Are AI Citations Stable or Do They Churn?
AI citations are not stable: 74% of the domains ChatGPT cites for a prompt change from one week to the next, and 56% change in Google’s AI Mode. Most advice says to watch every citation closely, and I think that is a waste of your time. For best results, track your share of answers over a month and ignore single citations.
AI Citations Churn Every Week, and Even Run to Run
AI citations churn at 74% a week in ChatGPT search and 56% a week in Google’s AI Mode. Those figures come from a SISTRIX study that tracked 82,619 prompts for 17 weeks, and it found that the typical ChatGPT prompt had no domain that showed up in all 17.
| Platform | Weekly churn in cited domains | What stays put |
|---|---|---|
| ChatGPT search | 74% | Almost nothing for the typical prompt |
| Google AI Mode | 56% | A core of 1 to 5 domains on 86.5% of prompts |
| Google AI Overviews | Low | 53% of prompts kept the same sources for 17 weeks |
And a week is the slow clock. Ask the same question twice in a row and you get a different answer. SparkToro and Gumshoe ran 12 prompts 2,961 times through ChatGPT, Claude, and Google’s AI and found less than a 1 in 100 chance of getting the same list of brands twice.
Google’s own answers move the same way. Ahrefs watched 43,000 keywords and found an AI Overview has a 70% chance of changing from one check to the next, with 45.5% of the cited URLs swapped for new ones.
But the meaning of the answer barely moved in that study. The links rotate while the message holds, which is the gap I explain in AI citations vs mentions.
So the platform changes the job. A citation in an AI Overview behaves closer to a ranking you can hold for a while. A citation in ChatGPT behaves like press coverage, which is worth earning and impossible to keep on a lease.
Most Citation Tracking Measures Noise
Most citation tracking measures noise: even five runs of the same prompt leave a margin of error of 27 points. That number is from Evertune’s test of 10,700 prompts, for a brand that shows up about 10% of the time. At 100 runs the error fell to about 6 points.
Most SEOs will argue with me on this. The standard advice is to set up a dashboard, watch your citations daily, and react when one drops. But a daily check is one run, and one run tells you less than five do.
A store owner sees their brand cited on Monday and gone on Tuesday, and it feels like a ranking loss. It isn’t one. The citation was never theirs to hold, and a client who panics over a lost citation on Tuesday usually has it back by the following week without anyone touching the page.
Stop tracking AI citations one at a time, and start tracking how often you show up across a month of answers. SparkToro reached the same place from its data: a visibility percentage across dozens of prompts, each run many times, is a reasonable metric, and a single ranking is not. Most AI visibility tracking tools can report it that way if you set them up to.
The Platforms Move the Floor Under You
The platforms move the floor under you a few times a year, and no page edit prevents it. In March 2026, seoClarity found the share of ChatGPT answers in the US with no citations at all doubled, from 28% to 48%. Citation volume fell 86% to 94% between February and April, then came back in May.
Nobody who lost citations that spring did anything wrong. The change was on ChatGPT’s side, and every site’s numbers fell together. The year before, Reddit went from showing up in about 60% of ChatGPT’s answers to about 10% in six weeks, which I cover in does ChatGPT use Reddit.
Compare that with Google’s regular search results. Google ran three core updates in 2025 and has run two so far in 2026, and you get months of calm between them. AI answers have no calm stretch, so the habits you built for rank tracking don’t carry over.
The Stable Part Is Your Domain
The stable part is your domain, and SISTRIX found it on 86.5% of Google AI Mode prompts. Those prompts kept a core of 1 to 5 domains in place for weeks and months, while the other slots turned over at 89% a week. The answer has a few regulars and a rotating cast of guests.
Getting cited once makes you a guest. The regulars in that study were sites like YouTube, Amazon, and the product and shop pages that fit the question, and 85% of the core content was evergreen. For questions about a brand, the brand’s own domain held its spot all 17 weeks in 43% of cases.
So one great post is a foothold and not a moat. The model decides it trusts a site, then keeps picking whichever page on it looks most current for the question. That’s why I’d rather see a store publish ten solid pages in one category than chase one page’s citation, and why content freshness matters for AI citations.
Track Share of Answers by Month
Tracking share of answers by month, on a fixed set of 20 to 30 prompts, gives you a number that survives the churn. Make each prompt a question a shopper would ask before buying, and keep the set the same so one month can be compared with the next.
From there, log three things for every prompt:
- How often your brand is named in the answer
- How often your site is cited as a source
- Which competitors show up more often than you do
Then read the trend once a month and make decisions once a quarter. A single day’s screenshot should never go in a report.
Also run each prompt more than once. A tracking tool does the repeats for you, and Evertune saw the error stop shrinking much past 100 runs. If you check by hand, five runs per prompt is the floor, and you should treat the result as a rough read.
The other number I trust is revenue. Sessions and orders from ChatGPT and the other assistants show up in your analytics, and they don’t swing with every reshuffle the way a citation count does. That reporting is part of my AI search services.
Want a Number You Can Trust?
AI citations churn every week, the platforms shake them up a few times a year, and the only part that holds is the trust a model has in your whole site. So measure the trend and build the site.
Book a call and I’ll show you which prompts in your category name your store today and which ones name your competitors. If you want the tracking and the content handled every month, that is what my AI search services for Shopify stores are for.
