AI Citation Decay Isn’t Random. We Watched Which 70% Vanish.

Why Citation Volatility Looks Random — But Isn’t

One of our clients, a cybersecurity firm, has a group of nine blog posts on the same topic, by the same authors, on the same site. One of them has 17 other posts linking to it with a sentence that says how they relate. It was cited by AI engines on all five of its tracked questions, five checks in a row.

Three other posts in the same group started with three such links each. Over the same weeks, AI engines cited them on about one question in five.

Same site, same topic, same authors, same weeks. If citation volatility were weather, it would have rained on all of them the same way.

The belief

“Citation volatility is random and can’t be managed.”

The industry calls it citation decay: 40–60% of AI citations turning over every month, and “70% of AI visibility gone in six months.” Every practitioner with their own tracking panel sees the treadmill: you win a citation, you lose it, you win a different one.

So the advice is to treat it as weather. Keep publishing, keep tracking, and don’t read too much into any one week.

What almost nobody measures is survival: which citations last, and what the ones that last have in common. That’s what we measured.

How we tested it

The VizzEx AI Citation Study tracks 10 live websites, including two famous marketing publishers we left untouched as a control group, on about 1,200 real buyer questions across Google AI Overviews, ChatGPT, Claude and Perplexity. One of the ten sites is ours, vizzex.ai. Every answer is stored, so we can follow each citation from check to check: when it appears, whether it’s still there next time, and what replaced it when it’s gone.

First: the page has to say something new

Before anything else matters, a page has to carry what we call a Unique Information Delta (UID): an information gain, something it says that the pages already in the answer don’t.

On every site’s questions in the study, 88 to 99% of the pages that got into an AI answer said something the pages already there didn’t. That was true for the sites using VizzEx, for the two famous publishers’ topics, and for a client site with broken page templates. Content that adds nothing doesn’t get in anywhere, no matter how it’s built or linked.

That’s the price of entry. So what keeps a page in the answer once it has earned its way in?

AI Engines Cite Your Site, Not Just Your Page — Here’s the Data

Counted by page, a lot of what looks like citation decay isn’t decay at all. In SEO, page competes against page: your URL ranks or it doesn’t. AI search works differently, so our study counts it differently. We track whether your site is cited on a question, not which of your pages is.

If an engine cited one of your posts last week and a different one of your posts this week, you didn’t lose anything. You still hold the question. A page-by-page report would show a loss and send you off to fix a page that didn’t need fixing.

Page-Level vs. Site-Level Citation Counting: A Real Example

Here’s what that looks like on one question: the cybersecurity client’s question about its own resilience framework, on Perplexity, check by check.

14 different client pages were cited on this question over these nine checks.

Counted page by page, that’s 7 losses in eight weeks, and seven reasons to go fix something. Counted by site, the client was cited at every check. The engine kept choosing the client; it just kept rotating among the client’s pages.

That isn’t one odd question. Here’s the same count across the eight sites using VizzEx, on all four engines:

The eight sites using VizzEx, every tracked question, check to check, through day 56.

How Often a Dropped Page Still Means a Retained Question

More than half the time one of these sites’ pages dropped out of an answer, the site didn’t lose anything. Counted page by page, they lost 25% of their citations from one check to the next—but that number only tells part of the story, because being cited and being recommended are not the same thing. Counted by site, 16%.

The two famous publishers look different. When one of their pages dropped out of an answer, another of their pages was still cited only 11 times in 90. The other 79 times (88%), they lost the question. On the sites using VizzEx, a dropped page meant a lost question 893 times in 2,120 (42%).

Top-Source Retention Rates Across All Four AI Engines

This isn’t just our sites. Every time we check a question, we save the whole answer, including every source the engine cited. So we can run the same test on everyone: take the top source an engine lists for a question, and see whether it’s still cited at the next check.

For each question, we took the top source an engine cited and looked again at the next check. Through day 56, we made that comparison 3,414 times on Google, 5,742 on Claude, 6,526 on Perplexity and 2,572 on ChatGPT.

Counted by page, ChatGPT keeps its top source at the next check only 34% of the time. Counted by site, it keeps it 56% of the time.

So in this study, a citation is lost only when the question goes to someone else’s site. Every number from here on counts that way.

AI Citation Churn Is Real — But Newcomers and Regulars Behave Differently

We measured the churn on every question in the study, check by check.

On each study site’s questions, this counts every other site the engines cited, leaving out the study site itself, so it shows how the engines treat their sources in general. Through day 56, we followed 11,439 newcomers on Google, 9,065 on Claude, 31,443 on Perplexity and 3,248 on ChatGPT.

Think of it as newcomers and regulars. At any check, some sites in an answer are newcomers: they weren’t there at the check before. The rest are regulars: they were. On Google, only 33% of newcomers were still there at the next check, against 69% of regulars. The two columns are separate groups, so they don’t add up to 100%.

A new citation is on probation. Most are gone by the next check: about two in three on Google, a little over half on Claude and Perplexity, three in four on ChatGPT. Regulars are far more likely to stay: 68 to 81% are still there at the next check, though the page cited may be a different one.

The engine matters too: Claude and Perplexity hold on to their sources longer than Google and ChatGPT do. So when a tracking report says “40% of your citations changed this month,” ask three questions before reacting. Were they counted by page or by site? How many were brand new? And which engines were they on?

What Structural Audits Reveal About New Citation Survival

Across the field, nothing we could see in a page’s layout predicted which new citations survived. Pages that scored perfectly on our structural audit were kept at the same rate as pages that failed it. Whatever decides it isn’t visible in how one page is built on its own. That’s where the links come in.

Most Lost AI Citations Aren’t Content Losses — 61% Had No Content Disadvantage

Every time a site in the study lost a citation, we compared the page that took it against the one in the study domain. Did the new page say something the study domain’s didn’t?

Each site’s latest comparison at mid-study (days 42 to 56). All 202 replacement pages in these comparisons were on other sites.

Only about four in ten lost citations went to a page that said something our study domain’s page didn’t. The other six went to pages with no content advantage at all.

What Won the Citation When Content Wasn’t the Difference

So what did they have? Here’s where the sites using VizzEx and the two famous publishers split:

On the sites using VizzEx, the pages that took a citation without saying anything ours didn’t were not as strong structurally as ours in 69 of the 78 cases we could score. More importantly, they included NIST, Merriam-Webster and Cambridge, healthcare.gov, big platform vendors like Palo Alto Networks and Fortinet, the big insurers, and other established companies in each client’s market. They won on standing: they’re the sources AI engines already lean on for those questions. That’s Belief 6 in this series.

It was the opposite for the two famous publishers. Every one of their no-content-advantage losses went to a better-structured page, most of them SEO-tool companies’ blogs.

Structure decides which of your own pages hold. Owned ground decides where the rest of the losses go.

Two Kinds of Links That Affect AI Citations: Authority vs. Semantic Structure

SEO has always been about links, and links mattered in this study. Just not the ones SEO counts. Before we get to which links matter, it’s worth understanding what links from other websites into yours actually measure—and why that metric tells a different story in AI search than it does in traditional SEO.

The second column matters because of what gets cited. 92% of the citations on buyer questions in this study went to blog posts, not to the sites’ other pages. The posts, and how they connect to each other, are what AI engines are choosing between.

Keyword Links vs. Semantic Relationship Links: What’s the Difference?

And links between your posts come in two kinds too.

A keyword link is a word in a sentence made clickable, like “segmentation.” The sentence around it says nothing about why you’d follow it.

A semantic relationship link is a sentence that says how two posts relate and why the reader should care. Here’s a real one from a client’s blog:

“It’s worth understanding why most marketing automation implementations fail before you build your segmentation system on top of one.”

That sentence tells the reader, and the machine, that one post is a prerequisite for the other. A machine reading it doesn’t have to guess the relationship from shared words. It’s usually one of 13 relationship types: one post defines a term another uses, contrasts with another, is the tool for the problem another describes, or is the evidence for another’s claim.

Which 70% of Citations Vanish? They Move Along Semantic Relationship Links

When an engine kept citing a site but switched to a different one of its pages, we checked whether the old page and the new one were connected by a semantic relationship link at the time.

A link counts only if it was already on the page at that check.

The page the engine moved to was linked to the one it left 63% of the time, against 33% if it had picked at random. That was true on seven of the eight sites. The links engines followed most often were the ones between a general concept and a more specific one.

Why Famous Publishers Lose Questions When Pages Drop

The two famous publishers have no semantic relationship links between their posts. Their pages rarely switched, 30 times in the whole study, and when one of their pages dropped out of an answer, they usually lost the question: 79 times in 90.

That’s the VizzEx GEOMesh at work. When your posts are connected by semantic relationship links, an engine that moves on from one of your pages tends to move to another of yours. When they aren’t, it moves to someone else’s site.

An example: one topic lane, ten weeks

Here is the cybersecurity client’s topic lane from the opening: nine posts about CISOs and cyber resilience, all on the same site, written by the same team. “Links in” counts the semantic relationship links pointing into each post from the client’s other posts. “Cited” means at least one of the four AI engines cited the client on that question.

Eight checks, days 21 to 70 of this client’s tracking.

What the Data Shows: Citation Rate by Semantic Links Pointing In

Grouped by links in at the start of tracking:

At day 28, Post A was cited on 5 of 5 of its questions, and most of the 0–3 posts on 1 of 5. Through day 70, Post A held 5 of 5 for five checks in a row, then slipped to 4 and 3 in the last two. The two most-linked posts don’t do equally well, and Post H gained links without gaining citations. Across the lane, the posts the rest of the site points to kept more of their questions.

It’s one lane on one site, so read it as an example. The finding is the one above: when a citation moves, the links decide whether it stays on your site.

Why it works: the dome and the box

Here’s an analogy for the difference semantic relationship links and structure make: two houses in the same storm.

A traditionally built house (a box house) takes the wind on one flat wall. A dome takes it on every strut at once, because every strut braces every other one. That’s why domes survive storms that flatten boxes.

The struts are the semantic relationship links, and each one says what it’s for. A post that sits inside a web of them is one an AI engine can understand quickly: what it’s about, what it builds on, what it’s the evidence for. Every page an engine reads costs it work to understand and confirm; we call that the VizzEx Compute Tax. When another source answers the same question for less work, the expensive one is the one that gets dropped.

Verdict: Belief 5 Did Not Survive — Citation Decay Is Architecture, Not Weather

Citation decay isn’t weather. It’s architecture.

First, the page has to say something new. That’s what gets it in. Then the links between your posts keep the citation on your site. When a citation moves, the links decide whether it moves to another of your pages or to someone else’s.

Watch Belief 5 Explained: Mid-Study Findings Webinar Segment

How to Manage AI Citation Decay: Three Actions to Take Today

1. Take your ten most important pages and ask what each one says that the pages already being cited don’t. If you can’t find it in a sentence on the page, nothing else on this list will help. Fix that first.

2. For the ones that pass, count the semantic relationship links pointing in. How many of your other posts link to each one with a sentence that says how the two connect, not just a clickable keyword? When engines switch pages, they mostly switch along these links. A post nothing links to makes it hard for an AI engine to understand which of your other posts also satisfy the query.

3. Count citations by site, not by page. If an engine moved from one of your posts to another, you didn’t lose the question. Then separate brand-new citations from established ones, and look at each engine on its own. Most new citations are gone by the next check, so judge a new citation after a couple of checks, not one.

Want to learn to write semantic relationship links? Our free quick course, What are Semantic Relationships and Why do they Matter?, walks through it.

What’s next

Belief 5 is one of seven things everyone says about AI search that we put on a scoreboard. Still to come:

• whether head terms are the prize (and the owned ground that took the rest of the losses above); If you want the full argument now, we’ve already written it up: the specific questions your buyer asks are the real prize in AI search, not the broad terms you’ve been chasing.
• whether more content is the answer.

The study ends October 17. At the end we’ll re-score every row on the board against the final data, including the predictions we filed at mid-study, whichever way they go.

Kim Albee is co-founder of VizzEx with Carolyn Holzman. The VizzEx AI Citation Study is a controlled, 84-day study of how Google AI Overviews, ChatGPT, Claude and Perplexity choose, keep and name their sources.