Citation is an outcome, not a process
The current conversation around AI visibility is fixated on a single question: “Is AI citing us?” Entire toolchains have emerged to answer it — citation trackers, mention monitors, share-of-voice dashboards for LLM outputs.
But citation is an outcome. Before an AI system cites a page, something else has to happen first: the system has to read the page, and it has to understand it. Understanding is the intermediate state between input and output — and it is precisely this intermediate state that nobody is observing.
The industry is optimizing for a result while remaining blind to the process that produces it.
The two observed layers — and the one between them
Observation of AI behavior toward a website currently exists at two layers:
The input layer asks: what does AI actually read? This is the domain of AI bot observation — recording which crawlers and retrieval agents access which pages, when, and how often. Frameworks such as EdgeShaping operate here. The input layer establishes the factual baseline: you cannot be understood, let alone cited, by a system that never fetched your page.
The output layer asks: what does AI say? This is the domain of citation trackers and answer monitors — sampling AI-generated responses and checking whether, and how, a domain appears in them.
Between these two layers sits a question neither can answer: given that AI read this page, what did it understand it to be?
A page can be fetched regularly and still be misread. It can be understood as something adjacent to — or entirely different from — what its author intended. The input layer will show healthy bot traffic; the output layer will show absent or distorted citations; and neither will tell you why. The gap between the two is the understanding gap, and it is currently invisible.
Defining PUL
PUL — the Page Understanding Layer — is a methodology for observing how AI systems understand each individual page of a site.
Three boundaries matter in this definition:
PUL is observation, not optimization. It does not prescribe how to write for AI. It reveals how AI currently reads what you have written. What you do with that revelation is a separate, subsequent decision — the same relationship that holds between analytics and marketing.
PUL is per-page, not per-site. AI systems do not understand “your site.” They understand — or misunderstand — individual documents. A domain-level notion of AI comprehension averages away exactly the signal that matters: which specific pages are understood as intended, and which are silently misread.
PUL is distinct from visit observation. Knowing that GPTBot fetched a page 40 times last month is an input-layer fact. Knowing what a model would say that page is about is an understanding-layer fact. They are different measurements answering different questions, and conflating them reproduces the industry’s existing blind spot at a smaller scale.
What PUL is not
PUL is not an AIO checklist. Checklists assume the failure modes are already known — add structured data, clarify headings, and understanding will follow. PUL makes no such assumption. It measures first.
PUL is not citation tracking. Citation tracking observes the output layer and can only confirm that something went right or wrong downstream. It cannot localize the failure to a page, nor distinguish “never read” from “read but misunderstood.”
PUL is not a scoring system. A single “AI-readiness score” is a compression of the very information PUL exists to expose: the specific, per-page divergence between intended meaning and machine-derived meaning.
PUL and ReverseHallucination
ReverseHallucination is the strategy of publishing targeted content to shape what AI systems say about a domain. PUL is its natural verification layer: if ReverseHallucination is the act of writing toward a machine’s understanding, PUL is the act of checking whether that understanding actually formed. Without an understanding-layer observation, ReverseHallucination can only be evaluated at the output layer — slowly, indirectly, and confounded by everything else the model has read.
The gap you cannot see until you look
Every page carries two meanings: the one its author intended, and the one a machine derived. When these coincide, nothing is visible — the page simply works. When they diverge, nothing is visible either — the page silently fails, cited for the wrong thing or not cited at all, while its bot traffic looks perfectly healthy.
The divergence between intended meaning and machine-understood meaning does not announce itself. It exists whether or not anyone measures it. PUL is the proposition that this gap should be observed — page by page — before anyone attempts to close it.
PUL (Page Understanding Layer) is a concept defined and published by Kenichi Uchiumi at mare-interno.com.