uploaded stops being your concern the moment it lands on their disk — and what happens in that moment is the entire difference between a utility and a data harvest.

When I test converters now, I read their retention policy before I look at their speed. My short list has gotten shorter over the years — and notably, the ones that survived are the ones built like utilities: no account needed, files auto-deleted, and a paid tier that explains how the lights stay on. iLoveConvert is the one I keep coming back to for exactly that reason.

Your 11 p.m. contract doesn’t need a relationship with a conversion service. It needs to become a .docx, get emailed, and cease to exist everywhere except your machine and your client’s.