Most sites that have migrated content checked that the redirects resolve and stopped there. That checks the plumbing, not the promise. A 301 can land on the same content at a new address, on a page with half the content, on a page that absorbed ten old ones, or on the home page. Each of those is a different outcome for the reader and for the search engine, and a status code cannot tell them apart.
This post gives you a four-way rubric, preserved, thinned, merged or lost, with a decision rule for each, anchored to Google's own site-move and redirect documentation. Then it gives the version you can run at scale: the eight fields an agent needs to classify every redirect pair, the sampled human check that keeps the agent honest, and what to do with each class. It promises no traffic recovery, ranks no tools, and quotes no numbers from any site.
- 01Google asks for a mapping of each old URL to a corresponding new one, and a permanent server-side redirect for each.Its site-move guidance, last updated August 20, 2026, describes preparing a URL mapping from current URLs to their new format and redirecting from old to new as indicated in the mapping. The rubric tests whether that mapping was honoured in content, not just in status codes.
- 02Four classes cover every redirect pair.Preserved: the same content and intent. Thinned: materially less content or a different intent. Merged: many sources into one destination. Lost: a category page, the home page, or an error in disguise.
- 03Eight fields are enough for an agent to classify a pair.Status chain, final URL, canonical, word-count ratio, heading overlap, title match, internal links in, and indexed state. All are computable from a crawl and your index tooling.
- 04The agent's classification is a hypothesis until a random sample confirms it.Draw a stratified sample across the four classes, have a person label it against the same rules, and publish the agreement rate with the audit. Below your threshold, fix the thresholds before you fix the redirects.
01 — The standardWhat Google's own guidance asks for
Google's site-move documentation makes the mapping the centre of a migration. It tells site owners to prepare a URL mapping from the current URLs to their corresponding new format, to decide where each old URL should redirect, and then to set up server-side permanent redirects from the old URLs to the new ones as indicated in that mapping. It also says each new URL should carry a self-referencing canonical tag, and that internal links should be updated from old URLs to new using the same mapping. Its redirects page adds that a permanent redirect is used as a signal that the target should be canonical, and the status-codes page calls a 301 a strong signal and a 302 a weak one.
Two consequences follow for an audit. First, the unit of work is the pair, old URL to intended new URL, not the redirect rule; a rule that matches a pattern can send a hundred old pages to one destination and still return a 301 for every one. Second, the question for each pair is whether the destination is the page the mapping meant. Google's search-console help describes a soft 404 as a page that returns a user-friendly not-found message without a 404 status; a redirect to an unrelated page is the same failure with an extra hop, and the rubric's lost class exists to catch it.
Google follows the redirect, and Google systems use the redirect as a strong signal that the redirect target should be processed.Google Search Central, HTTP status codes documentation, on the 301 status, last updated February 4, 2026
02 — The rubricThe four classes and their decision rules
Classify every pair into exactly one of four classes. The thresholds are yours to set and to publish with the audit; the ones named here are starting points, not measurements.
Preserved
The destination carries the old page's substance at a new address. Rule: final URL is not shared with other old URLs, word-count ratio near 1.0, heading overlap above your threshold, titles match. Nothing to do beyond updating internal links.
Thinned
The destination is a real page but carries less than the old one did, or answers a different question. Rule: word-count ratio below your threshold, or heading overlap low with a one-to-one destination. The reader lands somewhere plausible and finds less.
Merged
Several old URLs point at one page that is meant to cover them all. Rule: final URL shared by more than one old URL and the destination is a content page, not a hub. Legitimate when the destination actually covers each source; check that, source by source.
Lost
The destination is the home page, a category or tag page, a search page, a 4xx or 5xx, or a page whose canonical points to one of those. Rule: any of the above. The old page's content is gone and the redirect is hiding that fact.
03 — The fieldsThe fields an agent needs
An agent can classify pairs at scale if it is given the fields below for every pair and the rules above as a written decision procedure. The point of listing the fields is that each one is cheap, and each one decides something specific.
| Field | How to get it | What it decides |
|---|---|---|
| Status chain | Follow the old URL and record every hop and code | Lost if any hop is a 4xx or 5xx; flag chains longer than one hop for cleanup |
| Final URL | The last URL in the chain, normalised | Lost if it is the home page or a category page; merged if many old URLs share it |
| Canonical of the final URL | Read the rel=canonical tag on the destination | If it points elsewhere, classify against that page instead |
| Word-count ratio | Destination main-content words divided by the old page's | Thinned below a stated threshold, such as 0.6; preserved near 1.0 |
| Heading overlap | Share of the old page's H2 and H3 text found on the destination | Preserved above a stated threshold; thinned or merged below it |
| Old-page title match | Similarity of the old title to the destination title and H1 | Separates preserved from merged when word counts are similar |
| Internal links in | Count of site pages still linking to the old URL | Any class: those links should be updated to the destination |
| Indexed state | Whether the old and new URLs appear in your search-engine index tooling | An old URL still indexed weeks after the move is a signal the redirect is not being read |
Word counts and headings for the old page need a copy of it. Use the pre-migration crawl if one was kept, the old CMS export, or a web archive capture, and record which source each pair used. A pair with no old copy cannot be classified as preserved or thinned; mark it unclassified rather than guessing, and count how many there are.
04 — The checkThe sample that keeps the agent honest
An agent applying thresholds will be consistently wrong wherever the thresholds are wrong, which is why the classification is a hypothesis until a person has checked a sample. Draw the sample at random within each class, so that the small lost class is represented as well as the large preserved one. Give the reviewer the same four definitions, the old copy and the destination, and nothing else; do not show them the agent's label first. Then compute agreement per class.
Two outcomes are useful. If agreement is high across classes, the audit stands and the sample size and agreement rate go in the report. If it is low in one class, the threshold for that class is wrong, or the field is not capturing what the definition means; fix the rule and rerun before touching a single redirect. Doing it the other way round, fixing redirects on labels nobody checked, is how a migration audit becomes a second migration. The same discipline applies to any agent-run content review, as we describe in our content-audit template.
05 — The actionsWhat to do with each class
The audit is only worth running if each class has an action attached. These follow from the definitions and from Google's guidance that a redirect should lead to the corresponding page.
One more class of content deserves its own pass: documents that were PDFs on the old site and HTML on the new, or the reverse, which fail the word-count and heading fields for format reasons rather than content reasons. We cover that case in our post on publishing research as PDF or HTML. If your site migrated and the redirects were never audited on content, our agentic SEO service runs exactly this rubric, with the sample, as its first deliverable.
06 — Next stepFour classes and a sample turn a redirect list into an audit
Classify every pair, check a sample by hand, then act on the lost class first
Export the redirect mapping, collect the eight fields for every pair, and classify them with the four rules written down. Draw a random sample per class and label it blind. If agreement holds, fix the lost pairs first, because those are the ones sending readers to the wrong place today, then verify the merged pairs source by source. Publish the counts and the agreement rate with the audit.