Getting Mentioned by ChatGPT Wears Off in About a Month

Scrunch tracked 3.5 million citation events across six AI platforms between September 2025 and March 2026 and found that the median one is half gone in about four and a half weeks. On ChatGPT it’s 3.4 weeks. On Perplexity, 5.8. Google’s AI Mode, Gemini and AI Overviews land between 4.3 and 4.8.

That’s the number to hold next to whatever you’re being quoted for answer engine optimization, the work of getting your brand named inside an AI’s answer rather than ranked in a list of links. The retainer runs twelve months. The result runs about one.

A citation is a lease, and nobody tells you the term.

The decay isn’t the whole problem either. Between February 8 and April 27 this year, seoClarity measured citation volumes in ChatGPT falling 86% to 94% across five markets: the US, UK, Canada, Germany and Italy. The share of US prompts that returned no citation at all doubled in March, from 28% to 48%. In Germany it reached 85% by April. seoClarity puts the inflection points at March 8 and April 19 and reads them as OpenAI leaning harder on what the model already knows instead of fetching the live web.

Mitul Gandhi, Chief Architect and Co-Founder at seoClarity, drew the conclusion plainly:

“AI search will likely always be volatile”

Volatile is doing a lot of work in that sentence. Semrush ran a 230,000-prompt study from August to mid-September 2025 and watched Reddit go from appearing in roughly 60% of ChatGPT’s responses to 10% in six weeks, then recover. Sergei Rogulin, who runs organic and AI visibility at Semrush, read it as the platform trying to avoid over-citing certain sites. Nobody outside OpenAI knows for sure, which is the point. A 50-point swing in six weeks, on the single most-cited domain on the internet, with no announcement.

Now put that next to where the citations go in the first place. 5W’s Citation Source Index, published May 1, synthesized more than 680 million citations from six published studies and found the top 15 domains take 68% of the whole consolidated citation pool, with Reddit first at around 40%. Wikipedia, YouTube, Forbes, Business Insider and the NIH carry much of the rest.

So the shape of the thing is two bad facts stacked. Most of the citation supply belongs to fifteen websites you don’t own. And the slot you do manage to win on one of them is half gone in a month.

Ronn Torossian, who founded 5W, framed the shift in the grandest possible terms:

“For 25 years, the most consequential algorithm in communications was Google PageRank. That era is over.”

Two caveats before anyone repositions a budget on this, and they matter. 5W sells AI visibility audits, and its index is a synthesis of other people’s studies rather than its own measurement, so treat the 68% as directional. Scrunch’s study also found that citations from the 4,000-plus publishers in the Stacker Partner Network last roughly twice as long as everyone else’s, which is a finding that happens to describe a product Stacker sells. The half-life numbers were built from weekly decay curves with bootstrapped confidence intervals, so the method is real. The recommendation that follows it has an owner.

The mechanism underneath all of this is simpler than the acronyms suggest. A search ranking is a stored judgment about a page: once you earn it, it persists until something changes it. A citation is decided fresh every time somebody asks a question, by a model that was retrained last month, running retrieval rules that a vendor tunes without telling anyone. There’s nothing for the win to sit in. That’s also why attribution for this traffic keeps failing and why the industry has spent the year arguing about what to even count.

James Bishop, VP of Marketing at Vanillasoft, described the actual change to the work:

“Content that was once written only to get organic search results now needs to be written with the mindset of ‘what do I want the AEO engines to say about me’”

And Boris Dzhingarov, CEO of ESBO, named the reporting problem a founder-marketer runs into the week they try to measure any of it:

“AI-assisted discovery doesn’t leave a breadcrumb in GA4. No last-click. No source. Just a converted user with mysterious origins.”

For a 20-person B2B SaaS the practical consequence is a budgeting one, not a tactical one. A one-month half-life turns AEO from a project into a standing cost, the way paid media is a standing cost. Stop treating it as a fixed asset you build once. If your plan is a content sprint that ends, the results end roughly a month after the sprint does, and you’ll conclude the tactic failed when the lease simply ran out.

So the single move: before signing anything AEO-shaped, ask the vendor what their citations look like eight weeks after the work stops. If they’ve never measured it, they’re selling you a ranking model for a market that doesn’t have rankings.