How Secure Are AI Citations? 10 Issues We Discovered Monitoring 37,500 Citations

The article follows that order: how citations behave, tips on how to inform an actual loss from a blip, and what contributes to citations that final.

My colleague Xibeijia Guan requested 5 AI instruments the identical questions each day for 20 days. She took each web site cited on day one and adopted it. Was it nonetheless cited the following day, and the day after? If it disappeared, did it ever come again?

Xibeijia’s knowledge covers one month. To see whether or not her patterns maintain past that, we ran the identical measurements on questions we observe each day in Model Radar: 25 about AI visibility instruments and advertising conferences, and 69 about search engine optimization and advertising instruments. The place our outcomes again up or contradict hers, you will discover it in a “How stable is that this?” notice.

Let’s begin with essentially the most primary query: if an AI device cites your web page at present, will it nonetheless cite it tomorrow? And if the quotation disappears, does it come again?

Throughout all 5 instruments, a minimum of half of the citations disappeared by the following day: round 60% on Google AI Mode and ChatGPT, 50% on AI Overviews, and 80% on Gemini and Copilot.

Gemini and Copilot regarded much less constant, partly as a result of their subsequent reply usually cited no person. Exclude these solutions, and all 5 instruments regarded related: 50–60% of citations disappeared by the following reply that included citations.

Altering fan-out searches explains solely a part of the quotation churn. Earlier than answering, AI instruments usually run a number of internet searches, often called fan-out searches. ChatGPT repeated precisely the identical searches as yesterday, solely 26% of the time.

However even when the searches stayed the identical, its cited web sites modified 81% of the time, in contrast with 91% when the searches modified. This means a lot of the churn occurs after the search, when the AI decides which pages to quote.

Glen noticed the identical hole on the web page degree: AI Overviews cited about half of the identical pages once more the following day, and ChatGPT a couple of quarter. In Report A and in Report B’s August home windows, about half the citations on Google’s instruments and ChatGPT have been passed by day two, too.

Takeaway: Don’t decide AI visibility from one examine. For every query you observe, depend what number of days out of the final seven you have been cited.

A lacking quotation is not essentially gone for good. Throughout all 5 AI instruments, roughly 75% of citations that disappeared a minimum of as soon as returned throughout the 20-day research.

About half returned inside three days. On Google AI Overviews and ChatGPT, round 40% got here again after simply at some point with out being cited.

Of all web sites cited on day one, solely 20–27% disappeared and by no means returned through the research, relying on the device.

The Model Radar reviews we checked in opposition to confirmed the same sample, though barely fewer citations returned (round two in three). Our figures could underestimate the return fee, since citations disappearing close to the top of the research had much less time to return again. We additionally tracked solely the primary disappearance, so we could not measure how usually returning citations disappeared once more.

Takeaway: Do not change a web page due to one lacking quotation. Most come again inside days on their very own. If yours continues to be lacking after two or three weeks, then have a look at the web page.

The longer a quotation has lasted, the extra seemingly it’s to return after disappearing. Round a 3rd of citations that lasted simply at some point by no means got here again. However virtually all citations that lasted 8–14 days finally returned.

It’s kind of just like the Lindy impact: the longer a quotation has survived, the higher its probabilities of surviving, or coming again.

The following-day return charges inform the identical story. Solely 14–25% of one-day citations returned the next day, in contrast with 69–87% of citations that had already lasted a minimum of eight days.

Two Model Radar reviews confirmed that new citations have been significantly unstable. Round 70% disappeared the following day, and solely 34–45% of these returned through the research.

Takeaway: Do not depend a brand new quotation as an enduring win till it is appeared for a number of days. And if a quotation that is been there for per week disappears, give it time. There is a good likelihood it’s going to return by itself.

To date, we have counted each lacking quotation as a loss. However a few of these losses aren’t what they appear. Typically the AI did not cite any web site in any respect. Different occasions, it dropped your hyperlink however nonetheless talked about your model.

Greater than half of Gemini’s and Copilot’s solutions included no citations, making their citations seem much less secure than these of different AI instruments.

For Google’s instruments and ChatGPT, excluding empty solutions barely modified the outcomes as a result of they virtually all the time cited sources in our dataset.

To account for this, we recalculated the outcomes, excluding solutions with out citations. Gemini and Copilot then regarded very like AI Mode and ChatGPT, with round 60% of citations disappearing by the following reply that cited any sources.

On Google’s instruments and ChatGPT, skipping modified these figures by a couple of level or much less, as a result of their solutions virtually all the time cite one thing. There, a misplaced quotation virtually all the time means one other website took the spot.

Report B discovered the identical sample. On Gemini and Copilot, 68% of misplaced citations occurred when the following reply cited no person, in contrast with 1% on Google’s instruments and normally 11% on ChatGPT. That ChatGPT determine rose to 58% between August 19 and September 7.

Many obvious comebacks have been merely the instruments citing sources once more: 44% on Gemini and 52% on Copilot.

One limitation: excluding empty solutions means evaluating the following reply with citations, which could possibly be a number of days later. We additionally could not make a dependable 20-day comparability this fashion, since solely 4% of Gemini’s questions and 27% of Copilot’s had citations on all 20 days.

Takeaway: On Gemini and Copilot, examine whether or not the AI cited anybody earlier than treating a lacking quotation as a loss. Monitor each how usually your web site will get cited and the way usually one other supply seems as a substitute.

AI instruments can hyperlink to your web site, point out your model by identify, or do each. However most cited web sites weren’t really talked about by identify within the reply.

For instance, an AI reply about AirTags may hyperlink to ZDNet with out mentioning “ZDNet” within the textual content.

Even when a model was each cited and talked about, it usually misplaced one with out shedding the opposite the following day.

On Google’s instruments and ChatGPT, 12–16% misplaced their quotation however stored the model point out, whereas 9–12% stored the quotation however misplaced the point out. On Gemini and Copilot, shedding the quotation whereas retaining the model point out was far more widespread, partly as a result of these instruments usually cited no person.

Model names additionally stick round longer than citations. On each device, about 50-70% cited web sites confirmed up on simply at some point, in contrast with about one in 5 model names.

Model names fluctuate from each day: an AI device may say “HubSpot” at some point and “HubSpot CRM” the following. To keep away from counting these as losses, we used AI to group variants of the identical model, comparable to “Adobe Acrobat On-line” and “Adobe Acrobat”, earlier than measuring.

Glen’s guide examine of 55 manufacturers exhibits why this issues: 30–36% of obvious losses there have been the identical model underneath a unique identify. Some variants should still slip by, so deal with the model point out figures as barely conservative.

Takeaway: Monitor citations and model mentions individually. Earlier than counting a point out as misplaced, examine whether or not the AI merely used a unique model of your model identify.

If citations come and go this usually, which of them stick? We in contrast pages that stored their citations with pages that misplaced them, search engine optimization metrics, how properly the web page title matched the query, and the kind of query. These are correlations: they present what long-lasting citations have in widespread, not what makes them final.

Pages that rank in Google and already get search visitors stored their AI citations longer. On AI Overviews, pages cited for 15 days or extra had a median of 88 estimated month-to-month visits from natural search, in contrast with 8 for pages cited for less than at some point. Area Score and hyperlinks to the web page made little distinction: long- and short-lived citations got here from equally sturdy web sites (typical DR 64–78).

This builds on our earlier research of 1.9 million AI Overview citations, which discovered that 86% of the highest three cited pages ranked in Google’s prime 100, with a median place of three. That research confirmed that rating pages usually tend to get cited. This one suggests they’re additionally extra more likely to keep cited.

On AI Overviews and AI Mode, pages that ranked wherever in Google for a similar query have been about seven occasions extra more likely to hold their citations for a minimum of per week. The distinction was smaller on ChatGPT.

Rating larger helped, however merely showing in Google’s outcomes mattered extra. On AI Mode, round 40% of top-three pages stayed cited for a minimum of per week, in contrast with 20% of pages rating sixth or decrease.

Pages with extra present natural visitors additionally tended to maintain their citations longer.

Median natural visitors of cited pages on day one, grouped by how lengthy they stayed cited.

search engine optimization Authority metrics confirmed a a lot weaker connection:

Pages with no visitors, no backlinks, or very low authority have been extra widespread amongst short-lived citations. However past that, larger authority did not appear to enhance quotation stability a lot.

That does not make backlinks ineffective. They will nonetheless assist pages rank, and rating is related to longer-lasting citations.

Rating is not required to be cited, although. Round 75% of pages cited by AI Mode and ChatGPT did not rank in Google for a similar query. The identical was true for almost half of AI Overview citations.

The rating evaluation coated solely 230 questions, so these findings present a robust connection, not a definitive rule. Smaller variations between rating positions could also be random variation.

Additionally, rankings and visitors aren’t impartial indicators. Natural visitors largely comes from search rankings, so each findings could mirror the identical underlying relationship.

Takeaway: For Google AI Overviews and AI Mode, concentrate on getting your pages to rank for the questions you need citations from. Rating in any respect seems extra vital than the precise place.

Use natural visitors reasonably than DR alone to establish pages more likely to hold their citations, and prioritize hyperlink constructing that helps search rankings reasonably than chasing backlinks solely for AI visibility.

Pages with titles and descriptions that intently matched the query tended to maintain their AI citations just a little longer.

We scored every web page’s match with the query from 0 to 1, the place larger means a more in-depth match.

On Google AI Overviews and AI Mode, the everyday match rating elevated from 0.54 for one-day citations to 0.65–0.68 for citations lasting 15 days or extra.

This wasn’t simply true for the typical web page. Even the weakest-matching 25% of long-lasting citations matched the query extra intently than the everyday one-day quotation.

On ChatGPT, the connection was a lot weaker. Pages cited for at some point and people cited for 2 to seven days had virtually an identical scores (0.54 and 0.56). A clearer sample appeared solely among the many roughly 5% of pages that stayed cited for greater than per week.

Takeaway: Write clear web page titles and meta descriptions that mirror the questions you wish to be cited for on Google’s AI instruments. Deal with this as a small benefit alongside sturdy content material and search visibility, not a substitute for them.

When somebody asks a couple of particular model, comparable to “Is Asana good for small groups?”, that model’s web site tends to maintain its citations for much longer than different sources.

On Google’s AI instruments and ChatGPT, the named model’s web site stayed cited for all 20 days in roughly 30% of instances, 4 to seven occasions extra usually than the typical cited web site.

The model’s personal web site additionally appeared extra incessantly all through the research. On Google’s instruments and ChatGPT, it was cited on 60–68% of days, in contrast with simply 16–20% for different sources in the identical solutions. On Gemini and Copilot, different sources appeared extra usually, on 36–38% of days.

From at some point to the following, the model’s web site stored its quotation about 75% of the time (60% on Gemini). Different sources have been a lot much less constant, with solely round 33–40% showing once more, or about 50% on Copilot.

75% quotation stability is decrease than you may count on for a query about your individual model. The reason being the kind of query.

Report B tracked 22 questions on Ahrefs and located that the model’s personal web site was cited far more usually for factual questions than for opinions or comparisons.

Share of days an Ahrefs web page was cited, for the everyday query in every group, 28 August to 4 October 2026. Supply: Ahrefs through Letaido.

So, for factual questions on pricing, plans, and options, AI instruments virtually all the time cited Ahrefs’ personal web site. However for opinions and comparisons, Google’s instruments most popular third-party sources. ChatGPT continued to quote Ahrefs incessantly.

Glen discovered the same sample throughout his broader dataset. Manufacturers and domains cited most incessantly modified by solely 11–14% between days, in contrast with 61–78% for different sources. Nonetheless, the findings about info, opinions, and comparisons come from simply 22 questions on one model, so they could not apply equally to each firm.

Takeaway: Maintain your pricing, plans, and have pages clear, correct, and straightforward to seek out. For opinions and comparisons, take note of the third-party web sites Google’s AI instruments already cite, since these could also be extra vital in your visibility than your individual web site.

Past questions that point out particular manufacturers, three varieties of web sites stood out amongst citations that lasted all 20 days:

Takeaway: If a YouTube video, Reddit dialogue, or main writer constantly holds a quotation spot you need, getting your model featured in that content material could also be extra practical than making an attempt to interchange it.

How usually AI citations change could rely on the kind of query and the subject.

On ChatGPT and Google AI Mode, round 64–76% of cited pages modified from at some point to the following, no matter query sort. Questions mentioning particular manufacturers have been essentially the most secure, whereas “finest X” questions have been among the many least secure.

Glen discovered even larger variations when evaluating industries. Well being-related questions stored the identical first cited supply 56–79% of the time, in contrast with simply 13–23% for style questions.

One doable rationalization is competitors. A query like “finest undertaking administration software program” could have dozens of comparable pages competing for a couple of quotation spots, whereas well being questions could draw from a smaller pool of trusted sources.

Takeaway: Set completely different expectations for various questions. Model-related queries are inclined to have extra secure citations, whereas “finest X” questions are more likely to change incessantly, even when your content material performs properly.

In Ahrefs Model Radar, you possibly can observe your quotation efficiency intimately utilizing the Cited pages report.

Use the filters to change between AI platforms or see knowledge from all of them directly. The tabs under allow you to discover citations from your individual website, your rivals, different web sites, and pages you’ve got added to your watchlist.

The historic graph exhibits how your citations have modified over time, so you possibly can see whether or not they’re staying constant or fluctuating.

Under the graph, you will discover a number of helpful metrics that can assist you perceive how invaluable every quotation is and the way seemingly it’s to stay round:

Certainly one of my favourite methods to make use of this report is the watchlist. Consider it as rank monitoring, however for AI citations.

Every time I am optimizing a web page particularly for AEO or wish to monitor how a competitor’s web page performs throughout my tracked prompts, I add it to the watchlist. It is a simple strategy to hold tabs on the pages that matter most.

And whenever you mix it with tag filters, you possibly can concentrate on the precise immediate teams that want essentially the most consideration.

You possibly can take this evaluation even additional with AI.

Or, if you need a faster reply, simply ask Agent A instantly in Model Radar.

For instance, you can ask: “What is the common quotation fee of the pages on my watchlist based mostly in the complete monitoring interval?”

That manner, you will get helpful insights out of your quotation knowledge with out having to dig by the numbers your self.

Citations may not be the most effective KPI as a result of they naturally fluctuate loads. However they are often extra vital for branded queries, since AI programs, particularly ChatGPT, are more likely to search for details about your model, together with competitor comparisons revealed by yourself website.

This is a fast trick to seek out quotation gaps for branded queries.

Go to the AI responses report, set the Immediate comprises filter to your model identify, and set the Your model filter to Not discovered.

You will get an inventory of branded prompts the place your web site wasn’t even retrieved by the AI system. These are potential alternatives to enhance your visibility.