AI Citation Half-Life: Why Sources Drop Out in Weeks

When a buyer asks an AI assistant for a vendor, the answer arrives as a shortlist. Two or three names, typically. The named businesses become options. The unnamed ones are absent from the conversation. The buyer never sees them.

Most businesses measure the traffic that arrives. Few measure the shortlist they missed. That gap is where AI citation half-life lives. A source can be cited in an assistant’s answer on Monday and gone by Friday. The mechanism is straightforward. Assistants retrieve pages from a search index, then rank and synthesize an answer. Indexes refresh. Models update. Competitor pages gain new links or structured data. The retrieved set shifts. The citation disappears.

aeod.app assesses how assistants answer buyer questions. It records mentions and citations, and it tracks position across repeated runs. One pattern from those assessments: the same question produces different source lists over a period of weeks. The source that anchored an answer in one run does not appear in the next. That volatility is the subject of this article.

The definition below gives that churn a time scale. It turns a vague observation about changing answers into a number a team can track.

What an AI citation half-life measures

An AI citation half-life starts with a cohort. For a fixed set of buyer prompts, collect every source URL cited by an AI assistant during a given week. If the week 1 cohort contains 120 unique URLs, those 120 become the baseline. In week 2, repeat the same prompts and count how many of the 120 URLs still appear in the answers. Divide that count by 120. Repeat the process each following week. The half-life is the point where the retention share reaches 50 percent. If week 4 retention is 55 percent and week 5 retention is 45 percent, linear interpolation puts the crossing at 4.5 weeks.

A mention rate counts something different. It counts brand names in the answer text. If an assistant writes that Acme handles emergencies, the mention rate records Acme. The citation half-life records the source URLs underneath the answer, such as acme.com/emergency-plumber. The two measures diverge because one tracks text strings and the other tracks URLs. A brand keeps its place in the prose while its cited pages fall out of the source list. The brand stays visible in the answer; the evidence behind the brand does not.

Decay is a source-level phenomenon when prompts are repeated, because the cohort is defined by URLs and the prompt set is the control. If the question wording changes each week, a source disappears when the question changes. Holding prompts constant isolates source turnover. The same URL drops when the assistant swaps in a directory page, a competitor’s pricing page, or a newer article. The cohort method assigns that movement to the source layer, which is the layer an AI answer uses to justify its shortlist. That source turnover sits underneath the pattern in why AI assistants recommend the same businesses. The names in the shortlist look stable while the URLs supporting them rotate.

The measurement freezes a prompt set, captures cited URLs each week, and calculates retention against the original cohort. Applied to tracked prompt sets, the half-life lands at 4.5 weeks. The distribution has a bimodal tail.

The measured number: 4.5 weeks, and a bimodal tail

Scrunch and Stacker tracked 3.5 million citation events across six platforms from September 2025 to March 2026. The median citation half-life was 4.5 weeks. That means the typical cited source loses half of its citation volume in about a month. The panel spans six platforms, so the median reflects a broad event base across multiple assistants. The number is not a single smooth decay curve. Frozen prompt sets land on the same 4.5-week median, which makes the figure reproducible across independent prompt cohorts.

Separate persistence work shows the distribution has a bimodal tail. A source either returns on the next run or disappears for roughly a month. Middle ground is rare. Citations behave like a switch: present today, absent tomorrow, then present again after a few weeks when the source gets re-indexed or re-retrieved. Averages hide two distinct populations. One population turns over with the weekly run. The other drops out and waits for a fresh crawl. The tail matters for reporting because a source that drops out looks dead for a month, then reappears in the next cohort if the crawl catches up.

Amsive’s supporting benchmark adds context. Half of AI-cited content is under 13 weeks old. The 4.5-week half-life and the 13-week age ceiling describe the same pressure from two angles: sources age out of the answer set quickly, and the answer set skews toward recent documents. A page can be cited in one run and fall outside the citation window in the next. The bimodal tail is the visible signature of that window.

The practical reading is simple. A citation is a snapshot with a short shelf life. A weekly retention measurement catches the first drop. A monthly check catches the reappearance. The baseline to plan against is 4.5 weeks. The variance around that baseline is wide, and the shape is lumpy. That lumpiness is the reason a single average is a poor planning input.

The next section examines why sources fall out: fan-out, freshness, and index churn.

Why sources fall out: fan-out, freshness, and index churn

AI answers look stable from the outside. The source list underneath is not. Retrieval fan-out decides which queries the assistant sends to search or to its own index. Freshness ranking filters by modification date. Index refresh swaps the passage. Each mechanism removes pages that were cited last week.

Retrieval fan-out changes the candidate set. A query fan-out change can produce a sharp single-week drop in ChatGPT’s Reddit citation share. The same page remains live, crawlable, and relevant while falling outside the paths the retrieval system runs. That explains volatile citation counts when a brand’s website does not change. Tracking mention and citation position across buyer questions makes this redistribution visible as source churn.

Freshness ranking applies the next filter. Among cited pages, about 72% were updated within the past year, against 42% published in that window. The 30-point gap points to revision behavior. A 2023 comparison page that receives a 2026 pricing update carries a recent modification signal. A page published in 2025 with no updates does not. Assistants favor passages that match current facts. An older cited page loses position when a competitor publishes a more current passage on the same buyer question.

Index refresh handles replacement. The answer was built from a specific passage at a specific index state. When the crawler revisits, that passage is rewritten or removed. The assistant reconstructs the answer from whatever passage now sits in its place. The source does not need to disappear. The exact text that earned the citation does. This pattern sits underneath repeated shortlists. Repeated recommendation requires repeated retrieval of current passages.

A business name can persist in model outputs after the page that supplied the evidence leaves the index. The name has been repeated across enough retrieved passages that it survives as a model association. The citation does not. That gap between citation loss and recommendation loss is the next problem. A shortlist can keep the same names while the URLs that justified those names rotate underneath. The brand remains visible in the answer, and the evidence behind the brand becomes harder to trace.

The persistence paradox: fewer citations, longer life

RankCaster analyzed 5.22 million citations and found a pattern that runs against intuition. Pages that persisted for 139 days or longer averaged about 17 times fewer citations than pages that disappeared sooner. The durable pages also appeared across more providers and covered more topics. Lower citation volume accompanied longer life.

That result is a warning about narrow, high-volume citation patterns. A page cited hundreds of times by one assistant for one query class behaves like a single-purpose page. It churns. When the provider changes retrieval or prompt phrasing shifts, the page loses its only entry point. The high count creates an illusion of strength. The count is concentrated. The page has no fallback. It drops out of shortlists in weeks.

The 139-day threshold separates pages with staying power from pages with a spike. Persistence at that scale tracks breadth. A page that answers related buyer questions across multiple assistants has more retrieval hooks. It earns citations for different prompts and comparison queries. Pages built around comparison prompts such as X vs Y give assistants several ways to surface the same source. The goal shifts from maximizing citation count to widening the set of questions and providers where a page is eligible.

The assessment measures mentions, citations, and position across assistants, so breadth gaps appear next to raw citation totals. The prioritized action list then points to topics and providers that lack coverage. This is the practical response to the persistence paradox. A page with 40 citations from one provider for one query has a shorter expected life than a page with 3 citations spread across five providers and eight related questions. The second page has more ways to survive a retrieval update.

Why a single check cannot see decay

A single check captures presence at one moment. It cannot show turnover because it has no baseline to compare against. Decay appears only when the same prompt set runs again and the cited URLs are compared with an earlier cohort. Persistence depends on breadth across providers and topics, and that breadth changes as retrieval systems update. One snapshot can confirm that a source is cited today. It cannot show whether that source will be cited next week, or whether the citation is concentrated in a single retrieval path.

Decay differs by prompt type

Persistence varies by prompt type, so measurement belongs at the prompt level. A source that holds across re-runs on a broad category query follows a different decay curve than the same source on a constrained follow-up.

Broad category prompts draw from a wide retrieval set. Ask an assistant for “best project management software for agencies” and the retrieval layer pulls category pages and review roundups. The query has few constraints, so embedding similarity to generic intent stays stable across index refreshes. A source cited in one re-run remains eligible in the next because the top-k retrieval pool changes slowly. Decay here looks like gradual rotation.

Constrained follow-ups rebuild the retrieval path each turn. A buyer starts with the broad query, then adds “for a 12-person team using Slack and Xero under $15 per seat.” Query rewriting injects those constraints into a new search. The context window carries the conversation, yet the retrieval layer runs again. Pages that matched the first query drop out when they lack the new terms. The source vanishes mid-conversation and gets replaced before the buyer finishes asking. That behavior is the core of multi-turn AI visibility follow-up questions. Survival measured only on the opening prompt overstates persistence.

Pair prompts behave differently again. “Asana vs Monday” anchors retrieval to pages that discuss both named entities. Entity co-occurrence in titles and comparison tables drives selection. A page that covers one product well but mentions the other once has weaker footing than a page with a dedicated section for each. The retrieval set is narrow and stable because the two names define the query. A source persists on a pair prompt while dropping from category prompts. The mechanics behind that split are covered in AI comparison prompts: X vs Y verdict.

A blended survival rate across all prompts hides these differences. A business can hold on broad category prompts while dropping on constrained follow-ups and holding on pair prompts in the same audit window. Those patterns point to different actions. Broad category decay calls for maintaining category pages. Constrained follow-up decay calls for covering specific attributes and integrations. Pair prompt decay calls for dedicated comparison content. Segmenting survival by prompt class makes those patterns visible. Once survival is segmented by prompt type, the next decision is setting a refresh and re-measurement cadence.

Setting a refresh and re-measurement cadence

Set the cadence by retrieval path. Prompts that trigger live web retrieval, such as ChatGPT Search and Perplexity, read the current index and current page text. Run those prompts weekly. Prompts answered from model knowledge, such as Claude without search or a GPT snapshot with no browsing, draw on training data and cached behavior. Run those monthly. If a change is made to affect visibility, run the affected prompt immediately after the change. That includes edits to a cited page, new schema, a new comparison page, or a new review profile. The immediate run tests whether the retrieval layer picked up the change. A delayed run tests whether the change held.

Refresh the page a citation actually pointed to. A citation is a URL. Update that URL and the next retrieval sees new text at the same address. Publish a new page and the old citation continues to point at the old content, while the new URL starts without the historical signals the old page accumulated. The retrieval systems follow updated pages more than new pages because the pointer already exists. This is one reason the same businesses appear across audits. The same mechanism appears in multi-turn follow-ups: the model keeps using the source it already retrieved unless the new turn forces a new search.

The cadence has three clocks. Use the weekly clock for live retrieval prompts. Use the monthly clock for model knowledge prompts. Use the immediate clock for changes. The log tracks citation sources by prompt and model, and shows which URL was replaced. Log every run in one table: date, prompt, model, cited URL, position, and the source that replaced the cited URL. The replacement source is the diagnostic. If a competitor page replaces yours, that names the retrieval path that took over. If a forum thread or directory replaces yours, the path moved to community answers or structured listings. The replacement source tells you which surface the model now trusts for that prompt.

That log also segments by prompt type, which connects to the survival patterns from the previous section. Those prompt segments decay on different clocks. The replacement source explains the decay. The next question is what to do with this.

What to do with this

Earning a citation and holding one are different events on different clocks. A model names a source because a query fan-out hit a page once. The next run rebuilds retrieval from scratch. The page stayed the same; the assistant’s path changed.

Decay is the default state. A citation that disappears reflects a retrieval result that did not recur. That distinction changes the measurement. A weekly snapshot of mentions shows a binary. A survival curve over six to eight weeks shows whether a source persists, and where it drops.

Each drop needs a named cause. A fan-out shift changes which sources get queried. An unrefreshed page loses recency signals that several assistants weight during ranking. A retrieval path moves to another source when a competitor adds a clearer answer block. Without the cause, the drop rate is only a count.

aeod.app produces that reading: survival by prompt segment, with each drop tied to a logged change in the answer set. The log turns decay into a diagnostic list.

The method has a hard limit. Retrieval is rebuilt when the question is asked. No page edit guarantees a citation survives the next run. The useful target is a longer half-life, measured over weeks, with the cause of each loss written down.