Disclosure: I build and sell one of the three screenshot tools compared below, a hosted screenshot Actor on Apify. This article was written with the help of AI; the numbers come from the public benchmark data linked at the end, and everything here works in a plain Playwright or Puppeteer script.
page.screenshot({ fullPage: true }) looks done until you look at the output: blank rectangles where lazy images should be, the sticky header stamped down the page, a cookie banner across the middle, and emoji as tofu boxes. Here is what broke when I captured 30 real sites, and what fixed it.
1. Lazy-loaded images: scroll first
Images below the fold load only when they come close to the viewport. Scroll the whole page in steps, wait for pending images, then go back to the top before capturing:
await page.evaluate(async () => {
const step = window.innerHeight;
const max = Math.min(document.documentElement.scrollHeight, 30_000);
for (let y = 0; y < max; y += step) {
window.scrollTo(0, y);
await new Promise((r) => setTimeout(r, 150));
}
window.scrollTo(0, 0);
});
await page
.waitForFunction(() => [...document.images].every((i) => i.complete), null, { timeout: 10_000 })
.catch(() => {}); // a broken image should not fail the whole capture
Cap the scrolling (I stop at 30,000 px) or infinite feeds will scroll forever, and record that you hit the cap, so nobody mistakes a truncated page for a complete one.
2. Sticky and fixed elements
Fixed headers repeat on every slice of a stitched screenshot, and fixed layers get painted in the wrong place. What worked for me:
-
position: sticky→ forceposition: relative. At scroll position 0 the element sits in its natural spot anyway. - A
fixedbar at the top of the viewport → switch it toabsoluteat its current coordinates, so it appears once, at the top. - Other
fixedlayers (chat bubbles, back-to-top buttons, bottom bars, modal backdrops with a hugez-index) → hide them. - Check again right before the capture. Some headers only become
fixedafter you scroll, and some "please disable your ad blocker" modals show up late.
3. Capture speed and 100vh layouts
Chrome can capture beyond the viewport, but on Wikipedia's long Emoji article that took about 14 seconds per slice. So by default I resize the viewport to the page height (in slices of up to 8,000 px) and capture what is on screen.
That breaks on layouts built from 100vh sections: make the viewport 8,000 px tall and a "full-screen" hero becomes 8,000 px tall too. When the page height changes on resize, I fall back to scrolling a normal-size viewport and stitching the slices. Pages where body, not the window, is the scrolling element (getemoji.com is one) need yet another path.
Here is the bottom half of vercel.com, which uses that kind of layout, captured by three tools (left to right: mine, apify/screenshot-url, cheapapi/website-screenshot-api). The middle one repeats the fixed hero over the footer; the right one is blank in the lower half:
4. Emoji and CJK fonts in containers
Slim Docker images have no colour-emoji font, so 😀 becomes ☐, and Japanese or Chinese text turns into boxes. Two Debian packages fix it:
RUN apt-get update \
&& apt-get install -y --no-install-recommends fonts-noto-color-emoji fonts-noto-cjk \
&& fc-cache -f
5. Cookie walls and bot checks
Consent banners can be dismissed: block the consent manager's script, or click its "accept" button before capturing. CAPTCHAs are different. I don't try to bypass them; I detect the challenge page and report an error instead of a picture of it. (Careful: many normal pages load an invisible reCAPTCHA badge. Only treat a visible challenge as a block, or you will flag perfectly good pages. I learned that one this week.)
How three tools did on 30 sites
The same 30 pages (Stripe, The Guardian, BBC, Wikipedia, Vercel, getemoji, Japanese Wikipedia, Spiegel, IKEA and others), one run per tool on 18 Oct 2026:
| mine | apify/screenshot-url | cheapapi | |
|---|---|---|---|
| Complete and visually correct | 24/30 | 19/30 | 18/30 |
| Full page captured (28 reachable pages) | 27/28 | 25/28 | 21/28 |
| Bot-blocked pages (medium.com, unsplash.com) | 2 errors, nothing billed | 1 challenge-page image, 1 no output | 2 challenge-page images |
Caveats: one run each, with an earlier build of mine; 13 of the 30 rows were judged on captured height vs. true page height only, without visual inspection; the apify run was stopped at 29/30 for budget. Still imperfect in mine: booking.com shows grey skeleton placeholders, and The Verge has an empty ad band.
Per-site verdicts, run IDs, heights and the URL list are in the benchmark folder: github.com/tobbytools/website-screenshot-examples/tree/main/benchmark. The repo also has MIT-licensed calling examples for my tool.
Which sites break your screenshots? Drop them in the comments and I'll add them to the next test run.


Top comments (4)
One extra column I’d find useful in this benchmark is cost per usable screenshot. Retries and incomplete captures can make the actual cost quite different from the price per request, especially when you’re working within a testing budget.
I should mention that I work on APIFreaks. If you expand the comparison, our screenshot endpoint is another candidate: apifreaks.com/api/website-screenshot. It would need the same URLs, viewport and capture conditions before drawing any conclusions about how it performs.
Thanks Furqan, that's a really good idea. Price per request hides the cost of retries and broken captures, so I'd compute it as the price per shot divided by the success rate on the test set (e.g. $0.004 per shot at 80% usable works out to $0.005 per usable screenshot). I'll look at adding that column if I expand the benchmark.
Thanks Tobby, that formula makes sense. One small refinement: it assumes every failed attempt is billed. Some tools don’t charge for errors, like yours on the bot-blocked pages. So it might be worth splitting failures into “errored, not billed” and “billed but unusable” (skeleton placeholders, blank lower half, challenge-page images). The second group is what actually drives up the cost per usable screenshot.
If you do expand the benchmark, the 10,000 free credits on signup should easily cover running APIFreaks on the same 30 URLs.
The 100vh fallback is the right call, resizing to full page height breaks more than hero sections once the model gets involved too. A 2540 by 18410 capture lands at 216 by 1568 on older Claude models, body text around 1.4 pixels tall, or 355 by 2576 on Claude 4.7 and later, still only about 2.2 pixels. I build SlimSnap, a screenshot tool for Mac and Windows, and this is the actual failure behind most screenshot-looks-fine-but-the-model-cant-read-it reports. Wrote the full math here: slimsnap.ai/blog/screenshot-downsc..., might fit right after your lazy loading section since they are the same problem at two different stages.