The half-life of AI citations. How fast you stop being cited

AI visibility behaves nothing like a search ranking. What a language model cites today it almost certainly will not cite tomorrow. A place earned on the shortlist does not stay earned; it holds briefly and needs constant renewal.

GolOps tracked this over time. Every day for 10 months we ran the same choice scenarios through 8 AI systems and watched what changed. At the level of individual links, citations almost never repeat. At the brand level, mention counts swing from week to week. Two questions drove the work: does ChatGPT cite the same article tomorrow that it cited today, and does Perplexity mention a brand as often next week as this one?

Key findings

73.5% — one-and-done citations. Almost three in four URL citations appear exactly once and never return inside the observation window. Median lifespan: zero days. The volume of citations does not fall — models cite just as much, just different sources each time.

31 days — the brand half-life. The median brand loses half of its peak presence in a month. A citation does not accumulate; it decays.

51.8% — weekly swing. Even without anchoring to a peak, a typical brand's citation count swings by half week over week. The baseline rarely repeats twice.

13.2% — the stable share. Of brands observed for 90+ days, only one in eight holds a steady presence. The rest cycle through decay and rebound.

Ghost citations

The first lens is at the URL level. We tracked 108,650 citing links across 960 brands and asked one question: when an AI system cites a URL, how long does that citation hold?

The answer is stark. 73.5% of URLs appear exactly once and do not return. This is not a shrinking pool — models keep the citation volume but refresh the lineup. The links a brand earned yesterday may already be out of rotation today. This squares with independent data: a Profound study documents constant citation drift across every major AI platform.

Share of citations from the first observation window still appearing after N days. 108,650 URLs · 200 brands

Fewer than a third of links survive the first week. By two weeks, under a quarter. This is not fading interest in the topic — the aggregate citation volume per brand is stable. What changes is which sources carry that volume.

The decay curve

A single URL is a narrow lens. A brand has many links, and some of them stay. So the second lens is at the brand level. Across 857,000 daily measurements covering 10,991 brands over 10 months, we measured how long the citation count holds its peak. The median brand drops to half in 31 days.

That is decay from peak. But even without anchoring to a peak the picture is the same: the weekly citation swing for a typical brand is 51.8%. Whatever baseline a brand earns, the next week barely resembles it. A similar scale of instability shows up externally: according to SISTRIX (82,619 prompts over 17 weeks), Google replaces 56% of the sources in AI answers every week, and ChatGPT replaces 74%.

The distribution of decay speed shows how short the memory of AI output really is:

Days from observed peak to first 50% drop. 552 brands

Almost half of brands lose half their presence in a single week. A measurement taken once a quarter passes over two half-lives of drift before the next pass — and sees not the real position, but its random residue.

The stability premium

Of 1,314 brands tracked for 90+ days, only 173 hold a steady presence. The stable minority averages 74% presence; everyone else cycles through decay and rebound.

Almost half the sample sits in the 30%+ swing band. Stability is not bought with size. It rises sharply in the move from small brands to medium ones, then plateaus.

Volatility falls from 59% to 33% in the move from small to medium brands. After that the effect disappears: a large brand is no steadier than a medium one, and the mega tier sits in the same swing band as large. Past a certain threshold, extra citations stop buying extra stability. Stability is not the volume of presence — it is the manageability of presence.

What sticks, what vanishes

Sources decay at different rates. Some domains hold their position for weeks; others appear once and do not return inside the window. The stickiest domain in the sample held an average of 32 days between first and last appearance; the most volatile, under a day.

Source type explains the gap. Persistent sources skew toward niche industry tools and specialized guides — structured, slow-changing content. Volatile ones skew toward large marketplaces and social platforms, where content rotates quickly. It is the same pattern seen in the source map: AI holds onto what is marked up and stable, and quickly drops what is noisy and rewritten. What exactly makes a page "sticky" enough to cite is something we broke down in The anatomy of an AI citation. Freshness matters too: an Ahrefs study of 17 million citations across 7 AI platforms found that assistants prefer to cite fresher content.

Average days between a domain's first and last citation

Visibility has to be held

Stack three numbers. Most citations appear once. The brand half-life is a month. Only 13% of brands hold steady. One conclusion follows: AI visibility is not a milestone you reach but a position you hold.

That changes the work. A single citation proves a brand made it into the output now, but not that it will stay. A one-off AI audit run once a quarter misses two half-lives of drift before the next pass. And the teams that stay in the stable 13% behave less like campaigns and more like operators: they do not launch at the field of choice, they run it continuously.

Three principles follow.

Measure continuously. A quarterly audit cannot see a 31-day half-life. The same scenarios have to run on a weekly or daily cadence, so losses and recoveries land inside the window where they can still be acted on.

Watch the pair, not the snapshot. The one-and-done rate (73.5%) and the brand half-life (31 days) work together. URL churn tells you whether individual wins are sticking; the brand half-life tells you whether the position is holding overall.

A spike is not a win. A 51.8% weekly swing is normal noise, not a trend. A new citation is worth treating as provisional and counting as a win only after it has held for several measurement cycles.

Methodology

What underpins the numbers:

• 857,138 daily measurements across 10,991 brands from June 1, 2025 to March 30, 2026, across 8 AI systems. The brand-level half-life is computed on this data.
• 108,650 URL citations across 200 brands (sampled from 960), window October 4, 2025 to March 30, 2026, 177 days. Individual-link survival is computed on this data.
• 8 AI systems — ChatGPT, Claude, Gemini, Grok, DeepSeek, Meta AI, Perplexity, Google AI Overviews.
• Protocol — the same scenarios run daily against identical prompts; first-seen and last-seen dates recorded for every URL.

Honest gaps. URLs first seen close to the end of the 177-day window had less time to reappear, so their measured persistence can look shorter than it would in a longer study. The half-life is measured from each brand's observed peak, so part of the recorded decline is regression from an unusually high peak rather than a loss of visibility as such. Some brands stopped being tracked mid-window and some teams changed their scenario sets — those slices are excluded from long-term statistics. We do not yet have a clean external benchmark for the stability of classic search rankings, so the comparison to search stays qualitative.

The management implication

A 31-day half-life changes the unit of management itself, not individual tactics. If brand presence halves in a month, then a measurement taken once a quarter passes through two half-lives before the next pass and shows not the real position but its random residue. A one-off audit captures a photograph of a cloud: by the time the report lands on the desk, half of that position is gone. That is the case for continuous measurement over project work done in bursts: a position in the field of choice cannot be won once and closed; it can only be held.

This is the work GolOps puts into a continuous mode. We measure a company's position through the Choice Control Index, attribute it to the specific sources and scenarios that shape it, and translate the measurement into a prioritized plan. The Strategic Pilot closes the first cycle in 10–12 weeks. The Command Center keeps the loop running continuously — constant measurement across 7 AI systems with source attribution. That is the direct answer to the decay: not a single report, but measurement at the same frequency the output itself moves.

Decay is not the only asymmetry in AI citation. Page type also decides what makes it into the answer:

AI crawls your product pages. It cites your blog.

What infrequent measurement costs

Between two rare measurements the shortlist reassembles without you — and you do not find out until the next pass. Gartner forecasts 90% of B2B procurement under autonomous AI agents by 2028, and Semrush already shows AI-channel conversion running 4.4× higher than organic search: the cost of lost shortlist presence is rising faster than it looks. And it is measured in a simple unit — every 31 days without continuous measurement costs half of your presence, decayed unnoticed.

Request an index diagnostic → · Discuss a pilot →