Update, Rewrite, or Delete? A Decision Rule for an Existing Content Library

Almost every piece of AEO guidance is written for the next thing you publish. Structure the answer plainly, add original data, cover the decision rather than the keyword.

That is useful and it addresses maybe two per cent of a working content team’s actual surface area. The other ninety-eight per cent is a library — four hundred posts accumulated over six years, written by people who have left, against a strategy nobody remembers, most of it never touched again.

The real question is not what to write. It is what to do with all of that, and the answer for any given page is one of three things.

Why “just refresh everything” fails

The default response is a refresh programme: work through the library updating dates and tightening copy. It is the most common content initiative in existence and it is close to worthless here, for a reason worth being precise about.

Answer engines select sources on what a page uniquely contributes — information gain. Refreshing a page that says what forty other pages say produces a fresher page that still says what forty other pages say. Its substitutability is unchanged, which means its odds of being cited are unchanged. You have spent a quarter and moved nothing.

Worse, a large library of thin pages is a site-level liability rather than a neutral one. Retrieval has more chances to surface something forgettable with your name on it, and your average quality is what shapes an engine’s disposition toward your domain. AI slop is an AEO liability makes this argument about generated content; it applies identically to human-written filler that has aged.

So the triage has to be able to output delete, or it is not a triage.

The decision rule

Two questions, asked in order, and the order matters.

First: does this page contain anything only we could have published? Original data, first-hand experience, a customer result, a proprietary aggregate, a genuinely distinctive argument. Not “is it good” — is it unsubstitutable.

Second: is there demand for what it covers? Search Console impressions, sales conversations, support tickets, a tracked prompt that touches it.

That gives four cells.

	Unique substance	No unique substance
Demand exists	Update — the asset is sound, the packaging may not be	Rewrite — the topic is worth owning and this page does not earn it
No demand	Update lightly or leave — real but nobody is asking	Delete — the largest bucket, and the one nobody acts on

Most libraries are 10% update, 15% rewrite, 15% leave, and 60% delete. The last figure is uncomfortable and it is usually accurate.

What each action actually means

Update is the cheapest and it is mostly a structural job rather than a writing one. The substance is already there; the problem is that it cannot be extracted. Move the direct answer into a plain, self-contained sentence near the top. Break the page so each sub-question sits in its own delimited section — retrieval operates on passages, not documents, so a page whose pricing discussion is woven through a narrative cannot be retrieved for a pricing sub-query. Add the sources and figures you left out. Writing for AI citation covers the mechanics.

Rewrite means the topic deserves a page and this is not it. The honest version is starting over rather than editing — an editing pass over consensus content produces tidier consensus content. The brief for the rewrite has to answer the first question above with something concrete, which usually means going and getting the substance: interviewing a customer, pulling an internal number, running a small test. If you cannot answer it, the page is a delete wearing a rewrite’s clothes.

Delete means removing it and redirecting to the nearest genuinely useful page. Not noindexing, not “archiving” it behind a link nobody follows. Google’s guidance on creating helpful content is explicit that removing unhelpful content can improve how the rest of a site is assessed, and the same logic applies more sharply when a model is forming an impression of what your domain is for.

Deleting content is emotionally hard in a way the other two are not, because someone wrote it and someone approved it. That is not a reason to keep it.

Two things that make the triage tractable

Do it on a sample first. Auditing four hundred pages by hand is a project nobody finishes. Take thirty at random, run the two questions, and see the distribution. If it is 60% delete, you now have a defensible case for a decision that would otherwise sound reckless, and you have it after a morning rather than a quarter.

Let the evidence order the work. Two data sources make this much less subjective than it sounds. Search Console tells you what has demand — including pages with impressions and no clicks, which are frequently rewrite candidates. Citation data tells you which of your pages engines actually reach for, which is usually surprising and is the single best guide to what your citable material looks like. Citation intelligence is where that lives; content gap analysis covers the demand side.

What this does not solve

Being clear about the limits, because content triage gets oversold as a visibility strategy.

Pruning does not create citations. Removing forty weak pages does not make the remaining ones more citable. It reduces a liability and clarifies what your site is about; it adds nothing. The upside is real and it is second-order, and a team that prunes and expects a visibility jump will be disappointed.

There is no published figure for what pruning is worth. The mechanism — site-level quality assessment, dilution of retrieval — is well-supported reasoning rather than a measured effect with a number attached. Anyone quoting you a percentage lift from content pruning has invented it.

Freshness is not the variable. Updating a date changes nothing. Updating the substance changes something. The two get conflated constantly because one is cheap. Content freshness strategy is careful about which kinds of freshness actually matter.

The counter-argument

The reasonable objection: deleting 60% of a library is an enormous, irreversible action based on a two-question heuristic, and libraries carry long-tail traffic that nobody notices until it is gone.

That is right, and it should change how you sequence rather than whether you do it. Three protections: work in batches rather than a single sweep; redirect rather than 404, so any residual value is preserved; and hold a batch for a full measurement cycle before doing the next one, so you find out what you actually lost. If a batch costs you nothing, the next batch is easier to authorise. If it costs you something, you have learned that cheaply.

There is also a real case for leaving low-value pages alone entirely: if your team’s hours are better spent producing one genuinely original piece than pruning fifty forgettable ones, do that instead. Pruning is defensive work, and defensive work should never be the whole programme.

Where to start

1. Sample thirty pages and run the two questions. Record the distribution before deciding anything.
2. Start with rewrites where demand is highest, not deletes. It produces a visible win and builds the case for the rest.
3. Batch the deletes, redirect them, and wait a cycle before the next batch.
4. Re-read your top pages against the extractability test: does the central claim survive being lifted out of its paragraph?

AEO for content marketers covers where this sits in a content operation, and optimizing existing content for AEO is the page-level version of the update path.

The uncomfortable summary: for most teams the highest-value content decision available this quarter is subtractive, and it will not feel like progress until the quarter after.