Find broken links and spelling errors on a large site by tracing them to the systems that publish them.
Broken links exist on 21% of government webpages and 23% of news webpages. That’s a staggering number, and it doesn’t even include spelling mistakes. Odds are pretty high that your site has similar issues that so far have gone unnoticed.
To find and fix them, catalog unique link destinations and words, group findings by confidence and declared language, and collapse repeats into one fix per pattern: a redirect, a template’s lang attribute, a dictionary entry.
The method runs in five steps:
- Map the systems that publish to your site.
- Collapse broken links to unique destinations, sorted by confidence.
- Fix the page language before judging the spelling.
- Treat current pages, the archive, and PDFs as separate estates.
- Give each pattern one owner, and retest to confirm the fix.
Summary
- Broken links and spelling errors on large sites often trace back to a handful of publishing systems, templates, document libraries, redirect rules, and dictionaries, rather than individual pages.
- Sorting broken-link reports by unique destination and host turns thousands of findings into a few single repairs, since one dead URL affects every page linking to it.
- Link failures should be tiered into certain, likely, and potential categories, with certain 404s and 500s fixed first and ambiguous ones verified before action.
- Every finding should be resolved through one of four decisions: approve, ignore, fix in the CMS, or redirect, which assigns ownership and clears the backlog faster than exporting raw lists.
Large sites publish errors through a few systems
On a large site, broken links and spelling errors come from a handful of publishing systems, namely templates, the CMS, the document library, redirect rules, and dictionaries. Each system repeats its fault on every page it touches, so an error count in the thousands usually traces back to a few causes you can fix once.
The search starts inside your own site, with moved pages and deleted directories. On government pages, the vast majority of broken links are internal.
A link breaks when a page moves and nothing updates the pages that pointed at it, which is why moving pages during a site migration leaves broken links across the site. Spelling multiplies the same way, since a brand name missing from the dictionary adds an error every time it appears.
Finding errors, on a large site, means locating the pattern.
One dead destination is one job
Sort the broken-link report by destination URL and host. Each dead destination is a single repair, no matter how many pages link to it. Work the certain failures first, give likely failures a second check, and keep potential failures out of the sprint until someone confirms them.
A link inventory and a broken-links check answer different questions. The inventory lists every unique link URL, each occurrence, and every host you link to; the check tests whether those destinations respond. Grouping failures by host goes a level higher: When a partner retires its domain, every link to it fails together, and one replace-or-remove decision covers the set.
Your own dead URLs get a redirect. One redirect repairs every internal page that still points at the old address, along with the external sites and bookmarks still sending people there. Links to pages on other sites need a replacement source or removal, because only the other site’s owner controls that address.
Traffic sets the order. A dead link in the main navigation menu fails on every visit. It also hurts search performance. Google uses links to find new pages to crawl and to judge relevance, so a dead link in the navigation can cut crawlers off from the pages behind it.
Link failures split into three tiers:
- Certain for a clear 404 or 500,
- Likely for timeouts and connection failures, and
- Potential for ambiguous responses.
When a link loads for you but fails in the report, the cause is usually a geo block, bot detection, a login wall, a rate limit, or a host reachable only on the VPN. The check is a logged-out load in a private window, ideally from another network. If the page loads, approve the finding.
That approval protects the report, since phantom failures otherwise train editors to skip the real ones.
Placeholder links (href=”#”) and empty links belong in a separate pile, since they point nowhere by design.
Correct page language turns false typos into clean pages
You’d be surprised how often alleged spelling errors are caused by an incorrect lang attribute on a page.
Spellcheck is only as accurate as the language it assumes. When a page declares the wrong language, every correct word on it becomes a finding. The repair order follows from that: correct the page language first, then the dictionary, and only then the typos a person made.
Checkers read language from the code, the same way assistive technology does. Under Success Criterion 3.1.2 in WCAG 2.2, Language of Parts (Level AA), any passage without its own marking takes the default language of the page, which is why declaring the language of each passage matters to a spellcheck too. A template that hardcodes lang=”en” and serves translated content tests every French or Spanish sentence against English.
Fix that declaration, and thousands of findings disappear at the next scan.
The spelling check sorts findings by kind, because the kind usually points to the cause: likely errors, possible errors, words from another language, and capitalization. When AI spellcheck is on, it reviews unrecognized words to cut false positives such as names and coinages, and it skips pages with too many unresolved misspellings. The dictionary still does its own job.
| What the list shows | What it often is | First move |
| Thousands of misspellings in one section | Wrong declared language on the template | Fix the lang attribute or URL language rule, then retest |
| Words from another language | Content and declared language disagree | Mark the passage’s language in the HTML, or set a language override |
| The same likely error on every page | A brand, product, or organization name | One dictionary entry |
| Capitalization only (“silktide” vs “Silktide”) | A style question | Decide once: a case-sensitive dictionary entry, or leave it |
| Possible errors | Names, coinages, compounds, or regional spellings such as color and colour | Confirm which English the site uses, then review |
| Likely errors in captions, footers, old news | Real typos where nobody rereads | Edit in the CMS |
| A page with too many findings for AI spellcheck to review | Garbled text, binary content, or the wrong language | Diagnose the page first |
Typos cost listeners too, because of how screen readers announce text: “goverment” is read aloud exactly as written.
Sorting spelling findings by section shows where words from another language cluster. Fix that template first, add the names that still repeat, and what survives is the list worth an editor’s time.
Current PDFs get fixed while old press releases stay
A large site is comprised of three estates: current navigation, historical news, and documents. Scan all three, then fix them differently. A dead link in a policy PDF people still download is this quarter’s work, while the same link in a decade-old press release belongs to the record.
| Estate | What to do with findings | Cadence |
| Current navigation and pages | Fix in the CMS or redirect | Quarterly pass on certain failures |
| Historical news and retired PDFs | Approve as published | Decide once |
| Current PDFs | Fix the source file, re-upload under the same URL | Quarterly, with a human check on odd results |
Public archives already work this way. The End of Term archive has preserved federal websites at each administration change in 2008, 2012, 2016, 2020 and 2024. A University of North Carolina library guide notes that because the National Archives hosts those copies as historical material, broken links in them will not be updated. Archivists call that preservation: a record has to stay fixed.
Your own news archive and retired PDFs deserve the same decision. Approving those findings as historical clears them from the queue and leaves what the organization said at the time intact. When a current page cites a source that might move, Harvard Library Innovation Lab’s Perma.cc tool fixes the citation itself by saving a snapshot that keeps referenced links from changing or disappearing.
Current documents get repaired.
A policy PDF or fee schedule among the PDFs people still download carries its dead links and misspellings straight to the reader. These should be fixed in the source document, then re-uploaded under the same URL so existing links keep working.
Link rot keeps going after every cleanup, because pages you link to today will manage to move next year. Sections of an estate can be reported as if they were separate sites, which keeps the current and historical split visible in every scan.
Four decisions turn thousands of findings into tickets
A finding becomes a task only after someone decides what it is: approve it, ignore it, fix it in the CMS, or redirect it. Those decisions clear most of the list on their own. What remains collapses into a few cause-level changes, each with one owner and one change to make.
Exporting the raw list into a tracker skips the deciding step, and that’s where large-site quality programs tend to stall.
Decisions shrink the list fastest. Fixes follow a workflow that routes fixes to owners: content editors own words and CMS links, engineering owns redirects and template language.
Open one ticket per cause. A closed ticket proves someone did something, but only a passing retest proves the fix worked.
A few systems explain the one-in-five figure
Pew’s one-in-five error figure for government pages reads like thousands of separate failures, but on a large site it’s usually a few systems repeating themselves. Catalog destinations and words, sort by confidence, fix the language, and give each cause one fix.
And if you want a tool to help you work through it, Silktide makes a pretty great one (I’m probably biased though so don’t take my word for it – check out our glowing reviews).
Frequently asked questions
How do you spell-check a whole website?
Check each page against the language it declares, then group findings by kind and by section. Fix wrong language declarations on templates first, add recurring brand and product names to an account dictionary, and only then edit the typos left in captions, footers, and old news.
Should you fix every broken link?
Fix dead links in current navigation and current documents, starting with high-traffic paths. Historical material, such as decade-old press releases and retired PDFs, is generally fine as-is.
Do PDFs need broken-link and spelling checks too?
Yes. Current PDFs that contain dead links and typos need to be fixed in the source file, then re-uploaded under the same URL.
References
- Chapekis, A., Bestvater, S., Remy, E., & Rivero, G. (2024, May 17). When online content disappears. Pew Research Center. https://www.pewresearch.org/data-labs/2024/05/17/when-online-content-disappears/
- Content. (2026, October 3). Silktide.Com. https://docs.silktide.com/governance/content
- Finding Alternative Sources for Federal Information & Data: Archives Preserving Websites and Alternative Sources. (n.d.). The University of North Carolina at Chapel Hill Library. Retrieved https://guides.lib.unc.edu/federal-information/website_archives
- SEO link best practices for Google | google search central | documentation. (n.d.). Google for Developers. Retrieved October 26, 2026, from https://developers.google.com/search/docs/crawling-indexing/links-crawlable
- Understanding success criterion 3.1.2: Language of parts. (n.d.). Www.W3.Org. Retrieved October 26, 2026, from https://www.w3.org/WAI/WCAG21/Understanding/language-of-parts.html



