Half of AI Citations Vanish Overnight. What 37,500 Citations and 96,418 URLs Reveal About GEO

Key takeaways

• AI citations are unstable by design. In an Ahrefs study that tracked about 37,500 citations for 20 days, 50% to 83% of the websites cited on day one were gone the next day, depending on the platform. Only 4% to 7% stayed cited for all 20 days on Google's AI tools and ChatGPT.
• Most losses are temporary. About three quarters of lost citations came back within 20 days, and citations that had lasted a week or more almost always returned.
• A separate SISTRIX study of 96,418 URLs found that page structure is associated with being cited: a table of contents (1.7 times), Article schema (1.6), clear H2 structure (1.6) and FAQ schema (1.5). llms.txt showed no measurable advantage.
• The traffic at stake is still small. Across 100 websites, ChatGPT sent visits equal to 0.62% of Google organic sessions, though those visitors converted at a much higher rate.

The same question, a different answer every day

Ask ChatGPT, Gemini or Google's AI Mode the same question on two consecutive days and you will often get different sources. Until recently that was an anecdote. Two studies published in October 2026 now put numbers on it, and on what makes a page more likely to be cited in the first place.

The first, published by Ahrefs on 29 October, asked five AI tools the same questions every day for 20 days and followed every website cited on day one. The dataset covers 7,831 questions and about 37,500 citations across Google AI Overviews, AI Mode, ChatGPT, Gemini and Copilot. The second, published by SISTRIX on 22 October, compared 48,527 pages cited by ChatGPT, AI Mode or AI Overviews with uncited pages from the same domains, 96,418 URLs in total, to see which page features are associated with being cited.

Together they answer two questions that matter to anyone tracking "AI visibility": how much of what you see on a given day is noise, and which on-page changes are worth making.

50–83% of day-one citations gone the next day, depending on the platform (Ahrefs)

~75% of lost citations returned within 20 days

2.1× citation rate for pages with many structural features versus barely optimised pages (SISTRIX)

Why citations churn

The volatility does not come mainly from the AI rewriting its searches. Tools such as ChatGPT break a question into several background searches, a process often called query fan-out. Ahrefs found that ChatGPT repeated exactly the same fan-out searches as the day before only 26% of the time. But even when the searches stayed the same, the cited websites changed 81% of the time, against 91% when the searches changed. Most of the churn happens after retrieval, when the model chooses which of the pages it found to cite.

Gemini and Copilot add another layer. 64% of Gemini answers and 53% of Copilot answers in the study cited no sources at all, against less than 1% on ChatGPT, AI Overviews and AI Mode. A typical AI Overview cited 10 sources and a typical AI Mode answer 9, against 4 on ChatGPT. On Gemini and Copilot, many apparent "losses" were simply days when the tool cited nobody, and many "comebacks" were days when it started citing again.

The kind of question matters too. "Best X" questions had among the most volatile citations, because many pages compete to answer them. Questions that named a brand were the most stable. When a question named the brand, the brand's own website stayed cited on all 20 days in 28% to 33% of cases, against 4% to 7% for any cited site. In a smaller test of 22 questions about Ahrefs itself, the company's site was cited on almost every day for factual questions, but rarely on Google's tools for reviews and comparisons, where third-party sites took over.

Share of days on which the brand's own website was cited, 28 August to 4 October 2026. Source: Ahrefs. The sample is small and covers a single brand.

Brands are more stable than links

One of the more useful findings for marketers is that a brand name tends to survive in an answer longer than a link to the brand's website. Across platforms, most cited websites were not named in the answer text at all: 84% on AI Mode and 54% on ChatGPT. And when a citation disappeared the next day, the brand was sometimes still named. On Gemini this happened 40% of the time. Ahrefs found that 50% to 70% of cited websites appeared on only one day, against about one in five brand names.

Other trackers see the same concentration. Refine, which ran 17,691 prompts for 20 brands across ChatGPT, Gemini and Perplexity between May and August 2026, found that the top 20% of brands took 67% of all mentions, while 36% of brand-platform pairs were never mentioned at all. The same brand could score above 75% on one engine and close to zero on another, which the authors attribute to different retrieval systems rather than brand preference.

A citation on one day tells you very little. A citation that has lasted a week tells you a lot, and a brand mention tells you more than either.

What makes a page citable

The SISTRIX study looks at the problem from the page side. Because it compares cited and uncited pages from the same domain, it controls for brand strength, which is usually the biggest confounder in this kind of analysis. It tested 43 features with conditional logistic regression. A factor of 1.5 means the feature was about 1.5 times as common on cited pages as on comparable uncited ones.

Structure dominated. A table of contents (1.7), Article schema (1.6), at least five H2 headings (1.6), FAQPage schema (1.5), tables with a header row (1.5) and pages with ten or more bullet lists (1.5) were all associated with more citations. So were trust signals: an author box with a photo (1.4), a visible last-updated date (1.3) and at least ten external links (1.3). The current year in the title was the strongest single signal at 2.7, though SISTRIX treats title effects as preliminary.

Just as interesting is what did not help. llms.txt, a proposed file for telling AI crawlers what to read, showed no advantage: 11.7% of cited URLs came from domains with llms.txt, against 14.0% of all domains. Readability scores made no difference. Product schema was linked to fewer citations (0.8) and Recipe schema to far fewer (0.1), which says more about what AI tools like to cite, advice and explanation, than about the markup itself. Online shops did better with an advice section carrying Article schema than with optimised product pages.

Combined, the effects add up. Pages that met many of eleven key features were cited 2.1 times as often as barely optimised pages, yet only 10% of cited pages met many of them. SISTRIX also offers a warning about the field itself: in its tests, a sample of 833 URLs showed FAQ schema as a negative factor, while the full sample showed it as positive. Results only pointed reliably in one direction from about 5,000 URLs, so GEO studies smaller than that deserve scepticism.

The two studies agree on one point that matters more than any markup. Ahrefs found that on AI Overviews and AI Mode, pages that ranked in Google for the question were about seven times as likely to stay cited for a week as pages that did not. Pages with more organic traffic stayed cited longer, while Domain Rating and backlinks made little difference. A closer match between the page's title and description and the question was also associated with longer citations. For Google's AI tools, good classic SEO is still most of the work.

How much traffic is really at stake

Citation stability matters only as much as the traffic and revenue behind it, and the October data on that is sobering. Marie Haynes analysed twelve months of GA4 data from 100 websites and found that ChatGPT sessions equalled 0.62% of Google organic sessions, and all AI assistants together 0.73%. The SEO Works, looking at 95 UK sites, found that organic search sent 39 times as much traffic as AI. iPullRank, using a matched panel of 16,310 US desktop users, found that AI accounted for only 2.7% of a typical user's searches.

The visitors who do arrive often convert well. In the Haynes sample, ChatGPT visitors converted at 13.8% against 3.02% for Google organic, though Google organic still drove 96.5% of all conversions. The calculator below applies those ratios to your own numbers. It is an illustration based on one sample, not a forecast.

Measuring even that small slice is getting harder. Between 9 and 16 October, many ChatGPT links lost their utm_source=chatgpt.com tag after an interface change, according to SISTRIX, and because those links pass no referrer, the visits showed up as direct traffic. Google added utm_source=gemini to Gemini's cited links in early September and quietly stopped around 28 September. Links opened from the Gemini iOS app carry no referrer at all, so those visits also land in GA4 as direct. Kevin Indig, looking at twelve GA4 properties, found that direct traffic rose sharply in 2026 without any reliable link to AI visibility.

What this means for Indonesian websites

None of these studies used Indonesian data. The Ahrefs questions were mostly in English, and the SISTRIX sample comes from German-speaking countries. That matters, because the pool of pages competing for a question in Bahasa Indonesia is smaller, and AI tools may behave differently with it. Google AI Mode has been available in Indonesia, in Bahasa Indonesia, since September 2025, and AI Overviews before that, so the same dynamics apply here even if the exact numbers differ.

The practical lessons travel well. A single day's AI visibility report is mostly noise; the useful measure is how many of the last seven days a page or brand was cited or named. A citation that has lasted a week and then drops is likely to return, while one that lasted a day may not. For Google's AI tools, ranking in Google remains the strongest predictor of staying cited. And the page features linked to citations, clear headings, tables of contents, real FAQs, author information and update dates, are the same ones that make pages easier for people to read. We have seen how much AI answers now dominate the results page in our analysis of AI Overviews on branded searches; these studies show how unstable the sources behind those answers are.

What to watch next

• A SISTRIX follow-up. The authors plan a second study. Results that hold outside the DACH region, and for Perplexity, Claude and the Gemini app, would make the findings far more general.
• Tracking fixes. Whether OpenAI and Google restore consistent UTM tags or referrers for AI links, or whether AI traffic keeps disappearing into direct.
• Ads in AI answers. Google is expanding ads in AI Mode to more countries, and ChatGPT is widening its ad formats. Paid placements inside answers could change how much room is left for organic citations.
• Indonesian-language data. No published study yet measures citation stability for Bahasa Indonesia queries.

Frequently asked questions

How stable are AI citations?

Not very. In an Ahrefs study of about 37,500 citations tracked for 20 days, 50% of AI Overviews citations, 58% of AI Mode citations, 59% of ChatGPT citations, 81% of Gemini citations and 83% of Copilot citations were gone the next day. Only 4% to 7% lasted all 20 days on Google's tools and ChatGPT.

Do lost AI citations come back?

Usually. About three quarters of lost citations returned within 20 days, and about half returned within three days. Citations that had lasted more than a week almost always came back, while about a third of one-day citations never did.

What page features are linked to more AI citations?

In SISTRIX's study of 96,418 URLs, the strongest structural signals were a table of contents (1.7 times), Article schema (1.6), at least five H2 headings (1.6), FAQPage schema (1.5) and tables with a header row (1.5). Author boxes with photos, visible update dates and external links also helped. Pages with many of these features were cited 2.1 times as often.

Does llms.txt help get cited by AI?

Not in the SISTRIX data. 11.7% of cited URLs came from domains with an llms.txt file, against 14.0% of all domains, so there was no measurable advantage.

Does ranking in Google help get cited by AI?

For Google's own AI tools, yes. Ahrefs found that pages ranking in Google for the question were about seven times as likely to stay cited for a week on AI Overviews and AI Mode. Organic traffic was also linked to longer citations, while Domain Rating and backlinks were not.

How much traffic does ChatGPT send to websites?

Relatively little. Across 100 websites, ChatGPT sessions equalled 0.62% of Google organic sessions, according to Marie Haynes. Another study of 95 UK sites found organic search sent 39 times as much traffic as AI. ChatGPT visitors did convert at higher rates in the first sample.

Why does AI traffic show up as direct in GA4?

Some AI links carry no referrer or tracking tag. ChatGPT links lost their UTM tag for about a week in October 2026, Google stopped adding UTM tags to Gemini links around 28 September, and the Gemini iOS app opens links without a referrer.

Sources

• How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 Citations, Ahrefs, 29 October 2026
• Citation Worthiness: Which GEO Tactics Hold Up Across 96,418 URLs, SISTRIX, 22 October 2026
• What changed in ChatGPT in October: GPT-6, UTM tracking and sources, SISTRIX, October 2026
• 17,691 AI Answers: 20% of Brands Took Two Thirds of the Visibility, Refine, 27 August 2026
• AI Search Traffic Studies (October 2026), summarising Marie Haynes, The SEO Works, Kevin Indig and Semrush
• How the Same 16,000 Users Behave in AI Search vs. Google, iPullRank, 29 October 2026
• Google Gemini Stopped Attaching UTM Parameters To Cited URLs, Search Engine Roundtable, 30 October 2026
• Google AI Overviews & AI Mode 0.9% Click-Through Rate, Opt Out Study Says, Search Engine Roundtable, 30 October 2026
• Google Search: Memperkenalkan Mode AI di Indonesia, Google Indonesia, 9 September 2025