In March, ChatGPT recommends your brand when buyers ask for options in your category. You screenshot it. You share it in Slack. You count the project as done.
By June, you are gone from the answer. A competitor took your spot. Nobody noticed for weeks, because nobody was looking.
This is the most expensive misunderstanding in AI visibility right now: treating a mention as a permanent asset. It is not. It is a position you hold, and positions erode.
The evidence: visibility has a half-life
Research into prompt-space measurement puts numbers on this. The Prompt-Space Occupancy Score (PSOS) framework, published by Joseph Mas, tracks not just whether a brand appears in AI answers, but whether it keeps appearing at 30, 60, and 90 days. The decay findings should worry anyone running one-off optimization projects.
In one fintech case, a brand celebrated 700% visibility growth. Measured properly, it held 32% of category prompts and lost nearly half of that within 60 days. The incumbent it was chasing held 71% occupancy with almost no decay. In a retail case, a “visibility surge” decayed 70% within 90 days. The spike made a great dashboard. It did not survive a quarter.
Different industries, same shape. Mentions fade. The question is how fast, and whether anyone is measuring.
Why mentions erode
Three forces work against your visibility at all times.
Model retraining. AI providers now ship frequent model updates, and each one can reshuffle what the model recalls about your category. The version of the model that recommended you in March is not the version answering in June. Your mention was never guaranteed a seat in the next release.
Source updates. Retrieval-based answers rebuild themselves from whatever the index holds today. When a listicle updates, a review site reranks, or a competitor publishes a stronger comparison page, the raw material behind your mention changes. The answer changes with it.
Competitive crowding. You are not the only brand that discovered AI visibility this year. Every quarter, more competitors publish citation-ready content aimed at the same prompts. A static position in a crowded field is a shrinking position.
None of these forces care that you optimized once, twelve months ago.
The snapshot trap
Here is the uncomfortable part for anyone buying AI visibility reports. A one-time measurement cannot see decay at all.
Measure once and you get a photo of a moving object. The photo might catch you at your peak, right after a campaign, in the one week a fresh listicle carried you into answers. The report looks great. The decision it informs is wrong.
This is why “we checked, you show up in ChatGPT” is close to meaningless as a finding. The finding that matters is whether you still show up at day 30, day 60, and day 90, and whether your share is rising or bleeding. Any vendor who measures once and invoices is selling you the photo.
Why some brands barely decay
The same research shows the other side: brands that hold 65% to 80% occupancy for months. Established healthcare sources. Entrenched B2B platforms. Their advantage is not spending, It is anchoring.
Anchored brands share two traits.
Structured anchors. Their identity lives in the sources machines trust and revisit: an accurate Wikipedia presence, clean Wikidata entries, complete schema markup, and analyst coverage that names their category. Anchors give every retraining run and every retrieval pass the same consistent story to find. Without anchors, associations drift a little with every model update.
Consistent provenance. Joseph Mas’s work on authorship makes a point worth internalizing: models retain information better when it is consistently tied to a determinable source. When your category definitions, claims, and data get republished across the web with your name attached, the association compounds. When they circulate detached from you, your visibility leaks to whoever gets cited instead. Consistency of attribution matters more than raw reach.
Decay, in other words, is not fate. It is what happens to unanchored brands.
Stop the leak: a maintenance system
The fix is a cadence, not a project.
- Baseline properly. Measure your share of real buyer prompts across ChatGPT, Gemini, Perplexity, and AI Overviews. Log everything with dates.
- Re-measure at 30, 60, and 90 days. Same prompts, same wording, same engines. Decay only shows up when the measurement repeats.
- Reinforce anchors quarterly. Audit your Wikipedia and Wikidata accuracy, your schema coverage, and whether third-party sources still describe you in your category’s language.
- Refresh citation sources. Identify which pages carried your mentions and keep them current. A stale source is a mention waiting to expire.
- Watch the prompts you lost. Every lost prompt names the competitor who took it and the source that put them there. That is your next quarter’s worklist.
The economics favor maintenance heavily. Winning a prompt back after a competitor anchors into it costs far more than holding it would have.
The takeaway
One-time optimization fills the bucket. Retraining, source churn, and competitors drill the holes. Brands that treat AI visibility as a quarterly discipline compound their position. Brands that treat it as a project quietly hand it back.