Publishing for AI extraction
It's still unclear. Here's what we've figured out so far: Publish an editorial layer and canonical layer, and implement a correction loop.
If you work in media and have been following the last year of research on how AI systems filter and reshape information, you have probably already accepted the basic picture. Intermediation in large-language models (LLMs) is not going away.
It routes a growing share of how readers encounter your work, whether through assistants, search summaries, recommendation layers, or chat interfaces embedded in apps they already use. The intermediation layer also has its own rules for what it lets through and how it frames it.
Our earlier work has mapped those rules. A useful frame is a coordinate system with a handful of axes that together determine what passes, what gets filtered, and what gets reworded on the way out. Operators who understand the coordinate system can code-switch their vocabulary on the input side and make more of it through the intermediation layer than operators who do not.
Our research identifies four boundaries that determine if content passes or fails to get through LLMs:
1. Risk framing (technical vs. political)
2. Scale (individual vs. institutional)
3. Intent (objective vs. manipulative)
4. Scope (redistribution vs. restructuring)
By plotting inputs on these axes, we found that the same facts can either pass or trigger filters depending entirely on how they are framed relative to the "origin" of political safety.
Here we pick up where that frame stops. Code-switching is useful, but it is not enough. If you run an information service and want your work to reach readers through AI intermediation without losing what matters, you need a publishing practice, not just a vocabulary.
This is what that practice looks like in working order.
Jump: WHY code-switching is not enough | WHAT are the two layers of publishing | WHAT the canonical layer is and isn’t | HOW to implement a correction loop | WHERE to start
Code-switching alone does not scale
Code-switching can work as a one-time tactic: You write the story, then you pick different words so it passes through a filter. It works on a single piece. It does not scale, for three reasons.
First, it burns editorial time on every story. Every piece becomes two writing jobs instead of one. For a small newsroom this is a structural tax that will eventually be paid by cutting coverage.
Second, it hides the change from the team. The reporter writes one thing, someone else rewords it for AI, and over time nobody on the team has a clear picture of what actually reaches readers through the intermediation layer. The decisions drift into one person's head.
Third, it optimizes for passing through one system, today, in one language. The moment a vendor changes a base model, or a reader switches assistants, or the query is asked in a different language, the previous work has to be redone from scratch. You have paid the tax and got nothing durable for it.
A publishing practice fixes all three by treating AI legibility as a property of how you structure and maintain content, not as a per-story rewrite.
The two layers of a publication
We have come to believe that every piece of publication-grade content has two layers from now on, and we need to treat them as a pair.
The editorial layer is the piece as your readers will actually read it. It carries the judgment, the framing, the context, the voice, the institutions, the intent, and the structural analysis. It is the version you would hand to a thoughtful reader with an hour. It does not compromise. It is demonstrably written by a human for a human.
The canonical answer layer is a set of short, tightly structured units that a model can extract and serve cleanly. Each unit is something like: a question a reader might plausibly ask an assistant, a factual answer of around one to three sentences, a small set of supporting facts, and a link back to the editorial version. The answers are written so that the core of what you know survives intermediation, regardless of which model is on the other side.
You can publish both! Readers see the editorial layer. Machines index the canonical layer. The canonical layer is not a summary of the editorial layer. It is a parallel product, built from the same reporting, optimized for a different kind of reader, and stabilized in a way you can version and improve.
This is perhaps the new publishing for an intermediation layer that has its own rules for extraction that we need to get used to, much like SEO in the past.
What a canonical answer unit looks like in practice
We have not found any verifiable best practices here, so this is just how we envision this: A canonical answer unit has four fields, and each field does specific work. This of course is just a hypothesis based on our scrappy experiments.
Rather than simply being metadata or an info box on the webpage of an article, the canonical layer can be many things.
For example, a long-form investigation might be the canonical artifact, with a daily news brief, a video summary, and a translated edition as derivative layers pointing back to it.
Distinguishing these helps a newsroom decide where to invest depth versus reach.
- The question is a reader-plausible query, not a marketing phrase. If a real person would not type it into an assistant, it is the wrong question. Write the question in the reader's voice, not yours.
The question is the entry point. If it does not match how real readers ask, your canonical layer will never be surfaced by the intermediation system, because nothing in the real query stream will match it.
If you skip this discipline, you risk building a carefully maintained canonical layer that no reader ever actually reaches, because it answers questions nobody asked.
- The answer is short, declarative, and complete on its own. Two to three sentences. If the answer requires context that only makes sense inside the editorial version, the question was pitched at the wrong level. Go one step wider and try again.
Assistants and summarization tools will extract whatever is shortest and most self-contained. If your answer depends on context the intermediation layer cannot see, it will be dropped, rephrased, or replaced with something from another source. A tight self-contained answer is the form that survives extraction.
If you skip the tight form, you risk your answer being merged with another source's weaker answer on the way out, and the reader receiving a blend you would not stand behind.
- The supporting facts are a small number of verifiable claims with sources, stored alongside the answer. They are are the audit trail behind it, so that when a model, a fact-checker, or a reader asks where the answer comes from, you can point at something.
The intermediation layer is increasingly training on what it can verify. Answers that ship with their own audit trail are more likely to be trusted and less likely to be quietly rewritten. They also protect you, because when a reader or a critic asks how you know something, you have the receipts in the same place as the claim.
If you skip this, you risk a canonical layer that is fluent and confident and that you cannot defend when challenged.
- The link is a stable URL back to the full editorial version. The link matters because it is the one handle you have on the interaction: if a reader who reaches the canonical answer through an assistant wants more, the link is how they get home to your work.
Without a stable link, the intermediation layer becomes a terminal endpoint for your reporting. The reader gets the short answer and stops there, and the relationship between the reader and your publication never forms.
If you skip the stable link, you are letting the intermediation layer absorb the audience relationship you spent years building.
Some media practitioners may hear this idea of the canonical layer and worry that it is a softer version of the truth, sanded down to pass through filters, and that maintaining it will erode their editorial integrity.
The worry is legitimate and the answer is that parallel framing is not the same as dual truth.
The editorial layer carries the full picture plus a narrative thread. The canonical layer carries the core facts with the framing adjusted so that it survives the intermediation layer. The facts do not change. The institutions you name do not get unnamed. The critique in the editorial version is not erased. What changes is the rhetorical posture of the short answer, so that a model does not flinch at it and reshape it into something you would not recognize.
Journalists who work on this honestly end up writing sharper editorial pieces, because the discipline of the canonical layer forces them to separate the facts from the framing in their own heads. That separation is good editorial hygiene regardless of whether AI intermediation exists.
Adding canonical layers into your publication process may take a little bit more time and effort, but it produces four to eight canonical units that are versioned, maintained, and actually legible to the intermediation layer.
The correction loop
A publishing practice is not complete without a correction loop. You need a way to find out that the canonical layer is being surfaced incorrectly, and a way to push an update.
Without the loop, the canonical layer is a one-time intervention that will decay as models change and as the world changes around your reporting. With the loop, it is a living asset that stays current.
A minimal correction loop has three parts, and each part is worth investing in carefully.
- Monitoring
On a regular cadence, run your canonical questions through a representative set of assistants and summarization tools. Log the outputs. Flag cases where the answer you get back is materially different from the canonical answer you published. This is often dull, repetitive work. It can be partly automated.
The intermediation layer changes without warning! Reasons could be a model update, a platform policy shift, a new base model in a vendor's stack, any of these can change how your canonical answer is being surfaced. Monitoring is how you find out. If you skip monitoring, you risk running a publishing practice that was correct on the day you set it up and has been slowly misrepresenting your work ever since, with nobody on the team noticing.
- Triage
When a divergence appears, decide what to do. Sometimes the answer has drifted because the world has changed and your canonical version is outdated. Sometimes it has drifted because a model started reframing it. Sometimes it has drifted because the question was imprecise and the model latched onto a different interpretation. These three cases require three different fixes, and the point of triage is to sort them.
The wrong diagnosis leads to the wrong fix. Updating a canonical answer that was actually undermined by a model reframing will not fix anything, because the model will reframe the new answer the same way. Rewording a question that was actually just outdated will not help the reader. If you skip triage and treat every divergence the same way, you will burn time on updates that do not solve the problem you thought you were solving, and the practice will feel pointless to the team.
- Update
Push a fresh version of the canonical answer, linked to the same stable URL. Note the version and the date. Over time this produces a history of how the intermediation layer has handled your reporting, which is an asset for your own research and an asset for advocacy, because it is the most concrete form of evidence you can have about how these systems shape what readers see.
The cost is perhaps an investment in tooling, which for a small team is basically spreadsheets, a versioning system, and a handful of scripts. It is a small weekly discipline of running the correction loop. None of this is glamorous. All of it is tractable.
What you buy is an increased likelihood of durability. A story that you have written into the canonical layer and maintain with a correction loop will still be reaching readers through intermediation six months and twelve months after publication. A story that you have not done this work on will fade from the intermediation layer and its reach will quietly collapse, without you ever noticing, because nobody is watching.
If you believe that the intermediation layer is where a growing share of your audience gets information, this is the difference between writing into the current that carries your work and writing into the current that dissolves it.
The version history is documentation of how public information flows have been shaped by intermediation systems, which is information that is very hard to produce by any other method.
If you skip the versioning, you lose both the ability to improve your own practice over time and the ability to contribute to the broader conversation about how these systems behave with evidence that anyone can inspect.
Where to start
You do not have to do this for every piece of reporting. Perhaps pick your next article or, if you're ambitious, the five or ten pieces of reporting you have published in the last year that you most want people to still find when they ask an AI chatbot about the topic in twelve months: build canonical answer units for those pieces, wire up a simple correction loop, and run it for a month and see what you learn.
We've learned, first, that the canonical layer is easier to build than expected.. Second, the correction loop surfaces problems we did not know we had, and the fixes are often cheap once the problem is visible. Both of those findings are good signals.
This is what AI-legible publishing actually looks like when it is a second layer, a few disciplined fields per story, and a weekly habit. It does not require a research lab or a separate team of engineers. It requires a decision to treat the intermediation layer as part of your publication, and to maintain it.
If you have feedback or questions, don’t hesitate to get in touch at hello@gazzetta.xyz.