AI visibility doesn’t end, it’s maintained. That page ChatGPT cites today doesn’t stay there by inertia: freshness weighs, the model gets retrained every few weeks and, meanwhile, someone ships something newer than you. Leave it alone and the citation cools, and someone else takes your slot. This guide is about maintenance —the concrete action of refreshing content to keep the citation—, not about detecting the drop: that’s measuring, this is fixing.
Why AI visibility fades on its own (content decay in GEO)
Content decay is the silent loss of traffic and mentions a page suffers over time, without you changing anything. In classic SEO it’s slow: a good page holds for months or years. In GEO it goes faster, because the model favors what’s recent and rewrites its answer the moment it finds something more current.
The mechanics differ from Google’s. A page can keep ranking in search and at the same time vanish from AI answers: two independent drops. That’s why maintaining AI visibility is a separate job, not a byproduct of your usual SEO.
The timing varies by platform. According to Machine Relations, ChatGPT churns the most —it renews its sources around every 3-4 weeks—, while Perplexity holds a bit longer. The practical read is simple: there’s no «published and done», there’s a clock running from the day you publish.
The three reasons you lose citations: freshness, model change and competition
Losing a citation is almost never a punishment: it’s that you stopped being the best option for the model. Three things cause it, and it’s worth knowing which one is happening to you before you touch anything.
| Cause | What happens | What maintenance reverses it |
|---|---|---|
| Freshness | Your page ages and the model favors the recent | Refresh data, examples and the real update date |
| Model change | A retraining reshuffles which sources it retrieves | Reinforce entity signals and get back into the sources it reads |
| Competition | Someone ships something more complete or newer | Widen coverage and close the gap the other filled |
Freshness you fix on your own. The model change you don’t control, but you do control staying citable when it happens. And competition is fought by covering the topic better, not by shouting louder.
Maintenance cadence: how often to touch each page
Not all pages age the same, so not all get touched the same. The rule: the closer to the money and the more volatile the data, the shorter the cadence. A simple calendar avoids both abandonment and compulsive tinkering.
A cadence that works: money pages —services, comparisons, pricing— every month or two; evergreen guides every quarter; volatile data —figures, listings— the moment they change. Freshness isn’t touching for the sake of it: it’s keeping what the page says true today.
What to update in a refresh that keeps the citation
A refresh that works isn’t rewriting for the sake of rewriting. It’s adding what the model rewards: new information, true and easy to extract. Here’s what actually moves the needle each pass.
- New data and example. An updated figure, a recent case, a source from this year. That’s what tells the model «this is alive».
- Cover the gap. If a new question about the topic shows up, answer it on the page before someone else does.
- Citable structure. Use the refresh to leave each answer in a standalone fragment —this is citable content structure—, which the model extracts without reading three paragraphs.
- Real date. Mark the update only when you’ve actually changed something. The freshness signal counts if it’s backed by a real change.
And if the page has lost steam on the basics —weak title, buried answer, unclear listing— the refresh is the moment to optimize the whole page, not just to patch a stray data point.
From alert to action: what to do when the citation drops
Detecting the drop is half; the other half is acting fast, because in AI the window is short. When the alert fires —you stopped showing up in a prompt where you were— maintenance has an order.
- Confirm what dropped: which prompt, which model, since when. Without that you fix blind.
- See who took your slot: if a new competitor is cited, read what they have that you no longer do.
- Refresh the page with new data and gap coverage, not with a cosmetic facelift.
- Re-push the external signals if the cause is entity or consensus, not just text.
- Measure again a few weeks later: the model takes time to pick up the change, don’t expect an effect the same day.
The alert that triggers all this you don’t call by eye: a system of visibility change alerts builds it. That guide is about detecting the dip; this one, about what you do with the detection.
Maintenance mistakes that speed up the loss
There are ways of «maintaining» that make things worse. These are the failures that speed up the drop instead of slowing it.
- Publish and forget. The base mistake: treating the page as finished. In GEO nothing is finished, only updated or stale.
- Bumping the date without touching the content. The model compares the text; the fake date doesn’t fly and it costs you credibility.
- Refreshing everything at once and then nothing for a year. Maintenance is cadence, not a heroic sprint every once in a while.
- Adding filler each pass. Lengthening without new data dilutes the listing and makes it less extractable, not fresher.
If you’d rather have this maintenance run without depending on you remembering —cadence, refresh with new data and reaction when something drops— GEO monitoring watches continuously where and when you lose the citation, and GEO optimization does the labor of the refresh on your content.