Short answer, as of August 2026: pages lose citations for five reasons — the facts on them went stale, a rival page started answering the question more directly, the entity data behind them drifted, a redesign or edit removed the part that was being quoted, or the question itself changed shape. Only the first is really about age. The rest are about relative quality and integrity, which is why a page can decay without a single word of it changing.
Nobody sends you a notification. A citation you held in March is simply absent in July, and unless you were sampling on a schedule you will discover it by accident, months late.
Why citations decay at all
A citation is not a ranking you hold, it is a decision the engine re-makes every time it is asked. Each time, it re-evaluates which available sources best answer the question — so your page is not defending a position, it is re-entering a competition it happens to have won before.
That reframes maintenance. You are not preserving a thing you own; you are staying the best available answer while the field of available answers keeps growing. A page that stands still is, in relative terms, getting worse.
The selection logic behind that re-evaluation is unpacked in how ChatGPT, Perplexity and Gemini pick sources.
The five decay modes
| Mode | What changed | Warning sign | Typical fix | Effort |
|---|---|---|---|---|
| Staleness | Your facts, prices or dates aged out | Page contains a year, price or "current" claim now wrong | Update the facts and the date, re-verify | Low |
| Outcompeted | Someone published a more direct answer | Competitor now named where you were | Deepen specificity, add evidence | Medium |
| Entity drift | Name, address, category or ownership changed | You are named less across all prompts, not one | Re-align sources, correct upstream data | Medium |
| Structural edit | A redesign moved or removed the quoted part | Citation lost right after a release | Restore the passage, server-render it | Low |
| Question drift | Buyers ask it differently now | Impressions shift to variant phrasings | Add the new phrasing, re-scope the page | Medium |
The fourth is the most avoidable and the most common in practice. A team rebuilds a page, the new design drops the comparison table or moves the specifics behind a tab, and the exact passage engines were quoting quietly ceases to exist. Nobody logs it as a content change because the page "says the same thing".
Staleness is about dates, not age
An old page is not automatically stale. A 2023 page explaining a mechanism that has not changed is as valid as it ever was. What ages badly is anything with an implicit time claim inside it.
Three things date a page hard: a year in the title or headings, a price, and words like "currently", "new" or "recently" attached to something that has since been superseded. Engines are sensitive to these because a user asking a question now wants an answer that is true now, and a page announcing a 2025 figure is self-evidently at risk.
The maintenance move is small: revisit the facts, correct what changed, restate the date explicitly, and update the modified date in your structured data and sitemap so the change is discoverable. Superficial refreshing — swapping the year in the title and republishing — is worse than doing nothing, because it makes a stale page claim currency it does not have.
Being outcompeted is the hardest one to see
This is the mode that looks like nothing happened, because nothing did happen to you. Your page is exactly as good as it was. A rival simply published something that answers the question with more specificity, or with better corroboration behind it, and the engine now prefers it.
The diagnosis is straightforward once you look: ask the prompt, read who is named instead, and read their page against yours. Usually the difference is concreteness — they give a number, a range, a named process, a dated example, and you give a description. The recovery is to out-specify rather than out-length; the working method is in what to do when a competitor is cited and you are not, and the page-level upgrade pattern is in retrofitting existing pages for citation.
Entity drift shows up everywhere at once
If citation loss is confined to one page, it is a page problem. If you are named less often across many unrelated prompts, it is an entity problem — something about who you are stopped resolving cleanly.
The triggers are ordinary business events: a rebrand, a merger, a moved office, a new legal name, a location closed, a domain change. Each one creates a period where sources disagree, and disagreement makes an assistant cautious about naming you at all. The repair is source alignment, described in entity consistency and the knowledge graph, and the specific case of assistants repeating obsolete facts is covered in correcting stale facts in AI answers.
The severe version of this is a site migration, where URLs, templates and entity signals all move at once. That deserves its own handling, which is in protecting AI visibility through a site migration.
A maintenance cadence that is actually sustainable
The mistake is trying to maintain everything. Most pages do not earn citations and never will; the ones that do are a small, identifiable set, and that set is what you maintain.
Monthly, on cited pages only. Confirm they are still cited. This is a chart, not a task, provided you are sampling on a schedule.
Quarterly, on the top ten cited pages. Re-read the facts. Correct anything that aged, restate the date, and check the passage engines were quoting still exists in server-rendered HTML.
On every release. Diff the content of pages that hold citations. This one rule prevents the structural-edit decay mode almost entirely, and it costs nothing but a checklist item.
Twice a year, on the prompt set. Language moves. The phrasings buyers used last year are not necessarily the ones they use now, and a prompt set frozen at setup slowly stops describing reality — the refresh discipline is part of building a prompt set.
Annually, on everything else. A light pass for obviously dated claims.
That is a few hours a quarter for most sites, and it is the difference between compounding and treadmilling.
Make loss visible
The reason decay is expensive is that it is silent. Rankings falling produce a chart nobody can miss; citations disappearing produce nothing at all unless you are looking.
DigiRank's AI Visibility Tracker samples your prompt set on a schedule across ChatGPT, Perplexity, Gemini, Copilot, Claude and Grok and stores the verbatim answers and sources, so a lost citation appears as a dated change with the replacing source named — which turns decay from an archaeology exercise into a work item. Scheduled tracking starts on the $99/mo Starter plan; the content ROI engine and visibility timeline that tie citation movement to what you published arrive with Agency at $249/mo on a 14-day trial. For the baseline the whole exercise is measured against, AI visibility benchmarks and share of voice sets out how to establish one properly.
Frequently asked questions
Why did my page stop being cited by AI? Usually one of five things: the facts on it aged, a rival published a more direct answer, your entity data drifted, an edit or redesign removed the passage being quoted, or buyers started asking the question differently. Only the first is genuinely about age — the others are about relative quality and integrity.
How often should I refresh content for AI search? Check cited pages monthly, review the top ten quarterly, diff cited pages on every release, and revisit your prompt set twice a year. Refresh on substance rather than schedule: changing a date without changing a fact makes a stale page falsely claim currency.
Do citations expire on their own? Not on a timer, but they are re-decided on every query rather than held. As more pages compete to answer the same question, a page that stands still becomes relatively weaker even though nothing about it changed.
How do I know if I lost a citation? Only by sampling on a schedule and storing the results. There is no notification and no dashboard from the engines, so a citation lost in March is typically noticed months later unless you were recording answers with dates.
My rankings are fine but citations dropped. What does that mean? Ranking makes you eligible and citation is a separate selection step, so the two can diverge. Common causes are a competitor answering more specifically, or a crawler-access change that affects assistants without affecting search engines at all.
Is updating the year in a title enough? No, and it is actively harmful. Superficial republishing makes a page assert currency its content does not support, which erodes trust with both engines and readers. Change the facts or leave the date alone.
Which pages should I prioritise for maintenance? The ones that already hold citations, then the ones that rank well and answer commercially important questions. Maintaining pages that were never cited and never ranked is the most common way this discipline consumes time without returning anything.
Does a redesign put existing citations at risk? Yes, more than most teams expect. If a rebuild moves specifics behind tabs, removes a comparison table, or renders content client-side, the exact passage that was being quoted can vanish while the page appears unchanged. Diff cited pages before and after every release.