AI Citations Vanish Daily, but Three Quarters Return in 20 Days

By Nathan Brooks

About 75% of AI citations that vanished during a 20-day Ahrefs study returned within the experiment, even as 50% to 83% disappeared by day two.

Ahrefs marketing researcher Mateusz Makosiewicz said: “Treat citation volatility as normal.” He added: “Don’t judge AI visibility from one check.”

These were Ahrefs observations, not platform announcements. The study measured correlation, not causation.

What did the test cover?

Xibeijia Guan asked five tools the same questions daily for 20 days, then followed every website cited on day one.

Tool	Questions	Day-one sites	Gone by day 2	Lasted 7+ days	Lasted 20 days	No-source answers	Typical / maximum citations
Google AI Overviews	563	4,830	50%	18%	6.6%	0.9%	10 / 39
Google AI Mode	1,996	18,009	58%	19%	6.8%	0.5%	9 / 60
ChatGPT	3,173	11,225	59%	13%	4.0%	0.7%	4 / 20
Gemini	1,276	1,994	81%	0.3%	0.2%	64%	0 / 11
Copilot	823	1,447	83%	1.9%	0.5%	53%	0 / 15

The tools produced 37,505 day-one website citations. A Brand Radar check covered 25 AI-visibility and marketing-conference questions, plus 69 SEO and marketing-tool questions. It found similar churn, but a lower return rate of about two in three.

How quickly do citations change?

Excluding source-free answers, all five tools converged at 50%–60% loss by the next cited answer. Gemini and Copilot each fell to 58%; their seven-day rates rose to 9.5% and 15.3%. Google’s tools and ChatGPT changed by about one point or less.

ChatGPT repeated its exact prior searches only 26% of the time. Cited sites changed 81% when searches repeated and 91% when they changed, suggesting substantial churn after search, during page selection.

About 75% of lost citations returned within 20 days; roughly half returned within three days. AI Overviews and ChatGPT saw about 40% return after one missing day. Only 20%–27% of day-one websites never returned.

Citation duration before first disappearance	AI Overviews never returned	AI Mode never returned	ChatGPT never returned
1 day	34%	41%	35%
2–7 days	8%	10%	4%
8–14 days	4%	2%	0.7%

Next-day returns were 14%–25% for one-day citations, versus 69%–87% after eight days. New Brand Radar citations were less stable: about 70% vanished next day, and 34%–45% returned. End-of-study censoring may understate returns; researchers tracked only first disappearances.

What do empty answers and brand names change?

Report B found 68% of Gemini and Copilot losses followed an answer without citations, versus 1% on Google’s tools and usually 11% on ChatGPT. ChatGPT reached 58% from August 19 to September 7. Apparent Gemini and Copilot comebacks merely meant sources returned in 44% and 52% of cases.

Only 4% of Gemini and 27% of Copilot questions cited sources across all 20 days, limiting that comparison.

Tool	Cited, brand absent	Citation gone, name retained	Citation retained, name gone
Google AI Mode	84%	16%	11%
Copilot	81%	31%	3%
Google AI Overviews	77%	12%	12%
Gemini	60%	40%	4%
ChatGPT	54%	13%	9%

On every tool, 50%–70% of cited websites appeared for one day, versus one in five brand names. Researchers used AI to group name variants. Glen Allsopp’s manual check of 55 brands found 30%–36% of apparent losses were the same brand under a different name.

Which pages keep citations longer?

Pages ranking for the same question were about seven times more likely to retain citations for a week: AI Overviews recorded 29% versus 4%, AI Mode 28% versus 4%, and ChatGPT 11% versus 4%. In AI Mode, about 40% of top-three pages survived a week, versus 20% at position six or lower.

Citation duration	AI Overviews monthly visits	AI Mode monthly visits	ChatGPT monthly visits
1 day	8	1	0
2–7 days	27	10	2
8–14 days	47	44	17
15+ days	88	124	17

An earlier study of 1.9 million AI Overview citations found 86% of top-three cited pages ranked in Google’s top 100, at median position three.

Typical cited-site Domain Rating was 64–78 regardless of citation duration, while URL Rating stayed near five. Long-lived citations often had few referring domains. No traffic, backlinks or low authority clustered among short-lived citations, but additional authority did not materially improve stability. Backlinks can still support rankings, which had the stronger relationship.

About 75% of AI Mode and ChatGPT citations did not rank for the query; nearly half of AI Overviews citations did not. The analysis covered 230 questions. Rankings and traffic overlap, and smaller position differences may be random variation.

Title-description match scores rose from 0.54 for one-day citations to 0.65–0.68 for 15-day citations on Google’s tools. ChatGPT scored 0.54 versus 0.56 for one and 2–7 days; a clearer pattern appeared only among the 5% lasting over a week.

Two outside-study freshness cases favored updated pages: a competitor’s changing pricing page overtook Ahrefs, and a product page with fresher data drew 35 AI citations within days.

Which questions hold citations longest?

For branded questions, the named brand’s site stayed cited for all 20 days in 33% of AI Mode cases, 28% for ChatGPT and 28% for AI Overviews. Those rates were four to seven times the average cited-site rates of 6.8%, 4.0% and 6.6%.

On Google’s tools and ChatGPT, the named site appeared on 60%–68% of days, versus 16%–20% for other sources. It retained its citation the next day about 75% of the time; Gemini reached 60%. Other sources returned at 33%–40%, or 50% on Copilot.

For 22 Ahrefs prompts tracked from August 28 to October 4, 2026, factual questions cited the Ahrefs site far more often:

Ahrefs prompt group	Questions	AI Overviews	AI Mode	ChatGPT
Factual	7	97%	100%	100%
Reviews	3	5%	5%	97%
Comparisons	12	0%	5%	71%

Glen Allsopp’s broader data showed frequently cited brands and domains changed only 11%–14% between days, versus 61%–78% for other sources.

YouTube held all-20-day AI Overview citations across 33 questions, while Reddit did so across 12. YouTube often rotated about six videos within the same slot; Reddit more often retained one thread. Major publishers and government sites accounted for one quarter of all-20-day citations on questions that did not name them. ButterflyMX, AppFolio and Buildium held citations for multifamily-property-software questions; Crassula did so for cloud banking.

On ChatGPT and AI Mode, 64%–76% of cited pages changed daily regardless of question type. Named-brand questions were most stable, while “best X” questions were least stable. Glen found the same first source in 56%–79% of health questions, compared with 13%–23% for fashion questions. The researchers suggested stronger competition may explain the difference.

How should teams measure citations?

Count citation days across the last seven days, track brand mentions separately, and segment factual, review, recommendation and discovery questions. Do not credit one comeback to a page edit; inspect a citation missing for two or three weeks.

Ahrefs Brand Radar’s Cited pages report filters platforms, separates owned, competitor, other and watchlist pages, and charts citation history. Metrics include pages found and cited, bot visits, AI-search visits, organic traffic, and publication or update dates. Its watchlist functions like rank tracking, while tags separate prompt groups.

API and MCP access support custom dashboards, and Agent A can calculate watchlist citation rates. The AI Responses report can identify branded-prompt gaps by setting “Prompt contains” to the brand and “Your brand” to “Not found.”

Teams should monitor seven-day citation frequency, brand mentions, rankings and organic traffic; model updates may materially change the volatility.