AI-assisted cleanup: how I strip Wayback junk across thousands of pages

There is a version of website restoration that is purely manual: open each archived page, delete the Wayback toolbar, fix the broken links, swap the footer year, repeat 3,000 times. It works. It is also why a 3,000-page restore used to take two weeks and cost a fortune - and why, even then, it shipped with mistakes.

I do not do it that way anymore. Agents handle the repetitive cleanup. I handle the validation. Faster. Cheaper. More consistent than a human re-reading the same footer 800 times.

The problem with manual cleanup at scale

A recovered site lands with a predictable set of contaminants baked into every page:

• The Wayback Machine toolbar and its injected scripts
• Archive-internal links (/web/20210101000000*/example.com/…) that point back into the archive instead of to the live site
• Analytics beacons and tracking pixels from the original owner
• The original owner’s contact info, phone numbers, email addresses
• Purchase / “buy this domain” links left over from parking pages
• Footer copyright years from 2014
• Broken image references where the archive never captured the asset

On a 50-page site, a careful person can clean this in a day. On a 3,000-page site, they cannot - not without skipping things. And skipped things become the client’s problem three weeks after delivery, when a customer clicks a footer link and lands on a domain parking page.

Where AI agents fit (and where they do not)

The cleanup work splits cleanly into two categories: mechanical and judgment. Mechanical cleanup is deterministic - remove anything matching this pattern, replace this string with that string. Judgment cleanup is contextual - is this phone number the original owner’s, or a support number the client wants to keep? Is this link an archive artifact, or a legitimate external reference?

My pipeline handles them differently:

Mechanical cleanup: deterministic, page-wide, instant

Archive toolbar removal, archive-link rewriting, footer year update, beacon stripping - these are pattern-based and run across every page in a single pass. No AI needed, no judgment, no risk. A 3,000-page site is clean in minutes, not days.

Judgment cleanup: AI-assisted, human-validated

This is where it gets interesting. AI agents scan each page for things that might need removal - contact details, purchase links, references to the previous business - and produce a flagged list. They do not delete anything autonomously. They surface candidates, and I review the flags before anything changes.

The AI is a second pair of eyes that never tires and reads every page. The decision to remove is still mine.

On a 3,000-page site, this turns a 40-hour manual review into a 2-hour targeted review of flagged items. The same coverage, a fraction of the time, and nothing gets removed by mistake.

What this means for the final site

The combination - deterministic cleanup plus AI-flagged, human-reviewed judgment calls - produces a restored site that is genuinely ready to ship, not “mostly ready, you fix the rest.” Specifically:

• Every page has its footer year set to the current year, not 2014.
• Every link points to a live destination, not into the Wayback Machine.
• No page leaks the previous owner’s contact details or purchase links.
• No page ships with the archive toolbar or injected tracking.
• Broken assets are catalogued and flagged before handoff, not discovered by the client’s visitors.

Why this is cheaper, not just faster

The labor in a restore is not the capture. Capture is automated. The labor is cleanup and validation. When cleanup is automated and flags go to a human, a 3,000-page pass stops being two weeks of footer-hunting. That is why the job can ship in days, not a sprint of unpaid re-reads.

The AI does not replace the person. It removes the part of the work that was never really adding value - the part where a human re-reads the same footer for the 800th time looking for the same archive link. What is left is the part where a human actually matters: deciding what to keep, validating the result, and making sure the site ships clean.

The honest caveat

AI-assisted cleanup is not magic. It is a tool that makes coverage practical at scale. It does not recover pages the archive never captured, it does not invent missing images, and it does not fix structural problems with the original site. What it does is ensure that everything the archive did capture arrives in your WordPress install clean, consistent, and editable - without the weeks of manual labor that used to price small clients out of a proper restore.

If you want to see what real cleanup looks like, send the domain and the archive link on Telegram @wpservicelab. I will show you a sample of the flagged-cleanup output before you commit.