This site contains affiliate links. We may earn a commission at no extra cost to you.

Preview Pages Before You Buy: How to Test a Doujinshi You Can’t Read

You cannot read the work before you own it. You can read a few pages of it, for free, right now — and those pages will tell you almost everything you need to know about whether the purchase is going to work out.

Most English-speaking buyers use the preview wrong. They read it as a taster: is this the kind of story I want? That is the wrong question to ask of it, because the tags and the cover already answered it. The question the preview answers better than anything else is whether you can follow the work at all.

This guide covers what the preview actually is, how to run a text-density check on it in under a minute, how to dry-run your translation app on it before you spend money, where the preview systematically misleads you, and what to do when a work has no preview at all.

Two different things get called a preview

They are not the same, and mixing them up is the first place people go wrong.

Sample images (試し読み, tameshiyomi, “trial reading”). These are page images displayed on the product page itself. You look at them in the browser. Nothing is downloaded, nothing is installed, and no account is needed. Most works have them.

A trial version (体験版, taiken-ban). This is a separate free download — an actual file, containing actual pages, that you open the same way you would open the work you paid for. Not every work offers one. Where it exists, the product page has a download button for it alongside the buy button.

There is a third thing that is easy to confuse with both: some works can be read in DLsite’s browser viewer after purchase. That is a delivery format, not a preview. It does not help you decide.

If a trial version exists, use it. It is a larger sample, it arrives at full resolution, and it lets you test your reading setup end to end — unzipping, viewer, translation app — on files that came from the same source as the ones you would be paying for.

If there is no trial version, the sample images are what you have, and they are enough.

The thirty-second density test

Open every sample page. Then do not read them.

That instruction sounds strange, but it is the point. You are not assessing the story. You are measuring how much of each page is covered in Japanese, and you can do that without recognising a single character.

1. Look at coverage, not content. On each page, ask what fraction of the drawing is hidden behind text. Speech bubbles are white shapes sitting on top of art. They are visually obvious from across the room.
2. Find the busiest page in the sample and judge from that one. Not the average. The busiest page is the closest thing you have to the work’s ceiling, and the pages you did not get to see are more likely to resemble it than to resemble the quietest one.
3. Check where the text sits. Text inside clean bubbles is the friendly case. Text written directly over the artwork — thoughts scrawled beside a character’s head, notes crammed into the margins, sound effects drawn into the linework — is the difficult case, and it is difficult in a way no app fixes.
4. Look for handwriting. Typeset text has even strokes and consistent spacing. Handwritten text wobbles. You can see the difference at a glance in a language you do not read, and it matters more than the amount.
5. Look for a page that is all text. Afterword pages, prologue pages and setup pages exist. If one appears in a sample of six pages, there are probably more of them in the other ninety.
6. Count the bubbles on that busiest page. Not for a magic number — for a comparison. Do it on three works and you will have a scale of your own that is worth more than any threshold someone else hands you.

The output of this test is not a score. It is one of three verdicts: the art carries this, I would be translating this, or I cannot tell. The third verdict is a real answer, and the next section is what to do with it.

Dry-run your translation app on the sample, not on the work

This is the step almost nobody takes, and it is the one that saves the most money.

You can screenshot a sample page in your browser and feed it to whatever tool you intend to read with — a phone camera app, a screenshot translator, a manga-specific overlay tool. The sample is a real page from the work, at whatever resolution the storefront serves it. Whatever your tool does to that page is roughly what it will do to the other pages.

Three things to watch while you do it:

Does it find the text at all? Tools skip what they cannot segment. Text drawn into the artwork frequently gets ignored entirely rather than mistranslated, which is worse, because a silent omission looks like a page with nothing on it. If half the visible text on the page never appears in the output, you have your answer.

Is the output a sentence? Japanese speech in manga runs short, fragmentary and full of tone particles that carry meaning no OCR pipeline will reconstruct. A result that reads as a broken half-phrase is normal for one bubble and a bad sign across all of them. Our article on why OCR fails on manga covers the specific failure modes in detail.

How long did one page take? Time it honestly, including the moment where you re-aim because the first attempt caught two bubbles at once. Then multiply. A page that takes ninety seconds, across a forty-page work, is an hour of pointing a phone at a screen. That is the actual cost of the purchase, and it is not on the product page.

The arithmetic is the useful part. Nobody abandons a work because a translation was slightly off. People abandon works because the interruption every four panels stopped it from being reading.

Where the preview misleads you

The sample is real, but it is not a random sample. Four biases are worth knowing, and one of them we confirmed by measurement on our own catalogue.

It is selected to sell. The circle chose which pages you see. They chose the ones that look best, and pages that look best are usually the ones where the art is not buried under dialogue. The bias runs in a single direction: the preview is lighter on text than the work is.

It is usually the opening. Sample pages tend to come from the front of the work. Openings do the setup, and in a lot of works the setup is drawn rather than narrated, with the talking arriving later. When we measured text density across our own catalogue from sample pages, this showed up clearly enough that we now record the sample size next to every judgement rather than presenting it as a fact about the whole work.

Trial page counts get misread as the work’s page count. Product descriptions sometimes state both — a trial version of forty-one pages alongside a main work of ninety-two, for instance. We hit this repeatedly while extracting page counts from descriptions, and the numbers are easy to grab in the wrong order. When a description gives two page counts, the larger one attached to 本編 (honpen, “the main work”) is the one you are buying. Our guide to the Japanese words on a product page lists the other terms that turn up in the same sentence.

The images are resized. Sample images are served smaller than the files you receive. Fine handwriting that is legible at full resolution can be an illegible smudge in the preview — which cuts both ways. It can make a work look less readable than it is, and it can hide a page of dense margin notes that would have been obvious at full size.

None of this makes the preview useless. It makes it a floor rather than an estimate. If the preview looks text-heavy, the work is text-heavy. If the preview looks light, the work is light or heavier — never lighter.

What we do with exactly the same pages

We run this test on every work in the catalogue, mechanically, on the same sample images you can see.

The measurement takes four numbers out of an OCR pass: how many separate text regions are on the page, what share of the page area they cover, how many characters were recognised, and the average confidence of that recognition. We do not keep the recognised text, and the images are written to a temporary location, measured, and deleted.

The reason we look at region count and confidence, rather than character count alone, is that they separate two situations that a simple count treats as identical:

That distinction changed our own catalogue substantially. Judging by character count alone, 71 of 130 works landed in Minimal text — more than half the catalogue wearing the same label, which is useless for choosing. After separating “no text” from “text we could not resolve”, Minimal text fell to 51 and Text-heavy rose from 4 to 33. Those are the figures as of 10 August 2026, and they will move as works are added.

Two decisions behind that are worth stating plainly, because they affect how you should read our labels:

Borderline works get pushed toward Text-heavy, deliberately. Telling someone a work is easy when it is not costs them money and costs us their trust. Telling someone a work is harder than it turned out to be costs a sale. Those errors are not equivalent, so we do not treat them as equivalent.

Every judgement states how many sample pages it came from. A verdict measured from four pages is a weaker claim than one measured from twelve, and you should be able to see which you are getting.

You can browse the results directly: works with an English version, minimal text, OCR OK, and text-heavy. The last of those is not a warning list — it is where a lot of the best writing is. It is just an honest one.

When there is no preview

Some works have no sample images and no trial version. You are not out of options, but the order matters.

1. Check the work type first. A novel is pure text and no tooling saves you. A CG collection skews light. Manga depends on the individual work. This single check eliminates the worst outcomes in one glance — see CG collections, manga and novels for what each format actually contains.
2. Look at the circle’s other works. Circles are consistent in a way publishers are not. If three other releases from the same circle have previews and all three are dialogue-driven, you have your estimate.
3. Check for an English edition before assuming there isn’t one. It is a thirty-second check and it makes the entire question moot. Our guide on how to tell if a doujinshi has an English version covers where that information appears.
4. Read the description length as a weak signal. A long, detailed synopsis usually describes a plot, and plots are delivered in dialogue. A three-line description usually sells a situation, and situations are carried by art. Treat this as a tiebreaker, never as evidence.
5. If you are still unsure, buy the cheapest thing that circle sells. Being wrong about a ¥550 work is information. Being wrong about a ¥2,200 one is an annoyance you will remember.

And if the work is part of a series, start at the first entry. Later entries assume you already know the setup, and “assumes you know the setup” often means more text rather than less — recaps, callbacks, references you have no way to place. Series and reading order covers how to find the first one when the numbering is not in the title.

Summary

• Sample images and trial versions are different things. Samples display in the browser; trial versions are a free download that only some works offer. Use the trial version when it exists
• Judge the preview for text density, not for story. The tags told you what it is about; only the preview tells you how much Japanese is on the page
• Judge from the busiest sample page, not the average one
• Where the text sits matters more than how much there is. Clean bubbles are workable. Handwriting over artwork is not
• Run your translation tool on a sample page before paying, and time one page. Multiply by the page count to get the real cost of the purchase
• The preview is a floor, not an estimate. It is selected to sell, usually drawn from the opening, and served at reduced size — so it runs lighter than the work
• When a description gives two page counts, the one attached to 本編 is the work you are buying; the other is the free trial
• We measure the same sample pages mechanically, record how many pages each judgement came from, and push borderline works toward Text-heavy on purpose

Catalogue figures are as of 10 August 2026 and change as works are added. Storefront features change too — if a product page does not look the way this describes, trust the page.