• AI visibility

Guide · AI search

Why do AI search engines cite different sources every time I check?

Because AI search engines choose their sources partly at random each time they answer, even when nothing on the web has changed. In a 45-day study of four engines, about 65% of cited sources changed from one day to the next, and asking the same question minutes apart produced almost as much change. The churn is built into how these systems work, so it is not a sign that your content suddenly got better or worse.

By Underneath Team Guide Updated

The short version

  1. In a study of four AI engines over 45 days, roughly 65% of cited sources changed from one day to the next (Schulte and colleagues, University of St. Gallen).
  2. Re-asking the same question within 24 hours produced similar churn, with only 32% to 43% of sources shared, so most of the change is built-in randomness rather than news or website edits.
  3. Each engine has its own level of stability: repeated runs shared about 0.30 of cited websites on Gemini, 0.40 on ChatGPT search and 0.50 on Perplexity (Sielinski, who works for an AI visibility company).
  4. ChatGPT ran a mean of 3.7 hidden web searches per buyer question in our study, and none of 509 searches repeated the user’s question word for word, which gives the answer many ways to drift.
  5. Which engine you ask matters far more than which day: one vendor study found same-engine sources a day apart were about 42 times more alike than two engines’ sources on the same day.

How much do AI citations change from one check to the next?

A lot: most cited sources are different from one day to the next. Schulte and colleagues (opens in a new tab) tracked ChatGPT, Gemini, Google AI Mode and Perplexity every day for 45 days in early 2026, using German-language shopping and service questions sent from Swiss servers. Across 4,044 pairs of consecutive days, only about 35% of cited sources overlapped, so roughly 65% changed overnight.

Google’s AI Overviews, the AI summaries at the top of Google’s results, change too. Xu and colleagues (opens in a new tab) compared repeat appearances of the same trending searches on different days. When both showed an AI Overview, only 0.24% cited exactly the same set of pages, and only 0.06% had exactly the same text.

A vendor study by Tannenbaum (opens in a new tab) found the same scale of movement in a small benchmark of 15 prompts about AI visibility software. Between 5 and 6 June 2026, 67.0% of cited web addresses turned over for the same prompt on the same engine.

Is the churn caused by websites changing?

Mostly not: asking the same question minutes apart produces nearly as much change. Schulte’s team re-asked each question up to 10 times within 24 hours. Source overlap between those near-simultaneous runs averaged 32% to 43%, the same range as the day-to-day figures. The authors conclude that the engines’ own randomness accounts for most of the instability.

Sielinski (opens in a new tab) tested the other explanation directly. Sielinski fingerprinted the content of cited pages each day and found most pages did not change while their citation shares swung up and down. The conclusion: the variability is structural, not driven by content.

So a page can drop out of an answer tomorrow without anyone touching it, and come back the day after.

Where does the randomness come from?

From several steps: the searches the assistant writes, the pages it picks, and the answer it generates. Before answering, an assistant writes its own web searches. In our hidden-searches study, ChatGPT ran a mean of 3.7 searches per buyer question, Gemini 1.9 and Claude 0.76. None of the 509 searches repeated the user’s question word for word.

Each of those searches can return different pages, and the assistant then chooses which to use. ChatGPT does not always search at all: in the Swiss study, 57.8% of its runs had no citations, because it answered some questions from memory.

Fixing the settings does not remove the variation. A survey by Martinez (opens in a new tab) reports a peer-reviewed audit in which repeated runs with the randomness setting at zero still changed 9% to 28% of decisions. Small wording changes matter too. Grossman and colleagues (opens in a new tab) found that edits as minor as “what is” versus “what’s” cut the similarity of AI Overview sources by 28.99% compared with a plain rerun.

Do some AI engines change more than others?

Yes, and each engine has a fairly steady level of churn of its own. Sielinski sampled Gemini, ChatGPT search and Perplexity on three consumer topics over nine days. Repeated runs of the same question shared about 0.30 of their cited websites on Gemini, 0.40 on ChatGPT search and 0.50 on Perplexity, however many sources an answer cited.

EngineChange in cited web addresses, 5 to 6 June 2026
ChatGPT82.6%
Microsoft Copilot76.6%
Perplexity45.5%

Source: Tannenbaum, 15 prompts, one day-to-day comparison; the author founded an AI visibility software company and notes the change may partly reflect the collection system.

Google’s AI Overviews are less stable than Google’s ordinary results. Over two months, a peer-reviewed audit summarized by Martinez found only 18% of pages overlapped in AI Overviews, against 45% in organic Google. In Grossman’s tests from two US cities, two runs of the same AI Overview scored 0.69 on a 0-to-1 similarity scale, against 0.86 for the ordinary results.

Does a changing source list mean your brand changes too?

Not one for one: brand mentions are steadier than sources, but still move. In the Swiss study, consecutive days shared 45% to 59% of brands, against 34% to 42% of sources.

The link between sources and brands is loose. In our consistency study, 166 pairs of Perplexity runs cited identical web addresses, yet the brand list changed in 91.6% of them. The same evidence can produce a different shortlist. Engines also differ in how steady their brand lists are; see which AI engine is most consistent.

Switching engines changes far more than waiting a day. In Tannenbaum’s benchmark, 84.9% of pairs of engines answering the same prompt shared no cited web address at all. 96.4% of addresses appeared on only one engine. A same-engine source list from the next day was about 42 times more alike than another engine’s list from the same day.

What should you do about it?

Read every AI citation check as one draw from a moving system, not as a fixed position.

  1. Do not react to a single disappearance. Re-check the same question several times, on several days, before deciding anything changed.
  2. Track each engine separately. Their stability differs, and their sources barely overlap. If you need one number, see how to combine engines into one score.
  3. Watch your brand mentions as well as your cited pages. Mentions move less, and they are what buyers read.
  4. Keep the wording of tracked questions fixed, since small edits shift sources on their own.
  5. Aim to be one of the sources an engine returns to often. In these studies, a small core of sites recurs while the rest rotate.

If you want help setting up steady tracking, see our generative engine optimization service.

What does the research not tell us yet?

The studies measure how much citations change, but none can see inside the engines to say exactly why.

  • Most data come from short windows, from nine days to two months, so seasonal shifts and model updates are not well covered.
  • The Swiss study used German questions from Swiss servers, and Sielinski studied three consumer topics; other markets may differ.
  • Several key papers have commercial ties: Sielinski works for IQRush, Tannenbaum founded Aiso Boost, and Schulte is also affiliated with Aurora Intelligence. Tannenbaum’s day-to-day comparison may partly reflect collection changes.
  • No study links citation churn to what buyers do, so we do not know how much a rotating source list costs a brand.

Frequently asked questions

Is it normal for ChatGPT to cite different websites for the same question?

Yes. In one 45-day study of four engines, about 65% of cited sources changed from one day to the next, and one vendor benchmark measured 82.6% turnover for ChatGPT between two days.

Why did my page stop being cited by Perplexity?

Possibly for no reason related to your page. Perplexity was the steadiest engine in one study, but repeated runs still shared only about 0.50 of their cited websites.

Are Google AI Overviews more stable than ChatGPT?

They are less stable than Google’s ordinary results. Over two months, only 18% of AI Overview pages overlapped, against 45% for organic results, in a peer-reviewed audit.

Will checking again a few minutes later give different sources?

Often, yes. When the same questions were re-asked within 24 hours, runs shared only 32% to 43% of their sources, about as little as on different days.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work