AI Citation Half-Life

A metric that doesn't exist yet

Every published AI-citation study, including several referenced elsewhere on this site, reports a citation snapshot: what was cited, on a given date, for a given query set. None of them, as far as we could verify, re-check the same set of citations weeks or months later to see whether they're still there.

That's a real gap. A citation that disappears after two weeks, and one that persists for a year, get treated identically by every existing measurement. But they represent very different outcomes for whoever earned that citation.

Why "half-life" is the right borrowed term

Radioactive decay, drug metabolism, and viral-post engagement decay all share a mathematical shape: a quantity starting at a peak and falling away over time, often in a pattern well described by the time it takes to fall to half its starting value.

Borrowing the term signals a specific claim about the expected shape of the data. Not just "citations fade eventually," which is an unfalsifiable truism. Instead: "citations fade in a roughly quantifiable, comparable pattern across URLs and engines." That's a testable claim this study is actually positioned to confirm or refute.

Defining half-life

Applied here — Citation Half-Life is the number of days until a URL's citation rate across the tracked query set falls to 50% of its peak observed rate. A URL cited in 80% of relevant runs at peak and dropping to 40% has reached its half-life at that point, regardless of whether it disappears entirely afterward.

What a decay curve might look like

If real data eventually resembled this illustrative shape, the half-life point would fall somewhere between months 5 and 6, where the curve crosses 50 on the vertical axis. Is real citation decay this smooth? Faster? Slower? An entirely different shape, like a sharp cliff rather than a gradual curve? That's exactly what the study is designed to find out.

Study design

Pre-registered hypotheses

HL1's predicted direction, against persistence, is deliberately the more provocative bet. Suppose most citations are, in fact, durable for 90+ days. That would be a genuinely useful and reassuring finding for anyone investing in GEO. Registering the opposite prediction up front means nobody can accuse us of shaping the result after seeing it.

Why this matters practically

Suppose citations decay quickly and unpredictably. Then "getting cited once" is a much weaker goal than "maintaining citation." That implies an ongoing content-maintenance strategy, not a one-time publish-and-forget approach. Suppose citations turn out to be durable instead. Then the opposite advice follows: initial placement matters far more than continued refreshing. Nobody currently knows which world we're in.

Schedule

Tracking begins with the same URL set identified in Index v1's first collection, re-checked at months 4, 7, and 10. First meaningful half-life estimates publish in Q3 2027; the full 12-month survival curve lands in the State of AI Search annual report.

Untangling the causes, eventually

Measuring that a citation decayed is the study's primary goal. Explaining why a specific citation was lost is a harder, secondary question. This design isn't initially built to answer that with certainty. A lost citation could reflect the source page itself going stale, a competing page displacing it, an underlying model update shifting retrieval behavior, or simple query-result churn unrelated to content quality at all.

Once baseline decay patterns are established, a natural next step is a follow-up analysis correlating loss events with observable page-level changes, content updates, competitor citation gains. But distinguishing these causes with confidence would likely require its own dedicated, controlled design, not something read directly off this study's observational data.

What to do while the study runs

You don't need a finished half-life number to hedge sensibly. Treat every citation you earn as something to maintain, not a one-time achievement. Revisit and refresh your most important cited pages on a schedule, rather than assuming a citation earned once stays earned forever.

If you can, track your own cited pages informally: note when you first notice a citation, and check back periodically to see if it's still there. That kind of informal tracking is exactly the small-scale version of what this study is doing at scale, and it will make this study's eventual findings easier to interpret for your own site specifically.

Limitations

• A URL's disappearance from citation could reflect many causes — content decay, competitive displacement, or model updates — which this study is not initially designed to distinguish between.
• Twelve months is a long baseline for a field this fast-moving; the underlying engines may change meaningfully mid-study, which will be logged as a confound rather than hidden.