AI Citation Volatility: A 530,875-Citation Study Across 4 Engines

Your brand showed up in a ChatGPT answer yesterday. Someone screenshotted the mention, dropped it in Slack, called it a win.

Here's what the screenshot doesn't show: the citations underneath, the sources the engine pulled from to build that answer. Everyone watches the mentions. Almost nobody watches the sources. We watched the sources, 530,875 of them.

The industry talks about AI visibility in Google's language: your brand "ranks" in ChatGPT, a mention is a slot you hold and defend. But every answer sits on a set of sources, and if visibility worked like rankings, that set should hold still. So we ran the simplest test that could break the idea: ask the same questions, the same way, every day for a week, across ChatGPT, Google AI Mode, Perplexity, and Gemini, and measure how much the sources hold still.

They don't.

Here is the whole study in one line. AI visibility is not a rank you hold. It is a probability, and the rational response is breadth: enough credible pages telling the same story that the engine keeps finding you, whichever sources it reaches for that day.

Note: Row 1 is total change (sources that dropped plus new ones that appeared) & row 2 is how many of day-1's sources are still cited.

How we ran it: 2,398 queries, 4 engines, once a day for 7 days

Same prompt, logged-out default consumer session, fixed location, all four engines, every day. Within a query's seven runs nothing changes but the day, so the movement you're about to see is the engine, not personalization, geography, or session history.

One caveat before the fun part: this is one platform's data, one 7-day window, with source types labeled by a model. Read it as directional, not as the physics of AI search. The shape of the finding is what matters, and the shape is loud. Full protocol and definitions sit in the methodology at the bottom.

Finding 1: 69% of the sources behind the typical answer change overnight

Pool every answer in the study together and the headline is one number: 69% of the typical answer's sources change from one day to the next.

Churn counts both directions, sources that dropped out and new ones that appeared, so it measures total change, drops and arrivals alike. An answer that had sources one day and none the next counts as fully changed. We score each query separately, then average across all 2,398 queries, so every query counts the same.

Engine by engine, they stop looking like versions of the same thing:

Gemini cites the fewest sources and rebuilds nearly all of them every day. Perplexity is the steady one. Four engines, four different machines.

And the averages are not hiding a calm majority. Query by query, median churn is 82% on ChatGPT, 92% on Gemini, 78% on Google AI Mode, and 43% on Perplexity. On the first three, more than nine in ten queries turn over most of their sources between consecutive days, and on Gemini a full quarter of queries shared zero sources from one day to the next. The headline numbers are the typical case, not the work of a few wild queries.

Finding 2: a citation has a one-day half-life on three of the four engines

Churn counts arrivals as well as exits. Survival asks the question a brand actually cares about: of the sources cited on day 1, how many is the engine still citing later in the week?

Look at Gemini: more than 70% of today's sources are gone tomorrow. ChatGPT and Google AI Mode shed roughly 60% overnight. A citation there is one day's coverage, not a slot you hold.

Perplexity is the exception, and it's a real one. It keeps about two-thirds of its sources overnight and nearly half a full week later. On Perplexity, a citation behaves a little like a standing, relative to the other engines. On the other three, "we got cited" is news that expires overnight.

Finding 3: almost nothing survives the week, with 0.4% to 11.1% of sources going all 7 days

Sharper question: is anything constant? For every query, we counted how many of the seven days each source appeared, then pulled the two extremes.

On Gemini, 0.4% of a query's sources are constant. Functionally zero. Google AI Mode touches 40 different domains for a single query over a week and keeps almost none of them around. Perplexity, again, has an actual core: 11.1% of its sources show up every day. Small, but real.

This kills the mental model of a settled shortlist that AI gently reshuffles. The engine is rebuilding most of the answer from scratch each day, around a core that is close to empty.

Finding 4: editorial and business pages are 65% to 81% of every answer, and three engines redraw that layer daily

If the answer re-rolls daily, the next question is which sources get re-rolled. Start with the diet: where each engine's citations actually land.

Editorial plus business pages are the bulk of every engine's answer: from 64.9% on Perplexity to 81.3% on Gemini, with ChatGPT at 65.7% and Google AI Mode at 67.6%. News barely registers (0.3% to 3.4%), and community is a small slice (4% to 10%).

Now lay durability on top: of the sources of each type cited today, how many come back tomorrow?

On ChatGPT, Gemini, and Google AI Mode, the editorial and business layer that makes up most of the answer comes back only about 26% to 43% of the time. The biggest layer is also the least stable one.

That layer is still the surface you must be on: if you're not in it, you're not in the answer. But because it turns over daily, winning it is a breadth-and-freshness game, not one-and-done. Google AI Mode shows why: its news and community citations are sticky (58% to 70%) but a sliver of its diet, while the editorial and business sources that make up two-thirds of its answers come back at only about 30%. An engine can hold a source type perfectly well and still rebuild most of its answer daily, simply because most of what it cites lives in the layer that rotates. Perplexity, meanwhile, retains every type at 64% to 77%, so its diet barely matters to its stability.

You're working two layers at once. The big editorial layer is where the citations are; the durable community-and-news layer is thinner and only some engines have one, but where it exists, each placement lasts longer. You want both. On its own, neither is enough.

Finding 5: 55,387 domains, and no universal shortlist

You can't hold a position. You can't lean on a stable core. So maybe you can target the handful of sites AI trusts?

There is no handful. Across the study, the four engines cited 55,387 different domains.

Even the most concentrated engine, Gemini, needs 335 domains to cover half its citations. The widest, Google AI Mode, pulls from more than 31,000. Reddit is the single most-cited source on ChatGPT and Gemini; YouTube takes that crown on Google AI Mode and Perplexity. But no domain owns much: the top 10 combined are only 9% to 14% of an engine's citations.

So when someone tells you to "get on the few sources AI trusts," they have the shape wrong. It's a two-domain head, a thin shoulder of social platforms, and a tail tens of thousands of domains long. Hunting for a universal canon that does not exist is the losing move.

Finding 6: 84% of the sources for a question are cited by only one engine

Everything so far is churn inside one engine. The last pattern is between them, and it's the one I'd tattoo on every GEO deck.

Across the week, the four engines together cite about 72 different domains for a query. Sort those 72 by how many engines picked each one:

Eighty-four percent of those domains are cited by exactly one engine. Each engine is drawing from its own pool.

Even giving any two engines a full week to coincide barely helps:

And no engine is the exception:

This is the most important number in the study. Treating "AI search" as one surface you optimize once is a category error. The engines do not share a source set or a playbook, so you work them as separate channels. And this is only four engines; add more and the gap gets wider, not smaller.

So what do you actually do Monday

Four moves I would act on.

Stop reading a single answer as a result. One answer is one snapshot, and the sources under it change daily. Track an appearance rate across many runs, per engine, or you are charting noise and calling it a trend.

Treat each engine as its own channel. Any two share about 8% of sources over a week. Pick the engines your buyers actually use and work them separately. A ChatGPT win is not a Perplexity win.

Play both layers. Cover the big editorial layer and keep it fresh; keep earning the durable community and news placements where an engine has them (strong on Google AI Mode, community-led on ChatGPT, near-total on Perplexity, almost flat on Gemini). And don't skip YouTube video mentions: youtube.com is the single most-cited domain on both Google AI Mode and Perplexity in this study, and Google AI Mode in particular can pull from video transcripts, so a good video is a citable source, not just a brand asset.

Quit hunting for the canon. A two-domain head and a 55,000-domain tail means coverage beats targeting. There is no universal shortlist; there is a long list, and you want to be on as much of it as you can.

The engines start almost from scratch every day, so the presence has to be continuous. You don't win a citation once; you earn the odds, and you measure them across many runs, not one screenshot in Slack.

Methodology

About the data: produced by GetMentions AI from first-party AI-answer tracking across ChatGPT, Google AI Mode, Perplexity, and Gemini. Every figure here is computed directly from that dataset over the stated 7-day window.