A few years ago you showed up in AI answers, today you've disappeared — and you did nothing wrong. It's not a penalty: it's that models stop citing those who don't produce fresh signals, and the visibility earned in the past wears away over time. Meanwhile, the competitors who keep producing content push you down position after position, without you even noticing. Understanding the minimum level of activity needed to stay visible — and how to get back in if you're already out — is more accessible than it seems.
Three years ago AI cited you often. Today less, or not at all. It’s not an algorithm that penalized you: it’s the natural decay of an entity that stops generating fresh signals.
I see it with software houses, professional firms, and B2B manufacturers that between 2022 and 2024 had built a solid content presence and today find themselves “overtaken” by smaller but more active competitors. The good news is that it’s not a problem of lost authority, it’s a problem of missing maintenance of the entity. And maintenance can be put back in place.
In this article I explain how AI perceives the decay of an entity, what happens in the knowledge graph when you stop publishing, and how to build a minimal flow of signals that keeps you inside the answers.
What “entity decay” means for an AI model
When ChatGPT, Perplexity, or Gemini cite you, they’re not reading your site in real time. They’re reading a representation of your entity built on knowledge graphs and signals collected over time. If those signals stop updating, the entity “ages” relative to the real world, and the model struggles to resolve it correctly.
In the world of knowledge graph research this problem has a technical name: temporal degradation of entity linking. It’s the phenomenon whereby a system that in 2022 correctly linked you to your updated description, in 2024 starts to get it wrong because the context around you has changed and you haven’t.
Zhang et al. (2024) published on arXiv a model designed precisely to handle this problem, called CYCLE:
“CYCLE is a novel approach to solve EL task in temporal change.”
Translated: the authors recognize that entity linking (the process by which the model decides “which company is this text about?”) degrades over time and a cross-year contrastive mechanism is needed to mitigate it. They recognize it as a systemic problem, not as an exception.
The consequence for you is simple: the knowledge graph doesn’t remember you forever. If you stop generating new facts that connect it to you, the system starts to prefer more active entities when it has to choose who to cite in an answer.
Who suffers decay most quickly
In the same paper the authors observe an interesting detail: decay doesn’t hit everyone at the same speed. Entities with few connections in the graph — low-degree — are more fragile, because every lost signal weighs more on the total.
For an Italian SME this means: if your company has few external mentions, few citations, a thin Wikidata entry, and two social profiles, every month of silence counts as six for a competitor with a rich graph. Those who start low have to pedal harder to stay visible.
I connect this to a concept I’ve already covered in previous articles: author entity recognition and building E-E-A-T for AI are the way you add nodes to your graph. Without those, decay is a matter of months.
Six months of silence, then a burst of 10 articles in a week, then more silence.
What I observed across 20+ B2B brands over 12 months
Between 2025 and 2026 I tracked, longitudinally, over twenty Italian B2B brands — software houses, consulting firms, technical manufacturers — tracing their citations on ChatGPT and Perplexity with industry queries repeated month over month.
The pattern that emerged is fairly clear-cut. The brands that maintained a steady flow of content (at least one piece a week on their topic) and external mentions (at least one interview, guest post, or citation per month on third-party sources) stayed stable or grew. The brands that had “editorial pauses” of 3-4 months started to be cited less, on average from the fifth month of silence onward. By month twelve, almost all had been replaced in AI answers by a more active competitor.
It’s an observation, not a controlled study: twenty brands aren’t a statistical sample and the uncontrolled variables are many. But the pattern repeated often enough to rule out pure chance. Real analysis requires professional tools for continuous AI response monitoring, and even then it remains a directional indication.
Think of a software house in Trieste specialized in port logistics and rail-sea intermodal solutions: an extremely vertical niche, probably 4-5 real national competitors. If in 2023 they published a case study a month and today they publish a post every quarter, the graph around them is losing density while competitors in Northern Europe — who publish in English on LinkedIn every week — are increasing theirs. Who wins in Perplexity’s answer to “best software for Mediterranean intermodal container terminals”? Not whoever has the best product. Whoever has the freshest graph.
One new piece of content per week on your vertical area of expertise, not on the trend of the month.
The test you can run in 10 minutes
Before planning a refresh you need to understand where you stand. Three steps, all with free tools.
Open Wikidata and search for your brand. If you don’t have an entry, your public graph is weak regardless. If you do have one, look at the date of the last edit: if it’s older than 12 months, it’s a sign of decay.
Open Google’s Rich Results Test and paste the homepage URL. Look for “Organization” in the result: if it’s missing or has empty fields (founder, sameAs, foundingDate), the AI is working with an entity that has no backbone.
Open Google Search Console and look at the queries from the last 16 months, comparing 2024 and 2025-2026 on industry keywords. It doesn’t measure AI citations directly, but if the impressions curve on informational queries falls while the site is technically fine, your weight as an entity in the topic is declining.
Simple binary threshold: if two out of three of these signals are weak, your brand is in full-blown decay.
The mistakes I see most often
The “one-off” refresh. Six months of silence, then a burst of 10 articles in a week, then more silence. The AI doesn’t reward bursts, it rewards regularity. Two pieces a month for twelve months weigh more than twenty pieces in two months.
New content but off-topic. A port logistics software house that starts publishing about generalist artificial intelligence because “it’s trending.” It dilutes the area of expertise instead of reinforcing it. The graph associates you with a blurry cluster.
Schema.org never updated. The JSON-LD made in 2022 with the founder’s name, two sameAs to social profiles, and nothing else. Meanwhile you’ve opened a YouTube channel, been interviewed by an industry publication, added a location: none of this is in the markup. You’re telling the graph that you’re stuck in 2022.
Zero external mentions. Everything is bet on the company blog. But the graph is strengthened above all by third-party sources: interviews, guest posts, documented event participation, citations in industry magazines. Without external signals, you remain a self-referential entity.
What to do concretely over the next 90 days
Maintaining the entity isn’t a six-month project: it’s a habit. These are the minimums I’ve seen work on the brands that held onto their visibility.
- At least 2 external mentions per month on third-party sources relevant to the industry (interview, guest post, niche podcast, documented event with a web page).
- Quarterly update of the Organization schema and of external profiles (Wikidata, company LinkedIn, Google Business Profile if local). Each quarter, check whether there’s a new piece of data to add.
- One new piece of content per week on your vertical area of expertise, not on the trend of the month. Better 40 articles a year all on your core business than 80 scattered articles.
- Quarterly comparison with the 3-5 competitors the AI cites in your place: open ChatGPT and ask “best [your category] in [your geographic/sector scope]”, look at who comes up, check what they’ve published over the last 90 days.
It’s not a magic factor, it’s not enough on its own, and results don’t arrive in 30 days. But it’s the bare minimum to avoid suffering decay.
