Content Pruning: What To Delete And What To Keep

Every site over a few years old carries dead weight: pages published for a campaign nobody remembers, three near-identical posts on the same topic, and a long tail of thin articles written to hit a monthly quota.

The instinct is to leave them alone, because deleting feels risky and they are not obviously hurting anything. The instinct is wrong in one specific way. Duplicate coverage of the same topic splits signals across pages so that none of them is the definitive answer, and in a world where an assistant retrieves three to ten sources rather than ranking ten thousand, being the fourth-best page on your own site is worth nothing.

The Four Outcomes

Every page gets exactly one of these. Refusing to choose is how audits stall.

Keep and improve. Real demand, real topic, page is just weak.

Merge. Two or more pages covering the same intent. Combine into the strongest URL, move the unique parts across, redirect the others.

Redirect. No unique value, but a closely related page exists. Redirect to it.

Delete. No demand, no unique value, no close relative. Return 410 or 404 and remove it from the sitemap.

The most common mistake is redirecting everything to the homepage to preserve some imagined equity. That is a soft 404 in practice, it confuses users, and it is worse than deleting.

The Data You Need Before Deciding

Four columns, one row per URL. Impressions over twelve months. Clicks over twelve months. Whether any conversion or signup has ever been attributed to it. Last meaningful update.

Then add one judgement column that no tool can fill in: does this page contain anything only we could have written? A page with zero traffic that documents your own pricing history or a real incident postmortem is a keep. A page with 200 impressions that summarises a public API nobody asked you to summarise is a delete.

Twelve months matters because seasonal pages look dead for eleven of them.

The Cases Where Deleting Is Wrong

Legal, policy, and compliance pages. They exist to be findable when someone needs them, not to earn traffic.

Pages with backlinks. Redirect rather than delete, even if traffic is zero.

Documentation for features still in use. Support volume is the metric, not impressions.

Anything filed as a record. Incident notices, changelogs, deprecation announcements. Removing them looks like concealment, and in regulated contexts it can be.

If you have published an API deprecation notice, that page stays up long after the API is gone.

Merging Is Where The Writing Happens

The merge is the outcome with actual craft in it, and it is where most audits go wrong. Concatenating three thin posts produces one long thin post.

Before:

This post covers the basics of quantization. Quantization is a technique used in machine learning. In our previous article, we discussed model formats. Quantization can offer a number of benefits for users running models locally.

After:

Four-bit quantization cuts a 4B model from roughly 8 GB to 2.4 GB on disk and about 3 GB resident, at a small and measurable quality cost on writing tasks. That trade is why desktop tools ship Q4_K_M by default rather than the full-precision weights.

The merged page should be shorter than the sum of its parts and denser than any of them. If it is longer, you concatenated instead of merging.

A Wrivio Context for this could say:

Rewrite this merged draft to remove repetition. Delete any sentence that restates a point already made, references a previous article without adding information, or announces what the section will cover. Keep every figure, model name, format name, and date exactly as written. Do not add explanations, benefits, or comparisons that are not in the original.

Press Ctrl+Shift+Space, paste one section at a time, and read the diff. Watch for a deletion that removed the only place a qualifier appeared, which is the specific risk when compressing overlapping sources.

Sequence It So You Can Undo It

Export the current state first: URLs, titles, and full text. Then work in batches of twenty to fifty, not five hundred. Ship the redirects with the deletions in the same release. Wait four to six weeks and look at the aggregate before doing the next batch.

Expect total impressions to fall. That is arithmetic, not failure. The numbers to watch are clicks, conversions, and whether the surviving pages moved up. If those are flat or better while impressions dropped, the prune worked.

Do not run a prune in the same month as a redesign or a migration, because you will not be able to attribute anything afterwards. And if traffic moves in a way you did not expect, check the ordinary explanations before assuming your prune caused it, as covered in why your traffic dropped without an algorithm update.

The Version Of This That Is Now Urgent

If your archive contains a large batch of pages produced quickly to a template, this stops being housekeeping. Google’s spam policies name scaled content abuse specifically, and a thousand thin pages is a different kind of exposure from a hundred old ones. What is actually banned is worth reading before deciding how fast to move.

Common Questions

Does deleting pages hurt SEO?

Deleting pages that earn nothing and duplicate other pages generally helps, because it consolidates signals. Deleting pages with backlinks or legal function hurts, which is why redirecting is a separate outcome from deleting.

Should I use 404 or 410?

Use 410 when you are certain the page is gone permanently, since it removes the URL from indexes faster. Use 404 when you are not certain, and a redirect when a genuinely equivalent page exists.

How many pages should I prune at once?

Twenty to fifty per batch, with four to six weeks between batches, so that you can attribute the effect. Large single-shot prunes make diagnosis impossible.

What if a page has traffic but no conversions?

Keep it and check whether it is a definitional page doing its job as a citation source. Not every page needs to convert, but every page needs a stated purpose.

Download Wrivio for Windows to compress merged drafts down to what each source actually contributed, with the diff showing what was cut.