Last Updated on August 19, 2026 by Alex Birkett
If you’re planning organic growth strategy against AI search, most people focus on the question, “which pages are cited today?”
Or worse, which citations are cited most often in aggregate across industries.
This is nearly useless for your specific context. Of course Reddit, Wikipedia and YouTube are highly cited – they are gigantic websites with long histories and hyper specific threads that touch into nearly every topic. Does that help a B2B SaaS selling cybersecurity? Probably not.
An even better question is to consider the moving target, where the puck is going: “which pages will still be cited next month?”
We often assume stability in citation sources, but I got curious about citation drift.
What is Citation Drift?
Citation Drift is the time series measurement of volatility or stability of citation sources that artificial intelligence platforms like ChatGPT, Claude, Perplexity, and AI Mode use as references for AI answers.
It is, quite simple, a time-based comparison of the similarity in makeup of citation shares across distinct periods (weekly, monthly, quarterly, or yearly).
It is often split between URL drift and Domain drift, with most research showing high volatility in URLs but more stability in domains.
It has direct implications for your AEO / AI search strategy, especially as it relates to off page campaigns like brand mention acquisition, review site management, community engagement, and social media management.
Original Research: About Half the List Turns Over Monthly
In a consulting engagement through Stella Nova, I worked with an AI recruiting platform. We had made substantial progress in moving their share of voice and AI visibility (up from 19% to 29% with significant increases in owned content citation share and retrieval rate).
We focused on the blocking and tackling first, because they had apparent technical and content gaps.
After a few months, we layered on off page campaigns and outreach, which provoked the question: “which sources are most influential?”
A snapshot in time analysis revealed a pareto distribution, with the top 10 domains representing the bulk of retrievals.
This led me to ask, “how often does that top N source data change?” Or rather, what is the citation drift in our space?
I looked at the top 500 URLs cited in a large query set. This was roughly 100 prompts and tens of thousands of citations, yet I wanted to focus merely on the top 500 as a function of utility – most of what we do in growth should be focused at the highest impact levers. There is an abundant long tail, but that has been studied at length (see below).
I looked at two points one month apart. Only 245 of them — 49% — appeared in both snapshots. The other 255 dropped out entirely, replaced by 259 URLs that hadn’t been there before.
Treated as a set-comparison problem, the two lists share a Jaccard similarity of just 0.32, meaning about two-thirds of the combined universe of cited pages is unique to one snapshot or the other.
The picture changes when you zoom out to the domain level. Of the 204 unique publishers appearing in the earlier snapshot, 133 were still represented later — a domain-level stability of 65%, versus 49% at the URL level.
The sources AI systems reach for stayed largely the same. It was the specific pages within them that shifted.
Some domains moved dramatically. Reddit lost 14 cited pages between snapshots, dropping from 24 to 10. One vertical publisher fell from 49 to 43. Meanwhile, Pin.com went from 0 to 13, G2 from 3 to 11, and a cluster of niche publications appeared where none had been cited before.
This is exactly what you’d expect from systems that rotate through many candidates around a stable spine.
Editorial note: the research question here was less about the specific movements (e.g. Reddit versus G2 prominence) and more about volatility in a time period. I highly suspect the movements could reverse or reshuffle in unpredictable ways over a longer time period or repeat analysis. The research was designed to tell us where to focus on our attention if our goal was to build long term visibility.
The broader picture: this is what everyone is seeing
My result isn’t an outlier. It’s now the third or fourth independent measurement of the same phenomenon, from very different angles.
SISTRIX’s April 2026 AI citation drift study is the largest public benchmark to date: 82,619 prompts producing 1,548,213 snapshots across six countries, three platforms, and 17 weeks of weekly re-sampling.
The headline number is that citation churn runs 56% per week at the domain level in Google AI Mode, and 74% in ChatGPT Search. AI Overviews are more bimodal: for 53% of prompts nothing changes across the full 17 weeks, but for the remaining minority the sources rotate as fast as AI Mode.
Weekly drift stays consistently in the 54–59% band across every country SISTRIX tested.
Their most interesting finding for our purposes: for 86.5% of AI Mode prompts, there’s a “fixed core” of 1–5 reliably-cited domains, with the remainder rotating at 89% per week. That is essentially the pattern I saw (a stable domain spine, high page-level churn) measured on a totally different corpus.
SISTRIX also notes their numbers are conservative at the domain level; measured at the URL level, drift is about 15 percentage points higher, because even when a domain stays cited, the specific page on that domain frequently changes.
Profound’s July 2025 study measured domain level drift over a one-month window (June 11–13 to July 11–13, 2025) with roughly 80,000 prompts per platform.
Their numbers: Google AI Overviews 59.3%, ChatGPT 54.1%, Microsoft Copilot 53.4%, and Perplexity 40.5%.
Perplexity is the outlier because it anchors more heavily to a handful of high-authority sources.
Extended to a six-month window (January to July), drift climbs to 70–90% (roughly linear with time).
Notably, Profound’s monthly ChatGPT number (54%) is lower than SISTRIX’s weekly ChatGPT number (74%), which is what you’d expect if week-to-week fluctuations partially average out on a monthly view.
Stacked together, these numbers describe the same underlying reality. Week to week, AI systems rotate roughly half their domain citations on Google’s platforms, and closer to three-quarters on ChatGPT.
Month to month, the numbers hold at 40–60% depending on platform.
Over six months, 70–90% of the cited domain set turns over. The one thing that stays put is the “core,” a small set of high-authority sources that appear consistently across time and prompts.
My analysis of URL level turnover of 51% sits inside that range. My domain-level turnover of 35% is at the lower end of what everyone else is measuring, which makes sense given it is a snapshot of an individual client versus an aggregate pattern, but is consistent with the fact that my snapshots are spaced closer to one month than to six.
Why Citation Drift Happens
Citation drift isn’t a bug.
It’s the intended behaviour of probabilistic retrieval systems. Every AI response draws from a candidate pool sampled with some randomness, filtered by freshness signals, weighted by whatever the current ranking model believes is authoritative for the query at hand.
Small changes to any of those inputs (a model update, a re-index, a new page from a competitor, a fresh Reddit thread) cascade into different citation sets.
That has two important consequences:
- First, single snapshot rankings of “who gets cited” are inherently noisy. If you make strategy decisions from one week of data, you’re partly measuring signal and partly measuring the roll of the dice.
- Second, the assets that survive across snapshots are disproportionately valuable, because they’re the ones the system keeps re-selecting despite the randomness. Whatever qualities they share (domain authority, structural clarity, freshness, unique data) are the qualities to chase. Influence, as I say, is upstream.
What Drives Citation Stability
This is getting into conjecture, as I am not highly confident on the causal research here.
But some of the external studies point to some hypotheses, and I do think they are fairly uncontroversial and validated at least with correlation research.
One of the better supported levers comes from AirOps’s “Silent Pipeline Killer” report, which analysed 4,000+ ChatGPT cited pages across 900 high-intent queries in 15 industries.
Their key findings: 35% of ChatGPT-cited pages had been updated in the last three months, and 53% within six months.
On commercial queries specifically (the ones tied to buying decisions) 83% of citations came from pages updated within the past year, and 60% within the past six months.
Pages not updated for over a year are more than twice as likely to lose their citations, and pages that miss an annual refresh lose over half their chance of being cited at all. Freshness windows vary by industry: 3-9 months for SaaS, finance and news; up to 12+ months for slower verticals like education and research.
SISTRIX adds a second dimension: page type matters as much as page age.
Their classification of cited domains in Google AI Mode shows a clear hierarchy in “core rate,” the probability that a cited domain sticks in the stable core rather than the rotating carousel.
YouTube leads at 24%, big tech platforms at 16%, Wikipedia at 12%, marketplaces at 10%, forums/UGC at 3%, and news articles at just 1.4%.
News content is essentially a one way ticket into AI responses: cited this week, gone next week.
Evergreen product and reference pages systematically survive better than editorial content.
What This Means for AEO Strategy
Reading our data through this lens: the good news is that the domain footprint is much more durable than the URL footprint suggests.
Nearly two thirds of the publishers we care about are still in play across snapshots, and the “fixed core” behaviour SISTRIX describes is measurable and, in principle, targetable.
The bad news is that any specific page’s citation is closer to a lease than a purchase, though may have indirect effects on cumulative exposure (i.e. is it raises visibility in the short term, presumably this could lead to getting listed or mentioned in other unrelated citations).
But expect roughly half of your cited URLs to fall out on a multi month horizon, and plan content operations accordingly.
The practical implications: treat citation acquisition as a stock that decays, not a one time win.
Budget for continuous refresh, because AirOps’s data shows the freshness signal is real and measurable.
Prioritize evergreen product and reference formats over news style editorial when the goal is durable citation share.
Don’t over index on any single monthly snapshot; use aggregate presence over time as the metric, since one week signal to noise is unfavorable.
And measure at the level of analysis that matches your goal. Domain presence is measurable and controllable, but which specific URL surfaces on that domain in a given week largely is not.
The stability question is fairly clear to me. We are not operating in a linear, deterministic world anymore. We are engineering for probability, and much of the leverage comes upstream of any given citation source URL.