Old Website Content: Delete, Update or Keep?

Companies getting ready for GEO (Generative Engine Optimization) almost always stall on the same question: the website is piled with dozens or hundreds of old articles and product pages, some written years ago — should any of it be deleted?

The conclusion first: don't delete yet. Old content isn't a burden — much of it is an asset that hasn't been activated. Some of it does need dealing with, but the test isn't age, it's fit.

One boundary: "don't delete yet" doesn't mean "keep everything". Below is a three-way sort that separates what to keep, what to fix and what to deal with — run through it once and you're done.

Why old content often still works

Many people assume AI "only reads the newest things". The reality is more nuanced.

One study measured how complete AI's citations are (arXiv:2410.22349, preprint, not peer-reviewed): the sources actually listed make up only 20% to 25% of the material it genuinely read.

Which means: not being listed doesn't mean not being read. A lot of content feeds into AI's judgement without getting that visible slot. Those old articles that look like they get no traffic may still be read — you just can't see it in your analytics.

Scope note: this figure describes how complete the citation display is; it cannot be reversed into "all old content is useful". It's a reminder to be careful before deleting, because you can't see the whole picture.

Three kinds of content, three ways to handle them

Sort your website content by "would a customer ask this?" and the handling becomes clear.

• Category A — on-point: keep and update. This content answers questions customers genuinely ask (how to choose, how to use it, what to watch out for, what the industry standards are). It's the backbone of GEO and worth the time to update. Age doesn't matter — even something written three years ago should stay if the question still holds.
• Category B — neutral: keep, but keep the facts consistent. Company profile, history, awards and qualifications, news. This content doesn't answer buyer questions directly, but it's what AI uses to verify who the company actually is. The condition for keeping it is site-wide consistency — company name, founding date, address, phone number and business scope must not conflict across versions.
• Category C — no value: deal with it (rewrite, merge or remove). Outdated promotion notices, keyword-stuffed shell pages, machine-translated pages that don't read properly, pages that open on "coming soon". This content is a negative asset for AI.

The test is simple: read the first paragraph and ask yourself, "would a customer come to me with this question?" If not, it's Category C.

How do you "update" a Category A page? Not by changing the date

First, a common practice to drop: bumping the publish date to look fresh. It achieves nothing and makes your content messier.

Three things actually need attention.

• Whether the facts are current. Product specs changed, industry standards updated, policy shifted — all of these must be synced. This is the core of updating.
• The density of evidence. The KDD 2024 GEO paper (DOI: 10.1145/3637528.3671900, peer-reviewed) tested it: adding source citations and statistics to content affects how often it gets cited. The same paper found keyword stuffing is ineffective in AI answers, and can even hurt. So don't stuff keywords into old articles — add sourced facts instead.
• How the information is organised. Old content often runs in long unbroken blocks with no subheadings and no hierarchy. One study measured the correlation between structural factors and being cited at 0.63 to 0.68 (UC Berkeley, arXiv:2509.10762, preprint, not peer-reviewed); another measured roughly 22% higher retrieval hits after structural optimisation, while optimising body text alone without structure was negative (Yonsei University, arXiv:2602.12187, preprint, not peer-reviewed). Scope note: these are correlations or relative changes under specific experimental conditions, not promises of results.

Note this is about structure — how information is organised (what comes first, whether concepts are defined, whether it can be layered) — not formatting (font size, colour, borders). Don't mix the two up.

Before deleting anything in Category C, do two checks

Deleting outright is risky, because old content may hold value you can't see. Check two things first.

• Did it ever bring visits? Check your backend or analytics for traffic to that page over the past year. If it has traffic, someone is using it — don't rush to delete.
• Do external links point to it? If other sites link to that page, deleting it creates a broken link — bad for both sides.

If both are zero, it can be dealt with. And "delete" isn't the only option — in priority order.

1. Merge: if it overlaps with another piece, move the informative parts across, then retire the page.
2. Rewrite: if the topic has value but the writing is poor (keyword-stuffed, empty), rewrite it into something that answers the question.
3. De-index: the page can't be deleted yet (historical links, for instance) but genuinely has no value — stop it from being indexed rather than making it vanish.
4. Delete: only when there's no traffic, no backlinks and no value in the topic.

And one more thing: don't let old and new content contradict each other

In handling legacy content, the most dangerous — and least noticed — issue is conflicting facts.

Take one example: an article from three years ago lists the old address, while the Contact page shows the new one. In cross-verification, AI doesn't necessarily trust "the newer one" — it may lower its confidence in both.

Google's official documentation has a hard requirement too: structured data must match the visible text on the page — you can't write one thing in code and another on the page. The same logic sits behind it: inconsistency itself gets penalised.

So while you're taking inventory, pull company name (including English name), founding date, address, phone, email, main business and key milestones into one table and check them page by page. It doesn't take long, and it decides whether everything above cancels itself out.

Three common questions

The content is old but the question still holds — should we rewrite it?

Don't rewrite — patch first. Get the facts accurate, add data and sources, break long blocks into a structure with subheadings. After those three steps, most old content can compete again. Tearing it all down and rewriting actually throws away what's already accumulated.

How much content counts as "enough"?

Not by count — by coverage. The test: of the 20 questions your customers ask most, how many does the website answer head-on. Covering 15 or more beats writing 100 generic industry articles.

Will deleting content affect the whole site?

Not if you run the two checks in the fourth section. The risk comes from bulk mis-deletion — removing a large number of pages that have traffic or backlinks in one go makes the site structure lurch. The steady way: handle one batch, watch for a while, then continue.

In closing

To sum up: old content isn't something to clear out — it's something to sort.

Keep and update what's on-point, keep and align what's neutral, deal with what has no value in order. Sort first, then act; check impact first, then decide whether to delete.

One small thing to start with: open your article list, pick 10 at random, read only the first paragraph, and decide whether each is A, B or C. Ten minutes later you'll have a clear picture of your content assets.

Yihang Network, based in Shenzhen, previously worked in cross-border B2B and Alibaba.com-related services, and expanded into GEO from 2024. Our principle for legacy content: take inventory first, then sort — we don't favour blanket deletion.

Data and scope note: The citation-completeness study (arXiv:2410.22349), the UC Berkeley study (arXiv:2509.10762) and the Yonsei University study (arXiv:2602.12187) cited here are all preprints, not peer-reviewed; the KDD 2024 paper (DOI: 10.1145/3637528.3671900) is a peer-reviewed conference paper, and its findings are relative changes under lab conditions within a limited query set; the Google material comes from its official documentation. None of this represents real-market outcomes or any single company's results. The methods described raise the probability of being retrieved and cited; they do not guarantee a company will appear in AI answers, and no timeline or business result is promised.

1. Citation-completeness study, arXiv:2410.22349 (preprint, not peer-reviewed)
2. Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 (DOI: 10.1145/3637528.3671900)
3. UC Berkeley, arXiv:2509.10762 (preprint, not peer-reviewed)
4. Yonsei University, arXiv:2602.12187 (preprint, not peer-reviewed)
5. Google Search Central, "AI features and your website"