ChatGPT
Claude
Gemini
GrokGetting cited by ChatGPT, Gemini, Claude, or Perplexity feels like a win. It is a win, but it may not be a permanent one. The source that appears in an answer today can be replaced as the model searches again, new pages are published, or the question changes slightly.
The clearest public evidence so far comes from a study by Scrunch and Stacker that tracked 3.5 million citation events from September 2025 through March 2026. The average citation lost half of its activity in about 4 to 5 weeks. Read the full citation half-life study for the method and the platform breakdown.
The uncomfortable truth: AI citations are rented visibility
Traditional SEO trains teams to think in positions. You improve a page, it moves up, and you try to hold the spot. AI search is less stable. A citation is a source choice made for a particular answer at a particular time. It can be useful without becoming a lasting position.
The study found meaningful differences between platforms. ChatGPT citations turned over fastest, with a half-life of about 3.4 weeks. Perplexity was closer to 5.8 weeks. Google AI Mode, Gemini, and AI Overviews sat between roughly 4.3 and 4.8 weeks. These are not rules enforced by the products, and the study is industry research rather than a peer-reviewed ranking formula. Still, the gap is large enough to change how you plan content work.
Signal | What it tells you | Action to take |
ChatGPT | About 3.4 weeks | Recheck important prompts often and watch for source replacement. |
Perplexity | About 5.8 weeks | Build useful source pages, then test whether they remain cited. |
Gemini and Google AI surfaces | About 4.3 to 4.8 weeks | Keep facts, local details, and time-sensitive pages current. |
Any platform | About 4.5 weeks on average | Treat a citation as a signal to maintain, not a finish line. |
This is why GeoSnake's AI visibility tracking matters. The useful question is not just whether your brand appeared. It is whether it still appears for the same buyer question, in the same market, with the same facts and a source you would be comfortable sending a customer to.
A citation is not the same as influence
A second research paper makes the measurement problem sharper. In a controlled dataset covering 602 prompts and more than 21,000 search-layer citations across ChatGPT, Google AI results, and Perplexity, the platforms differed in both citation volume and citation influence. Perplexity cited more sources per prompt on average, while ChatGPT cited fewer sources but showed higher average influence among the pages it used.
The paper calls this difference citation selection versus citation absorption. A page can be selected and linked, yet contribute very little to the final answer. Or it can supply the definition, number, comparison, or step that shapes the answer. See the citation absorption research for the dataset and the limits of the finding.
That distinction changes what you should improve. Chasing more citations can be a vanity project if the answer does not use your evidence. A page with fewer citations but strong factual contribution may be more valuable than a page that appears in many low-impact source lists.
Why different models decay at different speeds
There is no single AI index. ChatGPT can decide whether to search and can return inline sources. Claude can use web search and provide direct citations. Perplexity is built around real-time, cited answers and also routes queries across different models. Their retrieval, refresh, and source-selection paths are different, so a citation that survives in one system may disappear in another.
OpenAI explains how ChatGPT Search uses web queries and citations; Anthropic documents Claude's web-search citations; and Perplexity describes its real-time, cited answer engine. The business implication is simple: do not treat one model's citation as proof of cross-model visibility.
What businesses should do before a citation fades
Do not respond by publishing low-value updates every month. The goal is to refresh evidence, fix inaccuracies, and strengthen the sources that answer your buyers' real questions. A recent critical survey of 45 GEO studies found that no reviewed technique had yet shown a stable, long-term, cross-platform causal effect on organic discoverability or downstream behavior. That is a useful warning against easy promises. Read the
new GEO evidence survey before buying any claim that one content trick guarantees lasting AI rankings.
Choose the 10 to 20 buyer questions where a recommendation or citation could change a deal, not just the questions with the highest search volume.
Run the same prompts across ChatGPT, Claude, Gemini, and Perplexity on a fixed schedule, recording model, date, market, competitors, sources, and the exact claim attached to your brand.
Separate selection from influence: mark whether your page was merely listed, actually used for a fact, or used to support a recommendation.
Refresh the page only when there is a real reason, such as a changed price, product, policy, proof point, comparison, or customer question, then test whether the updated evidence survives.
Use the GeoSnake weekly AI visibility workflow to keep this record consistent. GeoSnake can help you see the prompt-level movement, source changes, competitor mentions, and model differences that a single monthly score hides.
The real maintenance advantage
The half-life study found that editorial domains in the Stacker partner network lasted roughly twice as long as the average non-network domain. That does not prove that every brand can buy durable visibility through a publisher, and the partner network is part of the study's sample. It does suggest that trusted, independent sources may have more staying power than a brand page repeating its own claims.
So the durable strategy is two-part. Make your own pages clear enough for models to extract facts, then earn independent coverage that confirms those facts. Keep the facts identical across your site, profiles, reviews, partner pages, and public documentation. When sources disagree, models have less reason to trust any one of them.
“AI visibility lasts when clarity meets trust.”
If your team is still treating AI visibility as a one-time content sprint, change that now. The better operating model is a small maintenance loop: observe, diagnose, improve the evidence, and check again. That is slower than publishing a pile of generic pages, but it is the only approach that matches how unstable AI search actually is.
FAQs
How long does an AI citation usually last?
A Scrunch and Stacker study of 3.5 million citation events found an average half-life of about 4 to 5 weeks. The number varied by platform, with ChatGPT turning over faster and Perplexity lasting longer in that dataset.
Does updating a page guarantee that AI will cite it again?
No. An update can make evidence clearer or more current, but retrieval and citation are not guaranteed. Test the same buyer prompts again and look for a change in the answer, not only a change in the source list.
What is more important, citation count or citation influence?
Influence is usually the stronger signal. A page that supplies the key fact, comparison, or recommendation reason can matter more than a page that is listed but barely used. Track both when you evaluate AI visibility.