AI Citation Monitoring: 91% of Citations Change in a Month

What AI citation monitoring means

AI citation monitoring is the practice of repeatedly checking whether AI answer engines such as ChatGPT, Perplexity, and Google's AI Overviews name your agency, or link your page, when a buyer asks a question in your market, and recording that result on a fixed schedule instead of checking once and assuming it holds. It is not the same task as rank tracking. A rank tracker tells you where a page sits on a list of ten blue links. Citation monitoring tells you whether a machine trusted that page enough to quote it, name your agency, and hand a reader your answer instead of a competitor's.

The distinction matters because the two signals are starting to move independently of each other. A page can hold position one in Google's organic results and still lose the AI Overview citation for the exact same query, because the citation is chosen by a separate retrieval and synthesis process that increasingly ignores the organic ranking underneath it. Ahrefs measured this directly: across 863,000 keywords, only 38% of pages cited inside a Google AI Overview also ranked in the top 10 organic results for that query, down from 76% just seven months earlier¹. Two years ago, "rank well and you get cited" was a reasonable assumption. It is no longer one you can build a strategy on.

Most agencies treat AEO as a project: rebuild the site, add schema, publish a batch of answer first content, and consider the job finished. Citation monitoring is the part that comes after the build, and it is the part almost nobody does, because Google Search Console does not report AI Overview impressions, citations, or which page got credit. That reporting gap is not an oversight. It reflects that AI Overviews render differently for logged in users and change output between sessions, so there is no single, stable "position" to report the way there is for organic search. If you are not checking manually or with a dedicated tool, you have no visibility into this surface at all, and you will not know you lost a citation until a competitor mentions it or your lead volume quietly drops.

It also helps to be precise about what monitoring is not. It is not the same thing as traditional review monitoring, which watches for what people say about you on Google or Yelp. It is not brand sentiment tracking, which watches tone. And it is not a one time citation check run before a website launch to prove the new build "works," which is the most common way agencies encounter this idea and then stop. A single successful check tells you the page was citable on the day you looked. It says nothing about whether it will still be citable next week, and the data in the next section is the reason that gap matters.

How volatile AI citations actually are

The honest answer to "how often do AI citations change" is: far more often than most agency owners assume, and the instability shows up at every time scale that has been studied, from the same day to the same month.

Start with the shortest window. SE Ranking ran the identical 10,000 keywords through Google's AI Mode three separate times on the same day in June 2025 and compared the cited URLs across the three runs. Across 9,451 keywords that triggered an AI Mode answer in all three attempts, the average URL overlap was just 9.2%, meaning roughly 91% of cited URLs differed between runs of the identical query on the identical day. At the domain level, where the standard is more forgiving because it only asks whether the same website showed up rather than the same page, overlap was still just 14.7%, and 21.2% of keywords returned zero matching URLs across all three runs². Nothing about the buyer's question changed between those three searches. The answer engine simply chose differently each time.

Extend the window to a month and the pattern holds, just measured a different way. Semrush tracked 3,000 keywords, split evenly between US desktop and mobile, for 31 days in October 2024, focusing on the keywords where an AI Overview appeared on at least 20 of those 31 days so the comparison was measuring real churn rather than the Overview simply disappearing. 91% of the URLs that ever appeared in those Overviews were removed at some point during the month. Only 43% of removed URLs on desktop, and 41% on mobile, ever came back. Google swapped in a URL from a different domain entirely in 96% of the changes it made, and the average URL stayed cited for just 3.87 consecutive days on desktop and 3.33 days on mobile before being replaced³. Zero of the 3,000 keywords in that study held 100% URL consistency for the full month.

One more piece explains why the top 10 overlap number moved so sharply in seven months. Ahrefs attributes part of the drop to Google's query fan out process, where a single search is split into several related sub-queries behind the scenes, and the page cited in the final answer is whichever one showed up most consistently across that whole set of sub-queries, not necessarily the page ranking highest for the original phrase typed into the box. A page can be the best literal match for what a buyer typed and still lose the citation to a page that covers the topic more completely across the adjacent questions Google silently asked on the buyer's behalf¹. That single mechanic is the reason narrow, keyword matched pages are losing ground to pages that answer a full topic.

Why insurance agencies are especially exposed

The volatility above applies to every industry, but insurance carries an additional problem on top of it: agencies are not just losing citations to competitors, they are largely absent from the citation pool in the first place. A 2026 study by Wellows ran 1,960 real insurance related prompts through AI answer engines and logged all 34,238 resulting citations across 3,831 distinct domains. Insurer and agency owned domains captured just 4.6% of those citations combined⁴.

Community and user generated content, meaning forum threads, Reddit posts, and review sites, accounted for 6.2% of citations in that same study, outperforming every insurer and agency website combined. Reddit alone was the single most cited domain in the entire dataset at 3.5%, ahead of any individual carrier or agency site, and Sleepfoundation.org, a general health content site with no insurance offering at all, was the second most cited domain overall at 2.9%⁴. When an AI system answers a coverage question, it is currently more likely to quote a stranger's Reddit comment or an unrelated health publisher than any business actually licensed to sell the policy.

The same study broke down what buyers were actually asking. Informational questions, the kind that explain a plan type or a rule rather than push toward a purchase, made up 46% of the prompt set. Commercial questions, closer to comparison shopping, were 27%, and explicit coverage questions, meaning a buyer describing their own situation and asking what applies to them, were roughly 9%⁴. Nearly half of the volume where an agency could be cited is informational content explaining how something works, which is exactly the category most agency sites either skip entirely or bury behind a lead form instead of publishing as a plain, citable answer.

What each platform weighs differently

A monitoring loop that only checks one platform is monitoring a fraction of the problem, because ChatGPT, Perplexity, Google AI Overviews, and Claude do not pull from the same pool of sources or cite at the same rate. A Qwairy analysis of 118,000 AI responses collected between January and March 2026 found Perplexity includes an average of 21.87 citations per response, more than double every other major platform, while Google AI Mode averages 8.34, ChatGPT averages 7.92, and Claude averages just 5.67⁷. A page that never shows up in a ChatGPT answer might still be cited constantly on Perplexity, simply because Perplexity surfaces roughly three times as many sources per answer and therefore has three times the opportunities to include you.

The type of source each platform prefers also differs enough to change what "getting cited" should even mean for your content plan. News sites and established publishers dominate across every platform studied, running from 38% of citations on ChatGPT up to 51% on Claude. Beyond that baseline, Perplexity and Google AI lean hardest on established topical authority sites, at 35% and 28% of citations respectively, while ChatGPT and Claude draw a larger share from academic and research sources⁷. None of the four platforms behave like a single "AI search" you can optimize for once. They are four separate retrieval systems that happen to produce a similar looking answer box.

The same analysis found that only 11% of cited domains appeared across more than one of these platforms⁷. That single number is the clearest argument against treating AEO as a task you finish. If nine out of ten sources one platform trusts are ignored by the next one, a monitoring loop that checks only Google, or only ChatGPT, is blind to the majority of the surface a buyer might actually be using. This is also why the weekly loop described below checks all three major consumer facing platforms every time, not one, and logs them separately rather than as a single combined score.

What actually causes a citation to change

A citation usually does not disappear because your page got worse. It disappears because one of a small number of mechanics fired, and knowing which one happened decides what, if anything, you should do about it.

Three of those six causes have nothing to do with anything you control, which is precisely why a one time optimization project cannot be the whole strategy. You cannot fix "the model samples differently on different runs." You can only build a loop that notices when the citation you are relying on is gone, figures out which of the six causes it was, and responds in proportion. A change caused by sampling noise deserves a shrug and a re-check next week. A change caused by a competitor publishing a genuinely more complete answer deserves a rewrite. Those look identical from the outside on the day you notice you dropped, and the only way to tell them apart is to look at what replaced you.

The weekly monitoring loop

A monitoring loop does not need to be complicated to work. It needs to run on a fixed schedule that is shorter than the volatility it is trying to catch. Given that Semrush measured an average citation lifespan of under 4 days once an Overview starts churning, and SE Ranking found meaningful movement inside the same day, weekly is the longest interval that still catches a real trend before it has already run its course and reversed itself again.

Put a number on it so the loop feels concrete rather than abstract. A Medicare and ACA agency running 10 questions across 3 platforms is checking 30 individual answers a week, which is roughly 5 to 6 minutes per question if you are reading each answer rather than skimming it, or about 50 to 60 minutes total once a week. That is the entire cost of catching a churn cycle that, per the Semrush data above, can fully replace a citation in under 4 days if nobody is watching. Here is what four consecutive weeks of a filled in log can look like for a single question, to make the pattern concrete rather than theoretical.

Manual tracking vs a paid tool

Both approaches work. The right one depends on how many questions and how many markets you are tracking, and neither is complicated to start.

Whichever route you start with, a free Chrome extension called Google AI Overview Impact Analysis is worth installing regardless, since it flags on the search results page itself whether an Overview is present for a given query, which is useful context even for an otherwise fully manual process⁶. And Ahrefs' own guidance on monitoring ChatGPT specifically recommends starting with a one time manual audit before moving to automated tracking, and notes that a real correlation between a content change and a citation change typically takes 4 to 8 weeks to show up⁵. That lag is worth internalizing before you conclude a fix did not work. Four weeks of no movement is not failure. It might just be week three of an eight week lag.

What to do the day you lose a citation

The instinct when a weekly log shows a lost citation is to immediately rewrite the page that lost it. That is usually the wrong first move. The right first move is diagnosis, because the fix is different depending on which of the six causes from earlier actually happened, and rewriting a page for the wrong reason wastes the one resource that matters here, which is time before the next check.

If a competitor's page took the citation, open it next to yours and look specifically at whether it answers more of the adjacent sub-questions a buyer would ask next, since that is the exact mechanic Ahrefs' query fan out finding describes. If community content took it, such as a Reddit thread or a review site, that is a signal the answer engine currently trusts independent, first person accounts on that specific question more than any vendor's page, and the response is not always a rewrite: sometimes it is showing up inside that community directly, since Reddit's own citation share is large enough now that a genuine, non-promotional presence there is a legitimate part of an AI visibility strategy, not a side project. If your own page changed right before the drop, for example a URL move, a redesign, or a content edit, check that change first before assuming anything external happened at all.

Working the Week 3 to Week 4 scenario from the table above end to end shows what this looks like in practice. The competitor page that took the citation in Week 3 held it again in Week 4, which crosses the two consecutive week action threshold. The next step is not to rewrite the agency's subsidy page from a blank page. It is to open the competitor page and the agency's page side by side and check three things in order: does the competitor page answer more of the adjacent questions a buyer would ask next, such as what income counts toward the subsidy calculation or what happens if income changes mid year, does it carry a more recent visible update date, and does it cite a primary source such as HealthCare.gov or CMS.gov where the agency's page only asserts a number without one. Whichever of those three gaps is real is the one to close first, and the question goes back into the following week's log to confirm the fix actually moved the citation rather than assuming it did.

Where monitoring fits in a real content cadence

Monitoring only pays off if there is something to do with what it finds, which is why it works best paired with an actual publishing cadence rather than sitting on top of a site that gets touched twice a year. A citation lost to a fresher competitor page needs a fresher page in response, on a timeline measured in weeks, not the next redesign cycle.

This is the same reason our own Digital Foundation service builds new blog and location content every week as a standing part of the system rather than a one off project: a site that only gets new content around a redesign cannot respond to the kind of weekly churn documented above, and a monitoring loop with no cadence behind it just tells you, accurately, that you are losing ground you have no near term plan to win back. The data behind every page we build and every number in this guide runs through the same source, the Brain, so what gets published is never a stale copy of last year's numbers, which matters directly for citation odds since AI systems weight freshness and sourcing as part of what makes a page worth quoting in the first place.

Questions agencies ask