Citation Decay: How Long AI Answers Keep Citing Your Content

Citation decay means that AI answers in ChatGPT, Perplexity or Google’s AI features cite a source less often as time passes. The figures most people quote come from two vendors, Scrunch and Stacker: between September 2025 and March 2026, citation activity for the average source halved in about 4.5 weeks. Both sell AI-visibility services, and neither tested whether updating content slows the decline.

What is citation decay?

Citation decay is the decline in how often AI search systems cite a source they have cited before. It is about the linked sources shown in or next to an AI answer, not about a position in a classic list of search results.

That difference matters. An AI answer is built from pages the system retrieves for that particular question. For AI Overviews and AI Mode, Google describes retrieving relevant pages from its search index, reviewing them and showing them as links next to the answer, often after running several related sub-queries (query fan-out). Which pages end up as sources can therefore shift from one request to the next.

The numbers usually quoted on this come from Scrunch and Stacker. Scrunch expresses the decline as a half-life: the number of weeks until half of the citations of a group of sources that first appeared in the same week have disappeared.

How long do ChatGPT, Perplexity and Google cite a source?

Going by the published analyses, for weeks rather than months. Scrunch and Stacker jointly analyzed citations from AI platforms between September 2025 and March 2026 and published the results in two blog posts on March 26, 2026. Across all sources, citation activity halved in about 4.5 weeks, with clear differences between platforms.

Platform	Half-life according to Scrunch	Half-life according to Stacker (domains outside its own network)
ChatGPT	3.4 weeks	3.4 weeks
Google AI Mode, Gemini, AI Overviews	4.3 to 4.8 weeks (one shared range)	4.3 / 4.6 / 4.7 weeks
Perplexity	5.8 weeks	5.7 weeks
All sources	about 4.5 weeks	about 4.5 weeks

According to Scrunch, the values vary less by industry than by platform. Sources from insurance (5 weeks) and financial services (4.8 weeks) held on longest, while healthcare (4.1 weeks) and retail and e-commerce (4.3 weeks) turned over fastest.

A second figure often gets shortened in the retelling: domains in Stacker’s own network, which Scrunch describes as more than 4,000 news publishers, were cited for nearly 10 weeks instead of about 4.5. Stacker reports this as a 2.1x durability advantage. It compares domains, not updated pages with unchanged ones, and Stacker itself writes that it can’t claim causation.

How reliable are the figures?

Useful as an order of magnitude, not as a forecast for your own pages. The main limitations are stated in the posts themselves or follow from them:

• Vendor data: both texts are company blog posts, not peer-reviewed studies. Scrunch sells AI-visibility monitoring, Stacker sells content distribution through its publisher network – the very network that comes out ahead in the comparison.
• Averages across many questions: Scrunch stresses that the figures describe persistence across many prompts and weeks, not what happens when you ask the same question again and again.
• Unclear sample: neither post says which questions, languages or countries were analyzed. Whether the figures hold for non-English queries or for your market is open.
• Inconsistent details: even the two posts about the same data give different key figures – 3.5 million citation events and five named platforms at Scrunch, more than 3 million events and six platforms at Stacker.
• No causes, no test: the data show that citations decline, not why. Whether updates change anything was not examined.

Scrunch does call the 4.5 weeks a “refresh window”. That is a recommendation derived from the measured decline, not the result of an experiment comparing updated pages with unchanged ones.

Do AI search systems prefer fresh content?

Google and Bing say officially that freshness plays a role in their AI answers. Neither says how much it counts compared with relevance and quality.

• Google: according to Google, its generative AI features are rooted in the core ranking and quality systems of Search. Retrieval-augmented generation is used to improve the “quality, accuracy, and freshness of AI responses” by retrieving “relevant, up-to-date web pages” from the search index. For classic Search, Google describes dedicated freshness systems that show fresher content where people would expect it, such as reviews of a film that has just been released.
• Bing: Microsoft writes that accurate and up-to-date content is important for being included and cited in AI-generated answers. The sitemap’s lastmod field and IndexNow are meant to help changes reach search results and AI answers faster.
• OpenAI and Perplexity: their crawler documentation names one prerequisite – allowing the search crawlers OAI-SearchBot and PerplexityBot. It says nothing about the role of freshness.

Research offers a hint, but no proof for everyday use. In a lab experiment (a preprint by Fang and colleagues, last revised in September 2026), the authors added artificial publication dates to text passages. Seven language models asked to rank the passages by relevance consistently pushed the “newer” ones up; larger models weakened the effect but none removed it. That shows a date can sway a model. It does not measure how ChatGPT, Perplexity or Google handle dates in production.

Does updating the date help?

Only if the content has substantially changed. Google explicitly lists changing the date of pages to make them seem fresh, when the content has not substantially changed, as a warning sign.

Just as clearly, adding lots of new content or removing older content mainly to make a site seem “fresh” won’t help its rankings, Google says. Dates also follow clear rules:

• Visible and labeled: Google recommends a prominent date labeled, for example, “Published” or “Last updated”.
• Structured data: datePublished and dateModified belong in Article or BlogPosting. The visible date and the structured date should match.
• No future dates: the date describes when the page was published or updated, not the event the page is about.
• Sitemap: Google only uses lastmod if it is consistently and verifiably accurate, reflecting the last significant update to the main content, structured data or links. Bing asks you not to set lastmod to the time the sitemap was generated.

What keeps content citable?

There is no proven recipe that keeps citations coming. What is clearly documented are the prerequisites without which a page can’t be cited at all, and the kind of content Google recommends.

• Be reachable: to appear in Google’s AI features, a page must be indexed and eligible to be shown with a snippet, and the site must be included in the generative AI features setting in Search Console. OpenAI and Perplexity recommend allowing their search crawlers in robots.txt.
• Answer in your own right: Google advises content with its own point of view rather than summaries of what is already available elsewhere, organized into paragraphs and sections with clear headings.
• Stay verifiable: give figures, deadlines and versions with their source and date. When a primary source changes, you can see at once which statement needs checking.
• Report changes: after a real revision, serve the sitemap with an accurate lastmod and submit the URL via IndexNow. That reaches Bing and other participating search engines; Google is not on the list of supporters IndexNow names.

How to improve your chances of being named in ChatGPT and Perplexity in the first place is covered in ChatGPT SEO: How to Get Your Brand Cited in AI Answers. How SEO, AEO and GEO differ, and what is actually proven, is explained in Generative Engine Optimization vs. SEO and AEO.

What does an honest maintenance routine look like?

A good routine follows how quickly a page’s facts change, not a rigid calendar. Here is one way to set it up:

1. Pick the pages: list the pages you want to be cited for on your customers’ most important questions – service pages, guides, articles with deadlines or figures.
2. Judge the pace of change: check pages with legal deadlines, prices, model or version names roughly monthly, foundational articles quarterly, and everything else when something happens.
3. Check against primary sources: compare every figure and deadline with the original source, fix what is outdated, replace dead links and add new developments when they change the answer.
4. Show the change openly: only set the update date when something substantial has changed. A short note on what changed helps your readers too.
5. One date, one source: generate the visible date, dateModified and lastmod from the same field so they can’t drift apart.
6. Report the change: serve the updated sitemap and submit changed URLs via IndexNow.
7. Watch your citations: ask a fixed list of questions in ChatGPT, Perplexity and Google at regular intervals and note which sources are named. Bing Webmaster Tools’ “AI Performance” report shows how citations of your pages in Copilot and Bing’s AI summaries change over time. More on this in Measure AI Traffic in GA4: ChatGPT, Perplexity & Co.

This routine keeps your content correct. Whether it also extends citations, you’ll only see once you compare your own observations over a few months.

What is not known about citation decay?

Quite a lot. These questions in particular are open:

• whether regular updates extend citations – Scrunch and Stacker did not test it;
• why citations decline – both describe the decline but write that they can’t establish a cause;
• whether the figures apply to other languages, markets and individual industries;
• how heavily Google and Bing weigh freshness against relevance and quality;
• how ChatGPT and Perplexity treat freshness, since their crawler documentation doesn’t address it.

Anyone promising you a fixed number of extra weeks of citations from a refresh can’t back it up with the published data.

Honest dating can be built into the website itself: when the visible date, structured data and sitemap are generated from the same field, and your team can maintain content without a developer, upkeep stays a routine instead of a project. We build websites and relaunches like that, for example with a headless CMS, as part of our software development service.

Sources

1. The half-life of AI citations: What 3.5 million citation events taught us about AI’s memory – Scrunch, March 26, 2026
2. Most AI Citations Fade in Weeks. Distributed Content Lasts Twice as Long. – Stacker, March 26, 2026
3. Google’s Guide to Optimizing for Generative AI Features on Google Search – Google Search Central, last updated July 2026
4. Add a Byline Date to Google Search Results – Google Search Central, last updated December 2025
5. Creating Helpful, Reliable, People-First Content – Google Search Central, last updated October 2026
6. Keeping Content Discoverable with Sitemaps in AI Powered Search – Microsoft Bing Webmaster Blog, July 31, 2025
7. Introducing AI Performance in Bing Webmaster Tools Public Preview – Microsoft Bing Webmaster Blog, February 10, 2026
8. Do Large Language Models Favor Recent Content? A Study on Recency Bias in LLM-Based Reranking – Fang et al., arXiv preprint, version 2, September 21, 2026