Answer engine optimization · Original data
69% of AI citations change by the next day. We watched ours do it.
Eleven days ago, we ran forty-six questions through ChatGPT for a Miami med spa audit. One asked where to find the best laser hair removal and skin treatment in Miami. ChatGPT pointed to two sources, both from lavishlasermedspa.com: its homepage and its laser hair removal page.
Tonight we ran that exact same question again, same wording, same location setting, same live interface. Back came three sources: MedSpa Scout, DiscoverMedSpa, and BeautyPlaces. Not a single mention of lavishlasermedspa.com anywhere in the results. The overlap came out to zero, eleven days later, on an identical question.
We nearly dismissed it as single-query noise. Then we found a study built precisely to measure this.
The study that caught what we tripped over
GetMentions AI ran 2,398 real-world brand queries across 56 accounts for seven straight days in June 2026, logged out, fixed location, with fresh prompts run daily across ChatGPT, Google AI Mode, Perplexity, and Gemini. That yielded 67,144 answers and 530,875 individual citations across 181,225 distinct URLs, enough for 46,259 day-over-day comparisons of identical questions.
The headline: 69% of sources in a typical answer differ from what the same question returned yesterday. The actual list of cited websites shifts, regardless of the article or how it's framed.
That rate changes sharply from one engine to the next. Gemini rotated sources 88.3% of the time day-to-day, ChatGPT 79.2%, Google AI Mode 75.9%, Perplexity 44.4%. Perplexity is the outlier the other way too: 66.8% of its sources survived to tomorrow, versus 40.9% for ChatGPT, 40.4% for AI Mode, 28.1% for Gemini.
Stretch to a full week and nearly nothing sticks. Sources cited all seven days: 11.1% for Perplexity, 2.6% for Google AI Mode, 1.1% for ChatGPT, 0.4% for Gemini. A source that gets you cited Monday has roughly a one-in-a-hundred chance of still being there Sunday, if the engine is ChatGPT.
It's not concentration. It's churn.
We wrote about citation concentration three weeks back: a handful of brands and directories grabbing outsized citation share in any niche. This study measures something else, and the two findings sit deliberately at odds.
Across 55,387 distinct domains in the GetMentions dataset, the top 10 combined account for only 9% to 14% of citations per engine. That's a broad base rather than a narrow one. What's unstable isn't which domains matter generally. It's which specific ones surface for your particular question on any particular day.
Editorial and business pages comprise 65% of Perplexity's citations and 81% of Gemini's, and even that dominant category retains merely 26% to 43% of its own citations day-to-day. The winning category stays roughly constant. The individual pages inside it keep trading places.
A second study, a longer window, identical shape
We didn't want one vendor's numbers carrying the piece, so we verified with another measuring the same thing more slowly. Digital Authority Partners tracked 1,127 unique URLs cited across ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews for 30 queries, collected in three waves spaced 14 days apart over six weeks.
66% of URLs cited in week one vanished by week four. Of 1,127 URLs tracked across all three waves, only 119 appeared every single time, about 10.6% of the total. Four-week retention by engine: Perplexity held 44% of sources, Copilot 34%, ChatGPT 31%, Google AI Overviews 27%, Gemini 11%. Cross-engine agreement was thin too: at most 17% domain overlap between any two platforms, and 78% to 85% of cited domains appeared on only one platform.
That last figure matches what we'd already found in our own testing. We ran eight owner-anxiety queries through ChatGPT and Gemini side by side in September and found zero shared domains on seven of eight, five shared domains total out of 134 combined. Two vendors, two different measurement periods, same basic truth: these engines aren't working from a shared shortlist. Each runs its own search, largely independent, every single time.
What this means when you're the business being judged
None of this suggests citations are random or effort wasted. The category getting cited stays fairly stable, shown in the 9%-14% top-10 share and the 65%-81% editorial share holding firm. What moves is which specific page inside that category gets pulled today, and that's largely outside your control. A directory shuffles its listing order, a competitor updates a page, the model's retrieval samples a different index slice. None of it reflects whether your business improved or declined.
So a one-time audit is a snapshot of something already in motion. Getting cited once and never checking again tells you what happened the day you looked, and nothing about today. Businesses appearing consistently aren't the ones who found a hack. They're the ones with durable signals spread widest: accurate structured data, review counts that keep climbing, NAP details matching everywhere, service pages specific enough to survive being pulled into multiple answer types. That's what persists while individual citations churn underneath.
Want to see this yourself? Run the same question through ChatGPT twice, a week or two apart, and watch what actually changes. Usually it isn't your ranking. It's the sources beneath it.
Method: live re-check of one question from our 12 September LUX MedSpa Brickell audit (46 questions, chat_gpt/llm_scraper), re-run 23 September through the same tool. External data from GetMentions AI's day-over-day citation study (2,398 queries, 56 accounts, 7 days, June 2026) and Digital Authority Partners' AI Visibility Gap study (1,127 URLs, 30 queries, 3 waves over 6 weeks), both verified directly against publishers' own pages. Both are vendor research rather than peer-reviewed studies, disclosed as such.