Not too long ago, I discovered myself in one other dialog that has grow to be more and more acquainted. A senior govt had obtained a warning from an AI visibility evaluation vendor that the corporate was not sufficiently ready for AI search. Among the many suggestions was one thing I’ve seen showing extra ceaselessly: the corporate wanted an llms.txt file.
Abruptly, an rising, and still-debated publishing format had grow to be an govt concern. Somebody now wanted to find out whether or not the advice was legitimate, assess the potential affect, clarify why the corporate had not already carried out it, and determine whether or not advertising and engineering assets needs to be redirected to deal with it.
I not too long ago wrote about this phenomenon as a part of what I name the AI FUD Tax. The price of the person advice could also be comparatively small, however the organizational price of repeatedly responding to the most recent exterior AI audit can grow to be substantial. Each new audit, vendor pitch, protocol, acronym, or aggressive declare creates one other spherical of questions on whether or not the group is falling behind.
The issue just isn’t llms.txt, MCP, markdown, or any specific expertise. In actual fact, they do various things. Some assist machines uncover data, some present alternative routes to signify it, and others outline how techniques trade or entry it. Lumping them collectively, technically, can be inaccurate, however strategically, they create the identical organizational temptation: treating the most recent supply mechanism as the answer reasonably than analyzing the underlying data. The error is treating each as if it represents a brand new technique.
After watching variations of this cycle for many years, I imagine we’re as soon as once more focusing an excessive amount of consideration on the format and never sufficient on the data these codecs are supposed to speak. As an alternative of instantly asking, “Do we have to implement this?”, organizations ought to first ask a extra elementary query:
“Do we have now the data required to help it?”
That distinction turns into more and more necessary as AI creates extra methods for machines to devour organizational data. If the underlying data is incomplete, fragmented, inconsistent, or trapped inside particular person departments, including one other machine-readable format doesn’t resolve the issue. It merely creates one other place to publish the identical limitations.
The organizations greatest positioned to adapt is not going to essentially be people who implement each new protocol first. They are going to be people who arrange and govern their data nicely sufficient that supporting the following helpful format turns into a publishing choice reasonably than one other reconstruction venture.
Choice Protection Creates The Subsequent Query
Within the earlier article on this sequence, I launched Choice Protection as a technique to measure how fully a corporation has uncovered the proof AI wants to guage, evaluate, qualify, and confidently suggest its services or products.
The thought emerged from an issue I imagine many organizations are starting to come across. They could have an infinite quantity of product data but lack the proof AI must help an precise buyer choice. Specs can describe what a product is, however they don’t essentially clarify who it’s applicable for, when it needs to be really helpful, the way it compares with alternate options, or which trade-offs matter to totally different clients.
This turns into particularly obvious after we deconstruct complicated prompts. A request for the “greatest” product isn’t a single query. It’s usually a mix of said necessities, inferred preferences, constraints, comparisons, and eligibility standards that collectively decide which choices make the reduce.
Think about a buyer asking for one of the best family-friendly beachfront resort in Cancun. “Finest” just isn’t an attribute a resort can merely add to a web page. The advice could depend upon beachfront entry, household suitability, room configuration, facilities, value, availability, critiques, and different standards inferred from the request. The AI should consider these circumstances collectively earlier than deciding which properties qualify for consideration.
Choice Protection approaches the identical downside from the group’s facet. As soon as we perceive the variables influencing the choice, we will decide whether or not the group has authoritative proof to help them. If a essential criterion can’t be substantiated, the issue will not be that the model ranked poorly. It could by no means have offered sufficient proof to make the reduce.
This provides us a extra defensible technique to diagnose a scarcity of AI visibility. Quite than observing {that a} competitor was really helpful and instantly responding with extra “me too parity content material,” extra hyperlinks, or one other complicated technical implementation, we must always deconstruct the choice, determine the standards influencing qualification, and decide the place our supporting proof is incomplete.
In different phrases, we will start shifting from merely observing what occurred towards growing proof for why it occurred. As soon as we all know what proof ought to exist, nonetheless, Choice Protection raises one other query: The place ought to that data reside, and the way can we guarantee each system receives the identical full reply?
That’s the place the obsession with these new particular person codecs begins to create issues.
A New Format Can’t Repair Lacking Data
The simplest response to a Choice Protection hole is to place the lacking data into no matter format is receiving consideration for the time being. Each day LinkedIn is flooded with suggestions to increase the schema, create a markdown model, or construct an MCP endpoint, and sure, deploy llms.txt.
That will resolve an instantaneous publishing downside, but it surely doesn’t essentially resolve the underlying data downside. If a buyer choice relies upon upon 5 significant standards and the group can substantiate solely 4, publishing those self same 4 items of proof by way of one other protocol doesn’t immediately set up the fifth. We’ve got merely made the identical proof hole out there in one other format.
This distinction sounds apparent, but a lot of immediately’s dialogue of AI implementation reverses the order, specializing in citations reasonably than the sources or strategies of integration. An expanded schema can describe relationships between entities, but it surely can not decide what these relationships needs to be. MCP could make a number of organizational assets accessible to AI techniques, but it surely can not decide whether or not these assets comprise the data required to reply a buyer’s query. An llms.txt file can level machines towards data, but it surely can not compensate for data the group by no means created.
All of those codecs are mechanisms for speaking and transporting data. Their worth in the end is dependent upon the completeness and high quality of what organizations put into them. I’ve seen over 100 agentic readiness audits all flagging the presence of an llms.txt file, however none of them critique the depth or high quality of the file for people who had them.
That is additionally why the present dialogue round knowledge integrity is so necessary. In his latest Search Engine Journal article, Alex Moss argued that technical website positioning more and more must deal with maximizing knowledge integrity as AI techniques depend upon correct entities, specific relationships, machine-readable codecs, actions, and dependable notion alerts. He additionally made an necessary level that aligns carefully with my argument right here: Quite than making an attempt to foretell which rising protocol will in the end win, organizations ought to strengthen the underlying layers on which these protocols rely.
Information integrity, nonetheless, additionally creates a previous organizational query. Earlier than we will be sure that data stays correct, synchronized, and reliable, we have to decide what data ought to exist, how these items relate to 1 one other, who owns them, and which supply needs to be thought-about authoritative.
Information integrity helps be sure that data stays reliable as soon as it exists. Data structure helps be sure that the precise data exists, is related appropriately, and could be ruled as an organizational asset. Each grow to be more and more necessary as AI techniques depend on our data to make selections reasonably than merely return paperwork.
Construct The Base As soon as, Publish In all places
Throughout a latest webinar, I attempted to simplify this more and more complicated AI panorama right into a single precept: Construct the canonical base as soon as. Publish in every single place.
The thought is deliberately easy as a result of organizations don’t want one other acronym or protocol to handle. They should cease rebuilding the identical data for each new vacation spot.
On the basis are the details, relationships, insurance policies, experience, buyer choice standards, and supporting proof the group can authoritatively present. These parts should be captured in a canonical data supply the place they are often related, ruled, up to date, and reused independently of any specific publishing format. AI-focused organizations like Milestone have constructed this natively into their content material administration system because the canonical supply of reality, able to outputting to any present and rising markup codecs.
Choice Protection helps us decide whether or not that data is sufficiently full to help the client selections that matter. Data structure offers a technique to arrange and keep data. The assorted codecs then grow to be supply mechanisms for making the suitable parts of that data out there to the techniques that want them.
Right now, these supply mechanisms could embrace net content material, schema markup, Service provider Heart feeds, APIs, markdown, MCP, llms.txt, and different rising codecs. Tomorrow, the listing will virtually actually be totally different.
That ought to not require rebuilding the underlying data.
If each publishing format maintains its personal model of the group’s details and choice proof, each replace creates synchronization issues and each new protocol creates one other implementation venture. If these codecs draw from a standard, ruled data supply, the structure modifications dramatically. The tough work occurs as soon as on the data layer, whereas the publication layer adapts as applied sciences and necessities change.
That is what I confer with as Data Structure. It’s the organizational functionality that permits trusted data to be constantly assembled, ruled, and delivered wherever it creates worth.
Publication Is Not Functionality
This distinction additionally explains a few of my frustration with the more and more free use of phrases reminiscent of “AI prepared” and “agentic prepared.”
Supporting an agent-oriented protocol can actually make it simpler for an agent to work together with a corporation’s techniques. That’s helpful infrastructure, but it surely doesn’t routinely imply the group possesses the data required for an agent to make a helpful choice.
There are actually two totally different issues being conflated. The primary is organizational functionality: Can the corporate seize, join, govern, keep, and retrieve the data clients and machines want? The second is publication: Can that data be expressed by way of the format required by a selected search engine, AI platform, agent, software, or protocol?
Certainly one of my arguments about these markup codecs is that they arrive from a vendor’s resolution that guarantees a clearly outlined deliverable. They’re fixing an actual downside for AI techniques: the chaos of a contemporary marketing-focused web site, and that makes it simple to promote. A few of their instruments expound on how rapidly and simply they will spin out the format, with little to no point out of how they will wrangle organizational data, since no protocol can create it for us.
A protocol can not reconcile conflicting product data owned by totally different departments. It can not extract the experience residing inside a salesman’s head, decide which buyer objections matter, set up how a coverage impacts a selected product, or create the lacking proof recognized by way of Choice Protection. These are organizational data issues that should be resolved earlier than expertise can distribute the solutions.
Know-how can expose organizational functionality, but it surely can not substitute for it.
This is the reason organizations don’t grow to be AI-ready just by implementing extra protocols. They grow to be AI-ready by organizing their data nicely sufficient that every helpful protocol turns into one other publishing vacation spot reasonably than one other try and reconstruct what the group is aware of.
Manage Data Round Selections
For many of the net’s historical past, the main focus was on creating and managing collections of pages; CMS platforms bolstered that construction, and website positioning naturally optimized product pages, class pages, articles, FAQs, and touchdown pages as a result of pages have been the first items by way of which search engines like google retrieved data and clients consumed it.
AI is weakening that relationship.
A single buyer query could now trigger an AI system to retrieve data from a number of pages, product feeds, structured knowledge, critiques, exterior sources, databases, and different data repositories earlier than synthesizing a response. The focused webpage stays necessary, however it’s now not essentially the unit round which the choice is constructed.
The extra sturdy organizing precept is the client choice.
What does the client have to know? Which standards decide whether or not a product qualifies? What proof helps these standards? Which alternate options should be in contrast? What trade-offs needs to be understood? Which insurance policies, areas, availability constraints, or different relationships have an effect on the result?
These questions join instantly again to the Choice Protection framework. If we perceive the circumstances that collectively decide whether or not a product, service, or group makes the reduce, we will determine the proof required to help every situation. The canonical data supply then offers a ruled residence for that proof, whereas particular person codecs decide how it’s delivered.
This is the reason chasing output codecs will get the sequence backward. We should always not start with an empty protocol and ask what data we will put into it. We should always start with the client selections we have to help, make sure the proof required for these selections exists, after which decide which supply mechanisms make that data accessible to the techniques influencing the result.
This Is Infrastructure, Not One other AI Tactic
This argument ought to sound acquainted to anybody who has adopted my earlier Search Engine Journal articles as a result of it extends a place I’ve held for a while. In “website positioning Is Not a Tactic. It’s Infrastructure for Progress,” I argued that sustainable search efficiency is dependent upon capabilities embedded throughout the group reasonably than remoted optimization tasks. Search works greatest when product, content material, expertise, and enterprise technique are related round how clients truly uncover, consider, and select options.
I later launched the Search Fairness Hole to quantify the enterprise worth misplaced when organizations fail to seize the certified visibility they need to moderately earn. That article additionally highlighted the rising affect of zero-click experiences and AI-mediated search, by which visibility can stay even because the economics of buyer interactions change.
These concepts have grow to be more and more related as AI has advanced. Model Sovereignty establishes the necessity for organizations to grow to be the authoritative supply of their very own details and experience. Click on Worthiness helps decide the place continued engagement yields ample incremental buyer and enterprise worth to justify funding. Choice Protection asks whether or not we have now uncovered the proof AI wants to guage, evaluate, qualify, and confidently suggest us.
Data structure offers the sturdy basis beneath these capabilities.
With out reliable organizational data, model sovereignty turns into tough to determine. With out understanding buyer selections, it turns into tough to prioritize click on worthiness. With out full choice proof, Choice Protection stays incomplete. With out an structure able to governing and reusing that data, each new AI format sends the group again into one other implementation cycle.
This is the reason the following AI protocol is not going to save an website positioning technique that lacks the underlying data functionality. The protocol is downstream of the issue.
The Subsequent Protocol Gained’t Be The Final
I have no idea whether or not MCP, llms.txt, Brokers.md, UCP, or any of immediately’s different rising approaches will grow to be foundational requirements. Some undoubtedly will grow to be extra necessary. Others will evolve, merge into one thing else, or disappear as rapidly as earlier applied sciences that after appeared important. That uncertainty is exactly why organizations ought to resist constructing their AI technique round particular person codecs. The following protocol is not going to be the final one.
Organizations that construct their technique round immediately’s hyped codecs will ultimately have to rebuild round tomorrow’s. Organizations that construct a ruled data supply round their enterprise experience and buyer selections can be in a really totally different place. When the following helpful format seems, they won’t have to rediscover, recreate, and reconcile the data it requires. They might want to decide whether or not the format offers a beneficial new technique to publish what they already know.
That may be a way more resilient mannequin for AI readiness and, extra importantly, a a lot better use of organizational assets.
The purpose is to not help each doable format. It’s to make sure that the data required to make necessary buyer selections is full, authoritative, ruled, and out there for supply by way of the codecs that matter. The protocols will proceed to alter. The supply mechanisms will proceed to multiply, and consultants will undoubtedly proceed discovering new objects so as to add to AI readiness audits. The strategic response ought to stay remarkably steady.
Construct the bottom as soon as. Publish in every single place.
Extra Sources:
Featured Picture: Rawpixel.com/Shutterstock
