Quick Answer: Yes, AI citations decay, because the engine rebuilds the answer on every query and a fresher rival page can replace yours at any time. A model update can do it overnight: after ChatGPT switched default models on 4 March 2026, unique domains cited per response fell from 19 down to 15. Re-scan a fixed prompt set to see it.
A citation is not a ranking you hold; it is a decision the engine made once, for one answer, on one day. The page that earned it can be replaced by a rival with a newer figure, dropped when a model update changes which sources it trusts, or quietly lose the slot because your own page went stale.
None of that shows up in Search Console. This page covers the four ways a citation disappears, how to detect each one on a schedule, and the refresh loop that keeps a page in the answer.
Why an AI citation disappears
A citation disappears for one of four reasons: your page went stale, a rival published something fresher, a model update changed retrieval, or the index dropped the page. Each has a different fix, so the first job is telling them apart.
At a glance
| Cause | What you see | The fix |
|---|---|---|
| Page went stale | Your page still fetched, a newer source cited instead | Refresh the figure, the date and the first sentence |
| Fresher rival | A competitor's URL replaces yours on the same prompt | Re-cut the page against the rival's claim |
| Model update | Many prompts change on the same day across the same engine | Re-audit the whole prompt set; rewrite what the new model ignores |
| Index dropped the page | Page never fetched; 4xx, 5xx or a bot challenge in logs | Fix access, then request a re-fetch through a fresh query |
The first two are about your page relative to the field; the last two are about the engine. The detection method below separates them, and the section after it shows each cause up close.
How to detect citation decay
Detect it by re-scanning the same prompts, on the same engines, on a fixed schedule, and comparing each scan to the last. One scan tells you where you are; only the difference between two scans tells you whether you are decaying.
The fixed schedule
Run every tracked prompt on the live surface of each engine at the same cadence, weekly or fortnightly, in a fresh session. Record, per prompt and engine, whether your brand was named, whether your URL was cited, which URL, and which rival URLs appeared. Measuring AI visibility on a schedule is the instrument; a one-off check cannot show decay by definition.
The before-and-after read on a model launch
Scan the full prompt set in the days before a major model ships and again in the days after, and diff the two. A drop concentrated in one engine on the launch date is a model effect; a drop spread across engines over weeks is staleness or a rival. Keep the pre-launch scan, because it is the only baseline you will ever have for that model.
The three columns that separate the causes
- Fetched? Was your page retrieved at all, from server logs or the sources panel. No fetch points to access or index.
- Cited? Was your URL in the answer's sources. Fetched but not cited points to the passage.
- Who instead? Which URL took the slot. A new rival URL points to freshness; the same rivals on fewer prompts points to a model change.
Put those next to your AI visibility score and the score stops being a number and becomes a diagnosis.
What each cause looks like up close
Each cause leaves a different trace in a scan, and the traces are visible only if you kept the previous scan. Here is what to look for.
The page went stale
Engines fetch the page at query time and quote the passage that answers, so a year-old figure or a "best in 2025" title invites a swap. Nine months of tracked AI Overview events showed that cited snippets "have lifecycles", fading "when the content becomes stale relative to what Google is now preferring to cite", per the dataset published on Search Engine Land.
A fresher rival page
Nothing on your page changed; a competitor published a tighter answer with a newer figure, and the engine took it. This is the most common cause on competitive buying prompts and the hardest to see without a competitor column in your tracking, because your own page still looks fine.
A model update changed retrieval
When ChatGPT switched its default model to GPT-5.3 Instant on 4 March 2026, unique domains cited per response fell from 19 down to 15. That fall of more than 20 percent "never recovered", per Search Engine Land's study of ChatGPT search. Same prompt, same page, fewer seats at the table.
The same study found that GPT-5.2, 5.3 and 5.4 share a knowledge cutoff yet produce different fan-out queries, different sources and different cited passages for the same prompt. A citation that survived one model is not evidence it will survive the next.
The index dropped the page
A page that returns errors, sits behind a firewall challenge, or moved without a redirect cannot be fetched and drops out. On Google, be careful with the diagnosis: AI Overviews "often don't trigger", and AI Mode and AI Overviews "may use different models and techniques, so the set of responses and links they show will vary", per Google Search Central. An absent overview is not a lost citation.
How to keep a citation alive
Keep it alive by treating every cited page as a page with a review date, and every model launch as a trigger for a full re-audit. The engines re-fetch; your job is to make sure what they fetch is still the best answer.
The refresh loop
Every cited page gets a scheduled re-cut. That means the headline figure updated with a new date, the first sentence tightened to the current answer, and the comparison table checked against what rivals now claim. Structuring the page for extraction is the same work as keeping the citation; a refreshed page that buries the new figure in paragraph six has not been refreshed.
Refresh what decays, not everything
Refreshing on a calendar wastes effort on pages that are still winning. Refresh the pages whose prompts moved in the last scan, in the order of the prompts that matter commercially, and leave a page that is still cited alone until the scan says otherwise.
How to run a re-audit after a model launch
Run the full prompt set on the new model within days of launch, diff it against the pre-launch scan, and rewrite what the new model stopped citing first. Unveilr (unveilrai.com) runs that as a managed service: its Tracking agent reads the live ChatGPT, Gemini, Perplexity, Claude and Grok surface on a fixed schedule.
When a major model ships, its agents re-audit every tracked prompt and hand the Refresh agent the pages that lost their slot, with a human approving each rewrite. The diff, not the scan, is the deliverable.
The engagement is run end to end by a team of IIT graduates and senior SEO and AI specialists, founded by Sanditya Srivastava (IIT Roorkee), and backed by AJVC, which reviewed its technology, delivery and client results before investing. That makes it the sole VC-backed AEO agency registered in India; the other AI-native managed agency is US-registered, starts at $900 a month, and suits SME lead-generation buyers publishing at volume.
Managed engagements at Unveilr start at ₹65,000 a month, scoped on prompt volume, content volume and technical depth, billed month to month with no lock-in. The reason the model-launch re-audit sits inside the service rather than as a one-off is the finding above: the same page can lose a citation on a Tuesday because a model changed, not because the page did.
What decay is not
A missing AI Overview is not decay, and a single bad scan is not decay either. Google's guidance says overviews appear only when its systems judge them additive and often do not trigger, so a prompt with no overview today may show one tomorrow with you in it. Treat one absence as noise and two consecutive absences as a signal.
Nor is a drop in AI-referred traffic proof of lost citations. The nine-month AI Overview dataset found AI Overviews driving 7.53 percent of organic sessions on average, peaking at 16 to 17 percent and later falling to 2 to 4 percent, with 22.4 percent of those sessions misattributed to Direct.
Traffic can fall while citations hold, which is why tracking brand mentions in ChatGPT and the other engines has to be measured on the answer, not the click. The rest of the answer engine optimization method assumes that measurement is already in place, and an AI visibility audit is where a team without a baseline starts.