GEO-048·Issue 04·Measurement & Growth·48 / 132
Content Refresh: What to Change and What to Leave Alone
Most old posts do not need a rewrite. Four triggers, fact half-lives, and a change log decide what to correct, what to expand, and what to leave alone.
Issue focusAnswer Units: Writing Category and Product Pages for GEO
Consider a composite scene. Three weeks before gravel season, a US-facing bike computer brand plans its content. The growth lead wants to rewrite 140 posts. The engineer asks which pages Perplexity cited last week for pairing a computer with a phone. The sheet tracks dates, words, and sessions, not citations.
A page deserves a refresh when it is being quoted as current and a claim on it is no longer true — not because its publish date is old or its layout looks dated.
The problem with an old page is not age
Retrieval systems assemble answers from passages that read as present-tense facts. A 2024 post promising a two-year warranty is still a sentence in 2026, arriving without its publication date. The buyer hears a promise; the model repeated an expired claim. That is an accuracy problem, not a traffic one.
Two conditions must hold: the page is being quoted, and a claim has drifted. Otherwise you edit a working page or correct an uncited one. Recrawls also bend the report, and without a log of date, scope, and reason, nobody separates your edit from seasonality or a model change. A rewrite is legible — pages shipped, words written; a correction is three lines and looks idle.
Four triggers, checked in order
Run these in order. Accuracy first, demand later.
| Trigger | Signal | Window |
|---|---|---|
| Fact drift | Price, specs, model, certification, or warranty is out of date | Days to weeks |
| Citation loss | An answer that quoted you now cites a rival | Monthly |
| Question chain closed | Rivals answer questions you never addressed | Quarterly |
| Entry change | A new surface, language, or agent needs fields | Release cadence |
Fact drift outranks all: a stale warranty quoted by an assistant becomes a return, not a ranking issue. Fix it on discovery, outside the calendar.
Citation loss needs restraint: weaker content, a crawl failure, a rendering change, and a source switch look alike. Check access first — the AI crawler map covers those.
Question-chain completion wastes quarters. If the page is out of the candidate set and gets no referred sessions, new FAQs will not put it back.
An entry change needs structure: units in the spec table, a compatibility field, a returns table. Publish date, word count, and visual age are not triggers; they are reasons to redecorate.
Schedule by fact half-life, not publish date
A page’s date says when it was written. Its half-life says how long its facts stay true. Manage the second.
| Page type | Half-life | Refresh mode | Owner |
|---|---|---|---|
| Price, promo, stock | Days to weeks | Feed sync | Merchandising |
| Specs, certification, warranty | Quarterly | Correction | Product ops |
| Buying logic, comparisons, use cases | 1–2 years | Expansion | Content |
| Brand story, craft, founder | Rarely | Leave alone | Brand |
Four steps turn it into a queue:
- Inventory. For the 50 pages most likely to be quoted, record half-life and last-verified date — not last edited date.
- Triage. Run the four triggers; write one line saying change or not, and why.
- Change in the smallest mode. Correction for facts, expansion for a missing answer unit, rebuild only when the question changes. Most teams rebuild; most pages need the first two.
- Re-test. Rerun the same prompt set and record before and after.
Keep the URL and make the record honest
A refresh is not a relaunch: the URL carries citations, links, and history, so move it only when intent changes, with a 301. A sharper version is an edit in place.
| Object | Rule | Cost of breaking it |
|---|---|---|
| URL | Keep for refreshes; new URL and 301 only when intent changes | Citations, links, and history reset |
| Canonical | Self-referential per language | Language versions compete; AI gets the wrong market |
dateModified | Update only on substantive change; match the visible “updated” line | The field contradicts the page |
| Internal links | Update every inbound link when a URL moves | Dead paths and a broken crawl route |
dateModified is a declaration, not a lever. A schema.org property, it records change; a fresher date does not make a claim true. Update it on every save and it becomes noise; never update it and nobody knows whether the page holds. Change it when substance changes and keep the record in a change log.
When pages compete, keep the one with citations and links, 301 the other, and fix links. Invest by the page priority matrix, not the publish date.
Example: three pages, three decisions
Example — a hypothetical US-facing bike computer brand; no real products, prices, or results.
Page one, a model comparison, lists a withdrawn price and a discontinued mount. The correction removes the price, points to the product page for the live figure, and swaps the mount. Same URL, new dateModified.
Page two, “how to choose a bike computer for gravel,” is cited only for one line on battery life. The expansion adds an answer unit above it: who it suits, battery under load, maps versus turn-by-turn, and the road-unit difference. Original sentences stay. Same URL.
Page three is a founder story. The facts have not moved; the layout has. Leave it: reflowing would break the passage other pages quote, with no accuracy gain.
Two claims corrected, one rewrite avoided. The measure is not words shipped but whether the same questions return the right sentence.
Where this breaks
First, promotional and seasonal pages. Their lifecycle is a campaign: rebuild each season and archive the old version, not an expired price.
Second, marketplace listings. On Amazon or TikTok Shop the platform owns canonical and cadence; this policy is for owned sites.
Third, the pre-citation stage. With no citations, there is no loss to measure; write pages first.
Fourth, legal takedowns: remove the page; leave no empty shell.
Fifth, sampling noise. Answers vary by time and account; one failed test is not a trigger. Resample against the AI visibility baseline.
The rest of the line is on the measurement and growth hub; if the page is the gap, see the answer unit framework.
If you only make one change this month: take 20 cited pages, label each with a half-life and last-verified date, and calendar expired ones. Correct or expand only. Keep URLs, update dateModified only for substantive change, and log paragraphs, evidence, and result.