Citation Volatility: How Much Do the Sources AI Cites Change Over Time?

Get cited in an AI answer once and it is tempting to call it a win. Run the same prompt tomorrow and the sources can change completely. We tracked 28 shopping prompts about athletic shoes across ChatGPT and Google AI Mode, every day for six weeks, and recorded every domain each engine cited. The pattern was consistent: the list of cited sources is far less stable than a single check would suggest. On most prompts, the set of cited domains turned over by more than half from one day to the next. We call this citation volatility, and how much of it you see depends almost entirely on what the user asked.

Key takeaways

This is a study of how AI citations move, run on a public consumer brand so the numbers are relatable and nobody's competitive position is on display. Every method below works the same on your own brand.

Picture the same question asked of the same AI engine on five days in a row. Each day it runs a web search and cites a handful of sources. Some domains show up every single day. Others appear once and never again.

Citation volatility puts a number on that turnover. Each day, take the set of domains cited for a prompt and compare it to the previous day's set. If they are identical, the distance is 0. If they share nothing, it is 1. Average that across every pair of consecutive days and you have the prompt's volatility score. (Mathematically it is the Jaccard distance, 1 minus the shared domains over the combined domains.)

Run this across our 28 prompts and the scores landed mostly between 0.55 and 0.83 on Google AI Mode. A score of 0.69, the Google average, means barely a third of the domains in play on any two consecutive days were common to both. The sources behind a given AI answer are in near-constant motion.

The single biggest predictor of how much a prompt churns is the question itself. Narrow questions with an obvious frame of reference stay relatively stable. Broad, open-ended ones churn hard.

The mechanism is pool size. A question like "Jordan 1 vs Adidas Forum" has an obvious, finite set of relevant pages, so Google AI Mode drew on about 85 distinct domains across the whole six weeks and kept returning to the same ones (volatility 0.56). "Alternatives to Nike for high-performance running" sat on top of 445 distinct domains, a deep, interchangeable pool of best-of roundups and retailer pages, and a different slice surfaced each day (volatility 0.80). The broader the question, the larger the candidate pool, and the more the cited subset rotates.

This is the finding that travels. Absolute scores shift by category, but the ordering holds regardless of the brand: narrow, branded questions stay stable, while broad, open-ended ones churn. If you are tracking prompt-level visibility, expect your broadest, highest-intent prompts to be the noisiest.

Here is the part that a raw set-distance number hides. A prompt can swap out most of its cited domains every day and still cite the same two or three leaders every single time. The churn is real, but it is concentrated in the long tail.

To separate the two, weight each domain by how often it is actually cited, then measure volatility on those weighted sets. The membership-based score and the volume-weighted score tell different stories, and the gap between them is the size of the stable core.

This chart holds the most useful finding in the whole study, and it is a little counterintuitive. Google AI Mode has the higher raw set volatility (0.69 vs ChatGPT's 0.62) because it casts a wider net, citing about 24 domains per day to ChatGPT's 17. But its volume-weighted volatility is lower (0.40 vs 0.52). The reason is concentration: on Google AI Mode, the domains cited on at least 80% of days accounted for about 56% of all citations. On ChatGPT, that stable core captured only about 23%.

So the two engines work differently. Google AI Mode keeps a fixed cast of headline sources and rotates the supporting extras. ChatGPT spreads its citations more evenly across a smaller cast and rotates even the leads. For anyone doing answer engine optimization, that distinction sets the strategy: breaking into Google AI Mode's core is hard but durable, while ChatGPT's lead roles turn over often, so there is more room to break in and more risk of dropping out.

The practical takeaway is to never trust the set-membership number alone. A prompt that looks wildly volatile may have a rock-solid core with a noisy tail. Weight by citation volume before you read anything into the score.

The two engines diverge in a more basic way too: whether they search the web at all. Google AI Mode is a search product, so it grounds nearly every answer. ChatGPT decides per question, and the decision tracks intent.

For commercial questions like "best basketball shoes 2026", ChatGPT searched and cited on nearly every day, just like Google. But for "the most iconic sneakers of all time", a question of opinion and history, it answered from its training data and cited sources on only 6 of 42 days. Head-to-head comparisons sat in between: it grounded some days, recalled others.

That changes what you optimize for. Where an engine does not cite, there is no citation slot to win. The lever becomes whether the model already knows your brand from its training data, which you influence through broad brand mentions rather than a single citable page. Where the engine does cite, fresh, citable content is in play. The first step in any AEO plan is finding out which of your prompts are even grounded, on each engine, before you spend effort on them.

One caution this surfaces: a score built on six citing days is noise, not signal. The 6-day prompts can post extreme volatility numbers purely because there is so little data. Gate every score on a minimum number of observations, around 20 day-to-day comparisons, before you trust it.

Two surfaces, two different games. The whole contrast in one view:

Volatility describes how the cited sources move. A related question is who gets mentioned in the answer, and the data here is humbling even for a giant. Across these 28 prompts, AI almost never answered with Nike alone. It named a long roster of competitors, constantly.

Under Armour turned up in 81% of answers, Adidas in 59%, New Balance in 57%, with ASICS, Hoka, Brooks, and Saucony close behind. Even for the most recognizable name in athletic footwear, AI treats the category as a field, not a default. For a smaller brand the lesson is encouraging: the answer is a roster, and rosters have room. Tracking who shares your answers, your AI share of voice, is as important as tracking your own mentions. The per-engine split was nearly identical here, with Adidas in 57% of ChatGPT answers and 60% of Google's, so when rivals are all comparably sized brands, neither engine favors anyone in particular.

Pulling the threads together:

You do not need our dataset to run this. The method is simple and repeatable:

This is exactly what Elmo does automatically. It is an open-source, self-hosted AI visibility platform that runs your prompts across ChatGPT, Claude, Gemini, Perplexity, Google AI Mode, and more, records every cited domain, and computes the metrics in this post so you can watch how your sources, mentions, and share of voice move over time. Because it is MIT-licensed and you run it yourself, the underlying citation data is yours to query however you like, which is exactly how this study was made.

To go deeper on the fundamentals, start with our guide to answer engine optimization, then see how to track your brand in AI search.