Writesonic tracked 23 million sources cited by ChatGPT, Gemini, Perplexity, and Google AI Overviews between April and June 2026. The finding that should rewire how marketing teams report AI-search visibility: 44% of cited pages appear exactly once and then disappear. The median citation lasts 11 to 15 days. After that window, the seat is gone.
This is a different kind of number than the ones GEO dashboards usually surface. A share-of-voice chart tells you that your brand was cited last week. It does not tell you that the citation already expired and was re-sold to a competitor. Treating answer-engine presence as a position you hold, rather than inventory you rent, is the measurement error baked into most current reporting.
A citation is a rented seat, not a ranking
In classic search, a top-three organic position is sticky. You earn it, you defend it, and barring an algorithm update or a determined competitor, it persists for months. Teams learned to measure it that way: rank, hold, track drift.
Answer engines do not behave like that. The Writesonic data describes a fixed-seat model. An AI answer cites a small, bounded set of sources, and 90% of new citations arrive by displacing a brand that was there before. Every new citation is an eviction. The total number of seats per answer barely moves, so visibility is close to zero-sum at the source level. When a competitor earns a seat in the answer for your category, the most likely outcome is not that the answer grew longer. It is that you left.
That reframes the unit of measurement. Counting appearances tells you how many seats you held at one sampling moment. It says nothing about how long you keep them, or how fast you get evicted. A brand cited 40 times this month with a 5-day average tenure is in a weaker position than a brand cited 25 times with a 30-day tenure, even though the first brand looks better on an appearance chart.
Durability varies by engine
The 11-to-15-day median hides real spread across platforms, and the spread is governable.
Per the report, Perplexity retains citations roughly twice as long as ChatGPT. That single fact changes how you allocate effort. If a Perplexity seat lasts twice as long as a ChatGPT seat, the return on earning it is higher per unit of work, and the cost of losing it is higher too. A measurement layer that blends all engines into one “AI visibility” number erases exactly the signal you need to prioritize.
The engines also differ in how they attribute. Per the Writesonic data, ChatGPT cites third-party sources 96% of the time, with vendor-owned domains accounting for only 4.3% of citations. Your own website is structurally unlikely to be the thing ChatGPT quotes about you. The seat you are competing for is usually on someone else’s page: a review site, an industry roundup, an analyst note, a comparison article. That is where the durable citations live, and where most brand-controlled GEO programs are not looking.
Perplexity adds another wrinkle. The report finds 52% of brands cited by Perplexity are not mentioned in the answer text the user reads. The citation exists in the source list while the brand name never reaches the human. Appearance counters that scrape the source list will record a win the reader never experienced. Durability and visibility-to-the-reader are separate quantities, and conflating them inflates the scoreboard.
What holds a seat longer
If 44% of pages are cited once and vanish, the practical question is what separates the one-and-done pages from the ones that hold tenure across weeks. The pattern in the data points to four properties, and none of them are produced by publishing volume.
- Original data. Pages with first-party numbers, surveys, or benchmarks give an answer engine something it cannot reconstruct from ten other sources. A statistic with a name attached is hard to displace because the displacing page has to out-evidence it, not just out-publish it.
- Expert commentary. Named, credentialed perspective is a durable citation anchor. Generic explainer content is the most replaceable inventory there is, which is why it churns at the one-citation level.
- Deep specificity. Pages that answer a narrow question completely outlast pages that answer a broad question shallowly. Breadth is easy to displace; a definitive answer to a precise question is the seat competitors struggle to take.
- Off-site validation. Because the durable citations often live on third-party domains, presence in the sources that engines trust matters more than polishing your own pages. Earning a mention in the review site or the analyst roundup buys tenure your owned domain cannot.
These are not growth hacks. They are the properties of content that an answer engine has a reason to keep choosing when a fresh challenger shows up every two weeks.
What to measure instead of appearances
Most AI-search reporting today answers “were we cited?” The data says the better question is “how long do our citations survive, and what is evicting them?” Three metrics operationalize that.
Citation tenure. For every page that earns a citation, track how many days it holds the seat before disappearing. Report median tenure per engine, not a blended brand average. A rising appearance count with falling tenure is a program losing ground while its dashboard turns green.
Eviction rate. Sample your category’s answers on a fixed cadence and record how often a citation you held last period is gone this period, and what replaced it. The replacement is competitive intelligence. It tells you which competitor page out-evidenced yours, and on which third-party domain the fight is happening.
Reader-visible share. Separate “cited in the source list” from “named in the answer text the user reads.” On Perplexity especially, those diverge by half. The second number is the one tied to actual brand recall.
Do this now
Pick your three highest-value category questions. Sample the answers from ChatGPT and Perplexity today, and again in 14 days. Record which of your citations survived, which were evicted, and what replaced the evicted ones. That two-week comparison turns the abstract “11-day shelf life” into your own eviction curve, and it tells you whether your current GEO work is buying durable seats or one-time appearances. Once you can see the eviction curve, fund the four durability properties on the pages that hold tenure, and stop counting appearances as if they were rankings.
This analysis synthesizes New Writesonic data shows AI citations have 11-15 day shelf life (Foundation Inc / Writesonic, June 2026), and builds on our earlier work on the AI-search measurement reckoning, the real KPIs five companies use for AI search, and why marketing is becoming a governance team.
Victorino Group helps marketing organizations build durability-based measurement for AI-search visibility instead of appearance counters. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation