AI CITATIONS DECAY IN ABOUT 3 MONTHS. HERE'S WHY.

AI citations decay in about 3 months because engines that run live web search weight recency at retrieval time: roughly 50% of all content cited in AI answers is less than 13 weeks old, and pages that stop being refreshed lose retrieval priority to newer competitors within 90 to 180 days. The flip side is speed: new content can enter AI citation pools within 3 to 5 business days, and content updated in the past 30 days earns roughly 3x more ChatGPT citations than older material. That combination rewrites the economics of AI search visibility. A one-time optimization project produces a citation spike that fades within a quarter, while publishing velocity and a scheduled refresh cadence are now ranking levers in their own right. Below: why the decay happens, how fast new content gets picked up, how to detect erosion in your own numbers, and what a cadence that holds Share of Model looks like.

HOW FAST DO AI CITATIONS DECAY?

The numbers converge across independent studies. Half of everything cited in AI answers is under 13 weeks old. Content updated within the past 90 days gets roughly a 200% citation boost for commercial queries and 150% for informational queries, compared to content over a year old. Decay sets in between 90 and 180 days for most pages, with commercial topics decaying fastest because competitors publish there fastest. Once your page crosses the 3-month line without an update, it starts losing retrieval slots to whatever was published last month, even when the newer content is thinner.

WHY DO AI ENGINES PREFER FRESH CONTENT?

Three mechanical reasons, none of them sentimental:

• Retrieval pipelines rank recency directly. Engines with live web search (all four majors as of 2026) select sources at answer time. A recent publish or update date is one of the cheapest, most reliable quality proxies a retrieval system can use, so they use it.
• Fresh statistics outcompete stale ones. AI systems seek concrete data to extract, and a 2026 number beats a 2024 number for the same claim. Content with recent statistics and credible citations shows up to 89% higher selection probability in AI Overviews.
• The pool keeps refilling. Citation slots per answer are fixed (engines typically cite 3 to 10 sources), while new candidate content arrives daily. Your page does not get worse; the denominator gets bigger and newer.

HOW FAST CAN NEW CONTENT GET CITED?

Faster than traditional SEO ever worked. New content can enter AI citation pools within 3 to 5 business days of publication, because search-native engines like Perplexity retrieve live rather than waiting on index refresh cycles. Compare that to the 3 to 6 months a new page typically needs to rank meaningfully in classic organic search. The feedback loop tightened from quarters to days, which is exactly why monthly Share of Model measurement catches movement that quarterly reporting misses entirely.

WHY ONE-TIME OPTIMIZATION FADES

A single optimization project does real work: restructured pages, embedded statistics (worth a 30 to 41% visibility lift per the Princeton GEO-bench data), attributed quotes (28 to 32%), schema, extractable answer blocks. Citations rise. Then the clock starts. The statistics age, the publish dates age, competitors refresh, and the citation share you bought erodes on the same ~3-month curve as everything else. This is not a flaw in the optimization; it is the physics of a freshness-weighted system.

It also explains a pattern we see in measurement data: a company spikes from 0% to 20% Share of Model after an implementation, celebrates, stops, and is back under 10% two quarters later, while attributing the decline to "algorithm changes." Nothing changed except time. Rand Fishkin's line about the old metrics applies doubly here:

"Traffic is a vanity metric."

So is a one-time citation spike. The durable asset is the cadence that maintains the position, plus the off-site consensus (Reddit at ~40% citation frequency across engines, LinkedIn at #2, YouTube as the top single domain) that decays much more slowly than on-site freshness. Each engine weights these differently.

WHAT A REFRESH CADENCE LOOKS LIKE

• Monthly: re-measure. Share of Model per engine against a frozen prompt set. Movement down on one engine is your early warning.
• Every 60-90 days: refresh the citation drivers. Update statistics to current figures, swap aged examples, add one new attributed quote, and update the visible and structured dateModified. A substantive refresh, not a token edit: engines discount cosmetic changes.
• Every month: publish something new in the topic cluster. Google AI Overviews cites pages that appear across fan-out sub-queries, so cluster breadth compounds. One new extractable answer per month per priority cluster is a sustainable floor.
• Quarterly: refresh off-site surfaces. Third-party lists capture roughly 74% of AI citations; make sure the ones you appear on are current, and get on the new ones.

This cadence is, transparently, the argument for ongoing engagement over project work, ours included (services, examples). But the logic stands regardless of who executes it: in a freshness-weighted system, visibility is a subscription you pay with content, whether or not you pay an agency.

HOW TO TELL IF YOUR CITATIONS ARE DECAYING RIGHT NOW

Run the same prompt set you ran last quarter (the methodology documents the protocol). If your Share of Model dropped while nothing on your site changed, check three things: the age of the statistics on your most-cited pages, your last substantive update date, and what your competitors published in the past 90 days. In most decay cases we diagnose, the answer is simply that the winning pages are newer. If you have never measured at all, start with the baseline: the free scorecard includes a live Share of Model run, and 0% is the most common first result we deliver.