The half-life of AI citations: What 3.5 million citation events taught us about AI’s memory

AI citation half-life is the number of weeks it takes for 50% of a cohort's citations on an AI platform to disappear — and for the average source, that window is just 4–5 weeks.

Citations shape AI answers and provide a direct link from those answers to your brand and products. Naturally, teams want to be the source AI links out to — and to stay there. Understanding how long citations last, and which factors extend that lifespan, is now a core requirement for AI search strategy.

In partnership with the team at Stacker, we analyzed 3.5 million citation events across AI platforms — including ChatGPT, Perplexity, Google AI Overviews, Google AI Mode, and Gemini — from September 2025 to March 2026. We tracked cohorts of sources that appeared in a given week, measured how many were still being cited in subsequent weeks, and calculated a half-life: the number of weeks it takes for 50% of a cohort's citations to disappear.

The short answer: for the average source, citation activity drops by half in approximately 4.5 weeks. But that average conceals meaningful variation by platform, industry, and source type. ChatGPT cycles citations in as few as 3.4 weeks, while editorial domains in the Stacker Partner Network (4,000+ news publishers) can hold citations for 12+ weeks on Perplexity. The takeaway for marketers and SEOs: durable AI visibility requires both offense (earning new citations) and defense (retaining the ones you have).

Platform plays a major role in citation durability

Across all sources, citation activity halves in roughly 4.5 weeks, but platforms behave differently:

ChatGPT cycles through sources the fastest at about 3.4 weeks.
Perplexity citations last nearly 70% longer than ChatGPT’s at about 5.8 weeks.
Google’s AI Mode, Gemini, and AI Overviews cluster in the mid-range (about 4.3–4.8 weeks), suggesting a relatively consistent refresh cadence across Google’s AI ecosystem.

Implication: the platform you optimize for strongly determines how often you must re-earn citations. A team targeting ChatGPT needs to refresh content roughly every 3 weeks; the same team on Perplexity can operate on a 6-week cycle — nearly twice the runway.

Industry differences exist but are narrower than platform effects:

Insurance (~5.0 weeks) and financial services (~4.8 weeks) are the stickiest.
Healthcare (~4.1 weeks) and retail & ecommerce (~4.3 weeks) turn over faster, where timeliness and recency likely matter more.

Platform choice can outweigh vertical entirely. A healthcare brand cited on Perplexity (~5.8 weeks) still outperforms an insurance brand cited on ChatGPT (~3.4 weeks), despite insurance being the stickiest industry in the dataset. Platform selection carries more leverage than vertical alone.

Some sources have longer shelf lives

Citations from editorial domains in the Stacker Partner Network (4,000+ news publishers) last roughly twice as long as the average non‑Network domain. This advantage persists across every platform and industry we analyzed.

Perplexity + Stacker Partner domains showed the strongest combination at ~12.3 weeks—over 3x the average non‑Network ChatGPT citation and more than 2x the average non‑Network Perplexity citation.
Financial services partner sites were the stickiest in the dataset at ~10.8 weeks.

This creates a compounding structural advantage: while typical sources must re-earn citations approximately every 4–5 weeks, editorial domains in the Stacker Partner Network tend to hold positions for 10–12 weeks or more — reducing re-earning costs by roughly half or better depending on platform.

What this means for content, SEO, and attribution

Expect decay. ~4.5 weeks is the effective “refresh window” for average AI citations. Build refresh and re‑promotion into your content operations, especially for platforms with faster turnover.
Invest where durability compounds. Perplexity and trusted editorial sources hold citations longer, which can reduce the cost of re‑earning visibility.
Balance durability with demand. A long‑lasting citation is only valuable if the underlying prompts have audience interest. Pair citation tracking with AI search trend monitoring to prioritize what matters.
Align attribution with agent behavior. As AI answers drive more off‑site decisioning, you’ll need to track AI agent traffic and referral traffic together, and correlate visibility shifts with downstream outcomes.

How Scrunch helps you measure, compare, and act

Scrunch’s Monitoring & Insights and AXP give teams a full loop: measure visibility, diagnose issues, prioritize opportunities, and deliver AI‑optimized content back to agents.

Track visibility and decay over time
Monitor brand mentions, citations, sentiment, and prompt placement in 2–3 week windows to separate true trends from one‑off noise.
Compare periods (e.g., Q2 vs. Q1) with side‑by‑side charts to quantify improvement or decline. See how in our guidance on monitoring AI search trends in the platform.
Benchmark against competitors across models
Run the same prompt set across multiple AI platforms, then filter by platform to see where performance diverges.
Track share of voice, brand presence, and citation share for you vs. competitors to see where you’re winning or being displaced.
Find the sources and pages AI is most likely to cite
Use the Citations tab to assess citation consistency—the percentage of AI responses that reference a source—to uncover which domains and page types “stick” in your category.
Prioritize with Influence Score, which multiplies the percent of responses citing a source by the number of unique prompts it impacts. High Influence Score = high leverage.
Diagnose why you lost (or didn’t earn) a citation
Insights automatically flags relevance gaps, metadata inconsistencies, and structural issues.
Site Maps + agent traffic show what AI agents crawl vs. what actually gets cited. If crawled but not cited, run a Deep AI Audit to pinpoint content or structure gaps.
Decide whether to improve or create new content
Generally start with existing pages that AI already crawls; they’re faster to improve than starting from zero.
Create net‑new content when no page addresses the prompt or when a new angle is required to beat a durable incumbent.
Optimize delivery for AI agents with AXP
Ensure AI agents can crawl and parse your content (robots.txt, rendering, load).
Use AXP to deliver an AI‑optimized version of your page to agent traffic—without changing the human experience—so models reliably find, parse, and cite the right information.

Practical workflows and how‑tos

Get step‑by‑step directions to track citations in AI search.
Learn how to serve AI‑optimized content directly to LLMs using AXP.
See how many prompts to track: estimate X topic clusters × 12–15 questions each to get a representative set across the journey. Scrunch supports mixed branded and non‑branded prompts for full‑funnel coverage.
Understand which pages AI cites most often in your category by tracking citation consistency and using Influence Score to focus on sources that move the needle.

Pros and cons of using Scrunch for AI search optimization

Pros

End-to-end visibility: brand presence, citations, sentiment, prompt placement, agent traffic, and competitive share of voice in one platform.
Cross-model benchmarking: measure performance across ChatGPT, Perplexity, Gemini, and Google AI Mode, then zoom into each to tailor strategy to its specific refresh cadence and citation half-life.
Prioritization built in: Influence Score (citations × unique prompts impacted), citation consistency, and automated Insights help focus limited time on the highest-leverage fixes and sources.
Action layer with Agent Pages: deliver AI-optimized content to agents, improving parse quality and citation likelihood without redesigning the human-facing page experience.
SEO-friendly workflows: topic cluster planning, prompt tracking, and page-level audits map directly to existing content and technical SEO processes.

Considerations

Ongoing management required: faster-refresh platforms such as ChatGPT (~3.4-week half-life) benefit from more frequent monitoring and content updates than slower-refresh platforms like Perplexity (~5.8 weeks).
Prompt coverage matters: a representative set of prompts across the funnel is needed; under-sampling can skew citation consistency scores and misrepresent share of voice.
Ecosystem variability: platform behaviors evolve; Scrunch smooths noise with 2–3 week trend windows, but short-term volatility can still occur.
Third-party citations: winning durable editorial citations — particularly in the 10–12 week range seen with Stacker Partner Network domains — typically requires coordinated content and outreach beyond owned pages.

Is Scrunch user‑friendly for SEO and content consultants?

Yes. The workflows align with how consultants operate: define topic clusters, set representative prompts, benchmark visibility across platforms, then use page‑level audits and Insights to prioritize updates. Visual trend lines and side‑by‑side comparisons make it straightforward to show clients where visibility is rising or falling and why. If you’re comfortable with SEO tools and analytics, Scrunch’s learning curve is minimal, with new AI‑specific concepts like agent traffic and citation consistency clearly explained in‑product and in the help content.

How Scrunch measures brand impact vs. competitors

Scrunch provides analytics to quantify impact and compare against peers:

Brand and competitive presence across prompts and platforms.
Citation tracking and share of voice to see who sources AI answers.
AI agent traffic and referral traffic to connect visibility with on‑site behavior.
Filters by platform, timeframe, and topic to isolate what’s driving changes.
Visualizations (e.g., line charts) to compare periods and benchmark performance.

What to do next

Monitor your citation share and brand presence routinely—weekly or biweekly is a practical cadence for most teams.
Prioritize opportunities by Influence Score and citation consistency; balance durable sources with rising demand topics.
Improve pages AI already crawls before creating net‑new content, unless a high‑value prompt lacks coverage.
Fix crawl and parse blockers, then use AXP to reliably deliver AI‑optimized content to agents.

If you’re ready to operationalize this, you can start a 7‑day free trial or book a demo to see how Monitoring & Insights and AXP work together to help you win—and keep—AI citations.

A note on methodology

We tracked cohorts of cited sources across 3.5 million citation events from September 2025 to March 2026, measuring weekly decay rates for each cohort. Short-term volatility was smoothed using rolling averages; smaller source groups were stabilized via partial pooling to prevent outlier distortion. We used 200 bootstrap resamples to estimate confidence intervals and selected decay models based on holdout performance. Half-life estimates represent aggregate citation persistence across many prompts and weeks — not the result of rerunning any single prompt repeatedly.