Every Model Has a Different Diet
Why your website needs to become a canonical buffer, not just a landing page
There is a comforting story going around about AI visibility. It says the models are like pets. Feed them every day, feed each one its favorite food, and eventually they will learn to love you. Post enough, repeat your name enough, and ChatGPT will start saying it back.
I understand why the story spread. It gives you something to do every morning. It turns an opaque system into a creature you can tame.
The story is wrong, and the data on how it is wrong turns out to be more useful than the story itself.
The machine can eat your content and feed your competitor
In the first half of 2026, Ahrefs ran the cleanest experiment on this I have seen. Between February and May they published 34 self-promotional pages across five domains, promoting two of their own brands, and then tracked 9,886 AI answers from ChatGPT, Gemini, Perplexity, and Copilot to see what the pages actually did.
Here is the number that should end the pet story. Among the answers that cited one of the pages promoting their new conference, 43% never mentioned the conference at all. The AI used the page as a research source and then recommended competing events, sometimes events listed on that very page. When a page was pulled into the answer pipeline but not shown as a citation, the brand was skipped 74% of the time.
The page got eaten. The competitor got fed.
The same experiment showed something else. For a brand new entity, the self-promotional pages worked as a bridge from absent to present: 82% of the new mentions appeared in answers citing the brand’s own pages. For the already established brand, only 6% of new mentions came from its own pages. The other 94% came from what third parties had written.
Two entities is a small sample, so I will keep the conclusion narrow. In this experiment, self-published pages mattered far more for the new entity than for the established one. Once a market has a dense external record about you, your own pages may become one input among many, and not necessarily the decisive one. The model increasingly eats from other people’s plates. And at every stage, volume without a provable link between fact, entity, and category can inform the system without making it choose you.
More food is not the goal. Provable food is.
You are not feeding a pet. You are maintaining a habitat
Here is the correction that makes everything else make sense.
There are two kinds of memory in this system, and founders confuse them at their own risk.
Private memory belongs to a user relationship. An assistant may remember what one person told it, inside one account or one continuing conversation, and that memory can even shape how it searches for that person. Public memory belongs to the evidence environment. It is what allows a stranger, in a fresh session, to receive approximately the same current account of a person or business.
Private memory is not public legibility. Personalization can make an answer feel familiar. Only public legibility can make it reproducible.
For public visibility, you cannot rely on the system having met you before. When a fresh user asks about you or your business, the answer may come from prior training, active web retrieval, or a mixture of both. When retrieval enters the process, the system moves through the public information environment available to it and assembles an answer from what it can reach and support. Then the next question starts the process again.
You do not train the public model to remember you. You maintain the habitat from which strangers’ answers are rebuilt.
This is not a poetic downgrade of the daily-feeding idea. It is an upgrade, because habitats have properties that bowls do not. A habitat can be tended or neglected. It can be internally consistent or contradictory. It can contain fresh, dated, verifiable material, or a sediment of old bios and abandoned profiles that quietly outvote everything you published this year.
And crucially: the same habitat is foraged by very different species.
Five engines, barely one shared web
In a study published in mid-2026, SurfacedBy logged the sources cited by five AI engines answering the same set of real buyer questions: ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode. Across roughly 16,400 answers collected between March and June 2026, they recorded 127,198 citations pointing at 11,647 different websites.
Roughly seven in ten of those domains were cited by only one engine. Just 2.7% were cited by all five.
This was a vendor study of commercially oriented questions, counted at the domain level. It describes the citation layer, not the full retrieval process or the entire web. Still, the degree of divergence is too large to treat the engines as one distribution channel. A page one engine loves can be invisible to another. Optimizing for “AI search” as a single channel means optimizing for an average that no engine actually is.
It gets stranger. The diet is not even stable inside one system. Semrush and Kevin Indig ran 100 prompts through ChatGPT twice, once in minimal reasoning and once in high reasoning, and found that only 25.6% of cited domains overlapped between the two modes. In the deeper mode, Reddit’s share of citations fell by roughly half while government, academic, and official documentation sources gained ground, and the system ran nearly five times as many background searches.
The sample is small and the study is a commercial benchmark, not a law of nature. But whatever routing logic sits behind the interface, the test shows that changing reasoning depth alone was enough to produce a materially different source environment. Appetite changes with the question. The unit of strategy is not “the model.” It is model by mode by question by category. A founder, a clinic, a SaaS product, and a beauty brand do not just have different audiences. They have different evidence diets, and those diets shift with the seriousness of the question being asked.
Citation churn is not recommendation churn
Two findings, from two separate studies, deserve their own section, because together they explain what the habitat is actually for.
First, citations flicker. In the Ahrefs experiment, about a quarter of pages were cited once for a given question on a given engine and never appeared again. Even recurring citations showed up on only about one in three eligible days between their first and last appearance. Nothing about the pages changed. The sources under the answers simply kept rotating.
Second, a separate SurfacedBy analysis found substantial source turnover when the same buyer questions were repeated over time, ranging from roughly a quarter of sources changing between checks on Perplexity to well over half on Google AI Mode. In one manually inspected narrow query, the top recommendation stayed the same while the sources beneath it kept changing. The authors are careful about the limit, and so am I: that was one spot check, not a measured rate of recommendation stability. But it points at a distinction worth keeping.
Citation churn and recommendation churn are not the same event. Losing a specific citation does not necessarily mean losing the position. Keeping a mention does not mean controlling the argument the system uses to explain it.
This is why a habitat needs a fixed point. If the sources under every answer are in constant rotation, the only durable asset is a version of the facts that survives the rotation: one place where the current, dated, internally consistent account of the entity lives, so that whichever sources the engine happens to forage this week, they resolve back to the same story.
That place should be your website. But not the website most businesses have.
Your website is a canonical buffer
Most websites are built as landing pages: who we are, what we sell, leave your email. That remains a real job, and for commerce it is still the first job. But for public legibility, the website now has a second job that many businesses have not designed for.
I call that second job a canonical buffer.
A canonical buffer is the controlled layer between what is true inside a business and what AI systems can prove outside it. It holds the current, dated, verifiable version of an entity: its one-sentence definition, its facts, its terms, its history of updates. External platforms distribute and confirm that version. The buffer is where it is kept current.
The word buffer matters. A buffer absorbs the mismatch between two systems moving at different speeds. Your business changes weekly. The external record of your business, old interviews, stale bios, third-party mentions, changes slowly and unevenly, and the engines forage across all of it. The engines may reproduce the version that is most consistently supported by the sources they can retrieve, not necessarily the version you published most recently. The buffer is the layer that works to make those two versions the same version.
There is evidence that controlled and manageable sources can dominate in some commercial contexts. In Yext’s location-first study of 6.8 million citations across retail, food, healthcare, and financial services, 44% of citations came from first-party websites and another 42% from third-party listings that businesses could manage. That does not mean 86% of every founder’s or global company’s information environment is controllable; the study was built around local business discovery. But it demonstrates something important: in categories where structured facts, locations, services, and availability matter, a large part of the evidence layer may already be operationally manageable. It is just usually running unattended.
To be precise about the limits: a buffer is a necessary layer, not a proven ranking factor. No study I have seen demonstrates that any single structure guarantees citation, and OpenAI itself states that crawlability makes presence possible while guaranteeing nothing about prominence.
Maintenance beats production
If the habitat picture is right, the daily practice changes. Not daily publishing. Tending, on a steady rhythm.
Start with freshness, because it is where the pet story hides. In a 2025 analysis of roughly 17 million AI-cited URLs, Ahrefs found they were on average about 25.7% fresher than Google’s organic results for comparable queries. The average AI-cited page was still almost three years old. AI-cited pages tend to be fresher, but freshness is not a demonstrated causal lever on its own: the data does not prove that updating a page caused its citation. My practical conclusion is narrower. Do not manufacture freshness through daily posts or cosmetic date changes. Maintain the pages that carry your canonical facts, and update them when the underlying facts actually change.
And one caution from the academic side. A critical survey of 45 GEO studies from 2023 to 2026 describes visibility not as a ranking task but as a long stochastic pipeline: search activation, indexing, retrieval, reranking, citation, actual influence on the answer, and finally user behavior. The survey notes that no technique in the reviewed literature has shown a durable, cross-platform, causal effect on organic visibility and business results. It also records an uncomfortable effect: rewriting pages specifically for quotability sometimes hurt retrieval. Content can become more quotable and less findable at the same time.
So the honest promise is modest and still worth making. You cannot train the species. You can make your habitat the easiest place in their range to find the true, current, provable version of you.
Three questions before you build anything
The full audit method lives as a separate page: https://katyashalel.com/guides/buffer-audit/. For this essay, three questions carry most of the weight.
What must the public record make easy to prove? Not what you want said about you. What a system assembling an answer for a stranger must be able to support: who you are, what category you belong to, what you are the answer to.
Which version of you currently has the strongest evidence? Ask the engines who you are and watch which sources they reach for. The version with the most consistent support wins, and it is not always the current one. Find what is outproving you.
Which important claims exist only in your own words? A fact that appears solely in your own materials is weak food. Identify the two or three claims that most need independent confirmation, and go earn it there.
The pet story asks how often you should post. The habitat question asks something harder: what does the public record currently make easiest to prove?
You cannot force every engine to take the same route. You can arrange the evidence so that, whichever route it takes, the current truth is easier to reconstruct than the obsolete version.
That is the work. Not feeding the machine, but building an environment in which the truth survives retrieval.
I’m Katya Shalel, a founder and legibility strategist. Canonical version of this essay: https://katyashalel.com/essays/every-model-has-a-different-diet/. Observations about specific engines describe cited studies from 2025 and 2026 test sets and will change as the systems change.