TL;DR
| Definition AI citation decay is a sustained fall in how often AI engines cite a page for a fixed set of prompts, measured over several weeks. |
Not decay Run-to-run swings are normal. Only a drop that holds across weeks and beats your noise band counts. |
| Measure Fix the prompts and engines, rerun weekly for 4+ weeks, then compare baseline citation share with recent weeks. |
Diagnose Check in order: crawlable, indexed and ranking, facts changed, competitor page replaced you. |
| Recover Match the fix to the cause, then keep measuring. A fix is proven only when citations come back and hold. |
Track Per-page citation history shows which URLs AI engines cite and when that changes. |
AI citation decay is a sustained fall in how often AI engines cite a page for a fixed set of prompts, measured over several weeks. Run-to-run noise is the normal week-to-week swing in which sources engines pick. It only becomes decay when the drop holds across weeks and is larger than that noise.
This guide will show you how to differentiate between the two, how to calculate decay using an easy formula, what to look out for first and how to get your page back. Perfect for SEO leads/content owners who have found their page has disappeared from AI results and want a plan for Monday morning.
1What is AI citation decay?
AI citation decay, also called LLM citation decay, is the gradual loss of citations to a page across AI answers. The AI citation decay meaning is directional: the page was cited, then cited less, then rarely, across the same prompts and engines.
Definition: AI citation decay
A sustained drop in a page’s citation share across a fixed prompt set and fixed engines, lasting several weeks and larger than normal run-to-run variation.
It differs from a one-off miss. AI answers are probabilistic, so the same prompt can cite different sources each time (see our guide to normal answer variability). A page missing from one answer proves nothing.
Did you know?
| 34 to 42% overlap of cited sources between consecutive days |
7+ runs per prompt per day suggested for brand monitoring |
A University of St. Gallen study ran 32 prompts across ChatGPT, Gemini, Google AI Mode and Perplexity from January 24 to March 20, 2026. Cited sources overlapped only 34 to 42% between consecutive days. Source: Schulte, Bleeker and Kaufmann, 2026.
Decay vs run-to-run noise
| Signal | Decay | Noise |
|---|---|---|
| Direction | Falls steadily, week after week | Moves up and down around a flat line |
| Duration | Holds for three or more weeks | Reverses within a week or two |
| Spread | Shows across many prompts | Shows on one or two prompts |
| Engines | Often the same drop on several engines | Different engines disagree each week |
| Response | Diagnose and fix | Keep measuring |
2Why do AI engines stop citing a page?
Engines will always try to give new information from new sources. This means there are 4 main reasons why citations are not guaranteed. Firstly, the page may be banned, not indexed or ranked lower than before. Secondly, the information may be obsolete. Thirdly, a rival page may be giving better answers to the prompt. Lastly, external pages which previously mentioned you may no longer be citing you.
Only the first two sit on your own page, which is why a refresh alone often fails to bring citations back. The search engines will sometimes favor the new pages if they are similar, so you may find that your old page gets overtaken by a newer one.
3How to measure AI citation decay
Fix a prompt set and a list of engines, rerun them every week for at least four weeks, then compare baseline citation share with recent weeks. Keep the wording, the engines and the day of the week the same each time. Split the results by engine, because a page can hold steady in one engine while it fades in another.
| 1 | Choose 10 to 20 prompts where the page should be cited, and list the engines. |
| 2 | Run each prompt several times per week, so one answer never decides the result. |
| 3 | Record whether the page is cited, per prompt and per engine. |
| 4 | Repeat weekly for at least four weeks to set a baseline and a noise band. |
| 5 | Compare baseline with recent weeks using the formula below. |
How many runs is enough? The St. Gallen authors advise at least 7 runs per prompt per day for brand visibility monitoring. If that is too heavy, run each prompt at least three times per week and use the spread between those runs as your noise band.
Formula: citation decay rate
Citation share = prompts where the page is cited ÷ prompts tested, averaged across runs.
Decay rate = (baseline share − recent share) ÷ baseline share, using the first two weeks as baseline and the last two as recent.
Working rule: call it decay only if the drop lasts three weeks in a row and exceeds the spread you see between repeat runs in one week.
Worked example: a flat page and a decaying page
These are made-up numbers to show the math, not Wellows data. Both pages start near 30% citation share.
| Week | Page A (noisy, flat) | Page B (decaying) |
|---|---|---|
| Week 1 | 30%
|
30%
|
| Week 2 | 25%
|
28%
|
| Week 3 | 35%
|
22%
|
| Week 4 | 30%
|
17%
|
| Week 5 | 28%
|
12%
|
| Week 6 | 32%
|
10%
|
Illustrative data. Blue is Page A, amber is Page B.
Page A baseline is 27.5% and recent is 30%, so its decay rate is minus 9%: no decay despite the swings. Page B baseline is 29% and recent is 11%, a decay rate of 62%, and the fall holds for four weeks. Page B is the one to diagnose. Start with the pages that matter most to pipeline, since you cannot fix every page at once.
Doing this by hand across engines gets heavy. A tool that logs citations per prompt, such as Wellows LLM Citations, keeps the weekly history for you.
4How to diagnose why a page lost citations
Work through four checks in order, because each one rules out a cheaper fix before you rewrite anything.
- Can it be crawled? A 200 response, not blocked by robots.txt or a login, and the core text can be viewed without scripts.
- Is it being indexed and ranking? The page should be indexed and ranking for your query. Ranking issues tend to come first.
- Outdated information? Compare the information on the page with the current day. Prices, dates, features and figures may have changed. Old facts will lose out to updated facts.
- Replaced by competitor? Go to the pages that the AI engines are citing and look at the differences between their response and yours.
Do the checks in this order and stop at the first one that fails. A blocked page needs a technical fix, not a rewrite, and a stale page needs new facts, not new links.
For the first two checks, our technical SEO checklist covers the basics.
5How do I recover lost LLM visibility?
To recover lost AI search visibility, match the fix to the cause you found, then keep measuring. Updating a page without knowing why it lost citations usually wastes the update.
| Cause found | Recovery step |
|---|---|
| Not crawlable | Remove the block, serve the text in plain HTML and request a recrawl. |
| Not indexed or ranking | Fix the indexing issue, then rebuild relevance with clearer headings and internal links. |
| Facts changed | Update the facts, add a dated source and put the answer in the first 40 to 60 words. |
| Competitor replaced you | Add what the cited page has and yours lacks: data, examples, a clearer answer. |
| Third-party coverage faded | Earn mentions on the domains AI engines already cite for your topic, and keep your brand description the same everywhere. |
Changing only the date on a page is not a fix. Engines cite pages for what they say, so change the substance: new facts, a clearer answer, better evidence.
After the fix, rerun the same prompts for four more weeks. A recovery counts only when the citation share returns and holds, not when it shows up once.
What only Wellows adds
Most tools tell you a brand was mentioned. For decay, what matters is whether AI still cites you as a source, with a link or credit, because that is what sends traffic. Wellows LLM Citations tracks the URLs AI engines cite for each of your prompts across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity and Gemini, refreshes every 24 hours and flags every change in citation status.
This will also separate out those pages that explicitly link to you (or mention you by name) from those that implicitly link to you (via a third-party page about you). This is your shadow reputation. It shows what your brand looks like to an AI engine based on pages that are not yours. If your page falls out of date, these are the URLs that are still linking to you, so you know where to target your outreach.
Performance History will tell you if your page starts to gain citations after the fix has been shipped. You can measure, investigate, fix and measure.
Conclusion
AI citation decay is a trend, not a missing answer. This week, identify the 10 prompts that are giving you the most value on the page, establish a baseline and schedule those to run weekly. Over 4 weeks, you will be able to see if there is a decay or simply randomness, and which of the 4 tests to run first.
