An executive receives a vendor report that the company “isn’t ready for AI search,” and one of the recommendations reads: deploy llms.txt. It sounds like a technical checklist item—easy to scope, clear to assign—and suddenly scarce engineering time is diverted to a new format.
That reaction is understandable, but strategically risky. Formats such as llms.txt, MCP endpoints or expanded schema expose information to machines; they don’t create the authoritative facts, relationships or decision logic those machines need. Before you add another delivery mechanism, ask a more consequential question: do you actually have the knowledge AI must evaluate?
Decision Coverage: diagnosing visibility gaps
The right starting point is a diagnostic lens Bill Hunt calls Decision Coverage. A customer’s request—“best,” “cheapest,” “family-friendly”—is rarely a single attribute. It usually compresses eligibility rules, constraints, inferred preferences and trade-offs that an AI must evaluate together to reach a confident recommendation.
Take a prompt for the “best family-friendly beachfront resort in Cancun.” That query implies multiple signals: beachfront access, family-oriented room configurations, child-focused amenities, pricing and availability windows, guest reviews, and how the property stacks up against alternatives. If your content substantiates only a subset of those criteria, AI will likely rule you out—regardless of how many protocols you support.
Viewed this way, a ranking problem can be a missing-evidence problem, not a technical indexing failure. Map the decision variables that matter for your product or category and verify whether you hold authoritative evidence for each.
Why new formats won’t close evidence gaps
Formats do different technical work—llms.txt can point machines to resources, expanded schema describes relationships, MCP can make resources accessible—but none can invent facts or resolve gaps in organizational knowledge. Publishing the same incomplete evidence across more channels simply creates more places where the same deficiency appears.
In audits of agentic-readiness, Hunt frequently finds the presence of recommended files flagged without a critical evaluation of the depth and quality of the information they reference. Complementary thinking on data integrity matters here: technical SEO increasingly depends on accurate entities, explicit relationships and trustworthy perception signals, but data integrity presumes you already know which facts should exist, who owns them and which source is authoritative.
Build the canonical base once; publish everywhere
The practical organizing principle is straightforward: build a governed, canonical knowledge layer and publish from it. Capture facts, relationships, decision criteria, policies and supporting evidence in a single machine-readable source of truth. Then treat outputs—web pages, schema, feeds, APIs, markdown or emerging formats—as publishing decisions rather than reconstruction projects.
When multiple formats each maintain their own versions of facts, synchronization and governance collapse into ongoing maintenance costs. When those formats draw from a single canonical layer, the expensive work—defining what knowledge is authoritative, who owns it, and how it relates to customer decisions—happens once.
Start with three operational priorities:
- Map decision use cases. List the customer decisions and recommendation scenarios that generate value, and document the criteria AI must evaluate to qualify options.
- Inventory and assign ownership. For each criterion, locate existing evidence, assess completeness and assign a single authoritative owner responsible for maintenance.
- Design a publish layer. Consolidate authoritative data into a governed store that can export to the formats you need, with processes to ensure integrity and synchronization.
Protocols like llms.txt and MCP still have a role. The difference is sequencing: adopt formats after you confirm you have the right knowledge to expose. Formats then become efficient distribution channels, not band-aid fixes.
For teams with limited engineering bandwidth, this approach reduces churn. Instead of repeatedly responding to vendor audits with format-specific patches, you front-load knowledge architecture so future protocol changes are an operational task, not a strategic crisis.
What to watch next: as AI systems shift from document retrieval to decision-making, demand for verified, contextual evidence will grow. The immediate action is not to chase every emerging protocol; it’s to map the decisions your customers actually make, close evidence gaps, and make exposing that evidence a reproducible publishing step. That turns the next format request from an emergency into a routine configuration change.
