How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 Citations

Key takeaways:

The article follows that order: how citations behave, how to tell a real loss from a blip, and what contributes to citations that last.

My colleague Xibeijia Guan asked five AI tools the same questions every day for 20 days. She took every website cited on day one and followed it. Was it still cited the next day, and the day after? If it disappeared, did it ever come back?

Xibeijia’s data covers one month. To see whether her patterns hold beyond that, we ran the same measurements on questions we track every day in Brand Radar: 25 about AI visibility tools and marketing conferences, and 69 about SEO and marketing tools. Where our results back up or contradict hers, you’ll find it in a “How solid is this?” note.

Let’s start with the most basic question: if an AI tool cites your page today, will it still cite it tomorrow? And if the citation disappears, does it come back?

Across all five tools, at least half of the citations disappeared by the next day: around 60% on Google AI Mode and ChatGPT, 50% on AI Overviews, and 80% on Gemini and Copilot.

Gemini and Copilot looked less consistent, partly because their next answer often cited nobody. Exclude those answers, and all five tools looked similar: 50–60% of citations disappeared by the next answer that included citations.

Changing fan-out searches explains only part of the citation churn. Before answering, AI tools often run several web searches, known as fan-out searches. ChatGPT repeated exactly the same searches as the previous day, only 26% of the time.

But even when the searches stayed the same, its cited websites changed 81% of the time, compared with 91% when the searches changed. This suggests much of the churn happens after the search, when the AI decides which pages to cite.

Glen saw the same gap at the page level: AI Overviews cited about half of the same pages again the next day, and ChatGPT about a quarter. In Report A and in Report B’s August windows, about half the citations on Google’s tools and ChatGPT were gone by day two, too.

Takeaway: Don’t judge AI visibility from one check. For each question you track, count how many days out of the last seven you were cited.

A missing citation isn’t necessarily gone for good. Across all five AI tools, roughly 75% of citations that disappeared at least once returned within the 20-day study.

About half returned within three days. On Google AI Overviews and ChatGPT, around 40% came back after just one day without being cited.

Of all websites cited on day one, only 20–27% disappeared and never returned during the study, depending on the tool.

The Brand Radar reports we checked against showed a similar pattern, although slightly fewer citations returned (around two in three). Our figures may underestimate the return rate, since citations disappearing near the end of the study had less time to come back. We also tracked only the first disappearance, so we couldn’t measure how often returning citations disappeared again.

Takeaway: Don’t change a page because of one missing citation. Most come back within days on their own. If yours is still missing after two or three weeks, then look at the page.

The longer a citation has lasted, the more likely it is to return after disappearing. Around a third of citations that lasted just one day never came back. But almost all citations that lasted 8–14 days eventually returned.

It’s a bit like the Lindy effect: the longer a citation has survived, the better its chances of surviving, or coming back.

The next-day return rates tell the same story. Only 14–25% of one-day citations returned the following day, compared with 69–87% of citations that had already lasted at least eight days.

Two Brand Radar reports showed that new citations were particularly unstable. Around 70% disappeared the next day, and only 34–45% of those returned during the study.

Takeaway: Don’t count a new citation as a lasting win until it’s appeared for several days. And if a citation that’s been there for a week disappears, give it time. There’s a good chance it’ll return on its own.

So far, we’ve counted every missing citation as a loss. But some of those losses aren’t what they seem. Sometimes the AI didn’t cite any website at all. Other times, it dropped your link but still mentioned your brand.

More than half of Gemini’s and Copilot’s answers included no citations, making their citations appear less stable than those of other AI tools.

For Google’s tools and ChatGPT, excluding empty answers barely changed the results because they almost always cited sources in our dataset.

To account for this, we recalculated the results, excluding answers without citations. Gemini and Copilot then looked much like AI Mode and ChatGPT, with around 60% of citations disappearing by the next answer that cited any sources.

On Google’s tools and ChatGPT, skipping changed these figures by about a point or less, because their answers almost always cite something. There, a lost citation almost always means another site took the spot.

Report B found the same pattern. On Gemini and Copilot, 68% of lost citations happened when the next answer cited nobody, compared with 1% on Google’s tools and usually 11% on ChatGPT. That ChatGPT figure rose to 58% between August 19 and September 7.

Many apparent comebacks were simply the tools citing sources again: 44% on Gemini and 52% on Copilot.

One limitation: excluding empty answers means comparing the next answer with citations, which could be several days later. We also couldn’t make a reliable 20-day comparison this way, since only 4% of Gemini’s questions and 27% of Copilot’s had citations on all 20 days.

Takeaway: On Gemini and Copilot, check whether the AI cited anyone before treating a missing citation as a loss. Track both how often your website gets cited and how often another source appears instead.

AI tools can link to your website, mention your brand by name, or do both. But most cited websites weren’t actually mentioned by name in the answer.

For example, an AI answer about AirTags might link to ZDNet without mentioning “ZDNet” in the text.

Even when a brand was both cited and mentioned, it often lost one without losing the other the next day.

On Google’s tools and ChatGPT, 12–16% lost their citation but kept the brand mention, while 9–12% kept the citation but lost the mention. On Gemini and Copilot, losing the citation while keeping the brand mention was much more common, partly because these tools often cited nobody.

Brand names also stick around longer than citations. On every tool, about 50-70% cited websites showed up on just one day, compared with about one in five brand names.

Brand names vary from day to day: an AI tool might say “HubSpot” one day and “HubSpot CRM” the next. To avoid counting these as losses, we used AI to group variants of the same brand, such as “Adobe Acrobat Online” and “Adobe Acrobat”, before measuring.

Glen’s manual check of 55 brands shows why this matters: 30–36% of apparent losses there were the same brand under a different name. Some variants may still slip through, so treat the brand mention figures as slightly conservative.

Takeaway: Track citations and brand mentions separately. Before counting a mention as lost, check whether the AI simply used a different version of your brand name.

If citations come and go this often, which ones stick? We compared pages that kept their citations with pages that lost them, looking at SEO metrics, how well the page title matched the question, and the type of question. These are correlations: they show what long-lasting citations have in common, not what makes them last.

Pages that rank in Google and already get search traffic kept their AI citations longer. On AI Overviews, pages cited for 15 days or more had a median of 88 estimated monthly visits from organic search, compared with 8 for pages cited for only one day. Domain Rating and links to the page made little difference: long- and short-lived citations came from similarly strong websites (typical DR 64–78).

This builds on our earlier study of 1.9 million AI Overview citations, which found that 86% of the top three cited pages ranked in Google’s top 100, with a median position of 3. That study showed that ranking pages are more likely to get cited. This one suggests they’re also more likely to stay cited.

On AI Overviews and AI Mode, pages that ranked anywhere in Google for the same question were about seven times more likely to keep their citations for at least a week. The difference was smaller on ChatGPT.

Ranking higher helped, but simply appearing in Google’s results mattered more. On AI Mode, around 40% of top-three pages stayed cited for at least a week, compared with 20% of pages ranking sixth or lower.

Pages with more existing organic traffic also tended to keep their citations longer.

Median organic traffic of cited pages on day one, grouped by how long they stayed cited.

SEO Authority metrics showed a much weaker connection:

Pages with no traffic, no backlinks, or very low authority were more common among short-lived citations. But beyond that, higher authority didn’t seem to improve citation stability much.

That doesn’t make backlinks useless. They can still help pages rank, and ranking is associated with longer-lasting citations.

Ranking isn’t required to be cited, though. Around 75% of pages cited by AI Mode and ChatGPT didn’t rank in Google for the same question. The same was true for nearly half of AI Overview citations.

The ranking analysis covered only 230 questions, so these findings show a strong connection, not a definitive rule. Smaller differences between ranking positions may be random variation.

Also, rankings and traffic aren’t independent signals. Organic traffic largely comes from search rankings, so both findings may reflect the same underlying relationship.

Takeaway: For Google AI Overviews and AI Mode, focus on getting your pages to rank for the questions you want citations from. Ranking at all appears more important than the exact position.

Use organic traffic rather than DR alone to identify pages likely to keep their citations, and prioritize link building that supports search rankings rather than chasing backlinks solely for AI visibility.

Pages with titles and descriptions that closely matched the question tended to keep their AI citations a little longer.

We scored each page’s match with the question from 0 to 1, where higher means a closer match.

On Google AI Overviews and AI Mode, the typical match score increased from 0.54 for one-day citations to 0.65–0.68 for citations lasting 15 days or more.

This wasn’t just true for the average page. Even the weakest-matching 25% of long-lasting citations matched the question more closely than the typical one-day citation.

On ChatGPT, the connection was much weaker. Pages cited for one day and those cited for two to seven days had almost identical scores (0.54 and 0.56). A clearer pattern appeared only among the roughly 5% of pages that stayed cited for more than a week.

Takeaway: Write clear page titles and meta descriptions that reflect the questions you want to be cited for on Google’s AI tools. Treat this as a small advantage alongside strong content and search visibility, not a replacement for them.

When someone asks about a specific brand, such as “Is Asana good for small teams?”, that brand’s website tends to keep its citations much longer than other sources.

On Google’s AI tools and ChatGPT, the named brand’s website stayed cited for all 20 days in roughly 30% of cases, four to seven times more often than the average cited website.

The brand’s own website also appeared more frequently throughout the study. On Google’s tools and ChatGPT, it was cited on 60–68% of days, compared with just 16–20% for other sources in the same answers. On Gemini and Copilot, other sources appeared more often, on 36–38% of days.

From one day to the next, the brand’s website kept its citation about 75% of the time (60% on Gemini). Other sources were much less consistent, with only around 33–40% appearing again, or about 50% on Copilot.

75% citation stability is lower than you might expect for a question about your own brand. The reason is the type of question.

Report B tracked 22 questions about Ahrefs and found that the brand’s own website was cited much more often for factual questions than for reviews or comparisons.

Share of days an Ahrefs page was cited, for the typical question in each group, 28 August to 4 October 2026. Source: Ahrefs via Letaido.

So, for factual questions about pricing, plans, and features, AI tools almost always cited Ahrefs’ own website. But for reviews and comparisons, Google’s tools preferred third-party sources. ChatGPT continued to cite Ahrefs frequently.

Glen found a similar pattern across his broader dataset. Brands and domains cited most frequently changed by only 11–14% between days, compared with 61–78% for other sources. However, the findings about facts, reviews, and comparisons come from just 22 questions about one brand, so they may not apply equally to every company.

Takeaway: Keep your pricing, plans, and feature pages clear, accurate, and easy to find. For reviews and comparisons, pay attention to the third-party websites Google’s AI tools already cite, since those may be more important for your visibility than your own website.

Beyond questions that mention specific brands, three types of websites stood out among citations that lasted all 20 days:

Takeaway: If a YouTube video, Reddit discussion, or major publisher consistently holds a citation spot you want, getting your brand featured in that content may be more realistic than trying to replace it.

How often AI citations change may depend on the type of question and the topic.

On ChatGPT and Google AI Mode, around 64–76% of cited pages changed from one day to the next, regardless of question type. Questions mentioning specific brands were the most stable, while “best X” questions were among the least stable.

Glen found even bigger differences when comparing industries. Health-related questions kept the same first cited source 56–79% of the time, compared with just 13–23% for fashion questions.

One possible explanation is competition. A question like “best project management software” may have dozens of similar pages competing for a few citation spots, while health questions may draw from a smaller pool of trusted sources.

Takeaway: Set different expectations for different questions. Brand-related queries tend to have more stable citations, while “best X” questions are likely to change frequently, even when your content performs well.

In Ahrefs Brand Radar, you can track your citation performance in detail using the Cited pages report.

Use the filters to switch between AI platforms or see data from all of them at once. The tabs below let you explore citations from your own site, your competitors, other websites, and pages you’ve added to your watchlist.

The historical graph shows how your citations have changed over time, so you can see whether they’re staying consistent or fluctuating.

Below the graph, you’ll find several useful metrics to help you understand how valuable each citation is and how likely it is to stick around:

One of my favorite ways to use this report is the watchlist. Think of it as rank tracking, but for AI citations.

Whenever I’m optimizing a page specifically for AEO or want to monitor how a competitor’s page performs across my tracked prompts, I add it to the watchlist. It’s an easy way to keep tabs on the pages that matter most.

And when you combine it with tag filters, you can focus on the specific prompt groups that need the most attention.

You can take this analysis even further with AI.

Or, if you want a quicker answer, just ask Agent A directly in Brand Radar.

For example, you could ask: “What’s the average citation rate of the pages on my watchlist based in the entire tracking period?”

That way, you can get useful insights from your citation data without having to dig through the numbers yourself.

Citations might not be the best KPI because they naturally fluctuate a lot. But they can be more important for branded queries, since AI systems, especially ChatGPT, are likely to look for information about your brand, including competitor comparisons published on your own site.

Here’s a quick trick to find citation gaps for branded queries.

Go to the AI responses report, set the Prompt contains filter to your brand name, and set the Your brand filter to Not found.

You’ll get a list of branded prompts where your website wasn’t even retrieved by the AI system. These are potential opportunities to improve your visibility.