What we found across 240 citations

Between January and April 2026 we tracked 240 verified AI citations across ChatGPT (with browse), Perplexity, Gemini, Claude (with browse) and Google AI Overviews. Daily polling on a fixed prompt set, brand-presence logged as a binary, survival curves computed weekly.

The median citation lasted 18 days. The 25th percentile dropped at day 7. The 90th percentile was still alive at day 121. A long tail of evergreen reference pages held position for the full 90-day window without ever dropping.

The headline number most teams miss: at week 6, roughly half the citations earned in week 0 were already gone. Most AEO programmes are sold and budgeted as if citations behaved like backlinks — earned once, persisted indefinitely. That is not what the data shows.

This post is the methodology, the curves by platform and content type, and the defenses that work.

What “citation half-life” actually means

A citation is the appearance of your brand or URL inside an AI answer for a specific prompt. Half-life is the time it takes for half of a cohort of citations earned in week zero to disappear from the answer.

The crucial distinction — half-life is a per-prompt-cluster metric, not a per-page metric. The same URL can hold a citation on prompt A and lose the adjacent citation on prompt B inside the same week. Why — because the LLM regenerates the answer per query, with different retrieval candidates each time. The page did not change. The competitive set around it did.

This is why “we lost 30% of our citations this month” is a misleading sentence on its own. You need to know which prompts dropped and whether the URL is still earning the others. Otherwise you are panicking about a metric that does not map to the underlying reality.

Method — how we measure

The stack is similar to what we use on every retainer, with one addition for the half-life study specifically.

Daily polling. A fixed prompt set of 30–60 prompts per client niche. Every prompt runs once per day per platform. The full answer text is logged plus the source list (where the platform exposes one). We use Searchable Agent for the bulk of this, with a Perplexity API DIY layer for the fine-grained Perplexity tracking — Perplexity is the cleanest measurement surface because the source block is public.

Brand-presence binary. For each prompt-platform-day cell, we log a 1 if the brand appears in the answer text or the source block, a 0 otherwise. No partial credit. No sentiment weighting at this stage — that is a separate metric.

Survival curve. For each cohort of citations earned in week zero, we plot the percentage still alive at week 1, 2, 3, etc. The half-life is the point where the curve crosses 50%.

Manual cross-check. Friday afternoon, the SEO lead manually verifies a sample of 10 citations across the five platforms. Automated tools miss between 5% and 15% — usually citations where the LLM names the brand without the URL, or where Perplexity rotates which source-block position the URL appears in. The manual pass corrects the ledger.

DIY without commercial tools — possible. Use the Perplexity API for daily prompt runs, log the source list, persist to a sheet. ChatGPT and Gemini require a browser-automation layer (Playwright + a cookie management strategy). Doable, but the maintenance burden is real and you will spend more on engineer-hours than the $400–800/month a commercial tool costs.

Half-life by content type

The five categories below cover roughly 95% of what we have measured.

News and announcements. Half-life around 10 days. Funding rounds, product launches, regulatory announcements. They earn citations fast — sometimes within 24–48 hours of publication — because LLMs heavily weight recency on news-shaped prompts. They die equally fast as the next news cycle overwrites the slot.

Tactics and playbooks. Half-life 30–60 days. The AEO-friendly piece you are reading right now is in this bucket. They hold longer because the prompts that retrieve them (“how to measure X”, “best way to ship Y”) are not recency-anchored. But they get displaced when a competitor publishes a sharper or more recent version of the same tactic.

Evergreen reference. Half-life 90–180 days. Definition pages, taxonomy articles, glossary-style content, regulatory primers. These are the workhorses of an AEO programme. They take longer to earn citations (often 60–90 days for the first one) but they stay cited for the full quarter or longer.

Regulatory and legal. Volatile. Median half-life looks long until a regulation changes — then the entire cohort drops within a week as competitor pages with the new info take over. We treat this category as a special case in client reports because the regression risk is concentrated.

Product comparison and pricing. Half-life 14–30 days. These pages have to be refreshed constantly because the underlying competitive set changes. A pricing page that goes 60 days without an edit will lose half its citations.

Half-life by platform

The five platforms behave differently enough to need separate curves.

Perplexity. Most aggressive citation rotation. Fresh content beats stale within days, not weeks. Half-life on news 14 days, on evergreen 6 weeks. The upside — you can earn a citation in 7–14 days after a structural rewrite. The downside — you can lose it just as fast if a competitor ships a sharper version.

ChatGPT (with browse). Slower to grant citations, slower to revoke them. News half-life around 3 weeks, evergreen around 3 months. ChatGPT also re-uses citations more — once you are in the answer, you stay in it through several index refreshes unless something pushes you out.

Gemini. The most news-biased of the five. Half-life on news 14 days. Half-life on evergreen 4–6 weeks. Gemini’s retrieval seems to weight dateModified very heavily, so a stale page falls out faster here than anywhere else.

Claude (with browse). Closest to ChatGPT. Tactics half-life around 5 weeks, evergreen around 10 weeks. Claude is the smallest sample in our dataset because browse rollout was uneven across the period — treat the numbers as directional.

Google AI Overview. The half-life concept partially breaks. AIO regenerates the answer per query, with daily-or-faster volatility on most prompts. A page can be cited on Tuesday’s run, missing on Wednesday’s, back on Thursday’s — without the page or the prompt changing. We report AIO using a different metric — “percentage of last 30 days the page was cited” — rather than a survival curve.

Why citations decay

Five mechanisms, in rough order of frequency.

Newer competitor content. The single biggest cause. A competitor publishes a sharper take on the same prompt cluster, the LLM picks it up on the next index refresh, your slot disappears. We see this on roughly 60% of measured drop-offs.

LLM index refresh. Each platform refreshes its index on a different cadence — Perplexity is close to daily, ChatGPT closer to weekly with deeper monthly refreshes, Gemini somewhere in between. A refresh can re-rank the source set without anything else changing.

Recency-weighted retrieval. Even without competitor pressure, retrievers downweight old content. A page with dateModified from 14 months ago will fall behind a page from 3 weeks ago for the same prompt — assuming similar structural quality.

AI Overview re-clustering. Google occasionally rewrites the prompt clusters AIO is willing to answer. A query that yielded an Overview last month may show a regular SERP this month — and your citation simply has nowhere to live until the prompt comes back.

Schema and structural drift. The page itself changes — someone edits the H2 structure, removes a Quick Facts table, breaks the FAQ schema during a CMS migration. The page is intact for human readers but no longer extracts cleanly. Underrated cause; we have seen it kill half of a portfolio in a single deployment.

Defenses that actually work

The full playbook is in the Growth tier, but the highest-ROI moves are these.

Quarterly refresh with a visible date stamp. Add dateModified updates plus a human-readable “Updated April 2026” line in the visible header. We measure 40–60% half-life extension on evergreen pages with this single move. It is the cheapest defensive intervention available.

Republish with new examples. Every six months, swap two or three of the examples on a tactics page for fresher ones. This signals to the LLM that the page is being maintained — and the new examples themselves earn citations on adjacent prompts.

“What changed since last version” sections. A short block at the top — three bullets, last updated date, what changed. Both human readers and LLM retrievers reward this. We have shipped it on roughly 40 client pages and it is the single most over-performing format we use.

Track competitor citations. When a competitor overtakes you on a tracked prompt, you have 7–14 days to respond before the LLM “settles” on the new ranking. This is the operational reason the weekly report cadence matters. Catch it on Monday, ship a counter-rewrite by Friday, hold position. Catch it at month two — you are already three index refreshes behind.

Treat schema validation as a deploy gate. Run a schema-validity check on every CMS deploy. Roughly 1 in 8 deployments breaks a schema field somewhere on the site. If your AEO programme assumes schema is stable, build the gate.

How to report half-life to a retainer client

Citation count alone is misleading. We have killed two client dashboards because the count rose monotonically while half the underlying citations were already dead.

The metric that actually maps to share-of-voice is “actively cited in the last 7 days”. Run the prompt set Monday morning, count which URLs appear, that is the live ledger. Compare to last Monday’s ledger to see what gained and what dropped.

The reporting structure we use on Scale and Enterprise:

  • Weekly snapshot — actively cited in last 7 days, by URL and by prompt cluster
  • Survival curve — chart of week-over-week retention for the last 12 weeks
  • Drop-off log — every citation that dropped, with hypothesis (competitor, refresh, schema) and proposed action
  • Refresh queue — pages flagged for the next quarterly refresh, prioritized by half-life trajectory

The chart that makes the case to a CMO is the survival curve overlaid with refresh-intervention markers. You can see the half-life extension at the moment the refresh shipped. It is the cleanest visual proof that AEO is not a one-time push.

When to let a citation die

Not every citation is worth defending. Triage.

A citation on a long-tail informational prompt with no commercial intent is worth less than the engineering hours it takes to defend. We let those die. Same for citations on prompts that map to product lines we have deprecated, or markets we are no longer targeting.

The triage rubric — three questions. Does this prompt cluster map to revenue. Is the citation defending a positioning we still want. Is the refresh cost less than 25% of the original page cost. Two yeses out of three — defend. One or none — let it go.

Roughly 30% of the citations we earn for a typical Scale client are eventually triaged into the “let it die” bucket. That is not a failure. That is portfolio discipline.

The single biggest mental shift teams need to make — citations are not backlinks. A backlink, once earned, persists until something actively breaks it. A citation expires by default. The retrieval layer rolls forward every day; if you are not in the active cohort this week, you are not cited this week.

Treat citations like a magazine subscription, not a deed of property. The subscription requires renewal — on a quarterly cadence for evergreen, monthly for tactics, weekly for news. A retainer that does not include refresh work is selling you the first wave of citations and letting the second wave die.

That is also why “set and forget” AEO does not work. Every page is a perishable asset. Some perish in 10 days, some in 6 months, but every one of them perishes if you stop maintaining it. The programmes that compound are the ones built around that fact, not against it.

If you have an active AEO programme without a refresh cadence — start one this quarter. The numbers in this post should give you the budget conversation. If you do not have an AEO programme yet, the Starter audit is the cheapest way to see where your half-life curve is sitting before you commit to anything bigger.