Run this experiment: take your best-selling product and write down everything a customer needs to know to be confident it's the right choice. Not the spec sheet — the real list. Which setups it's compatible with. Where the materials come from. Whether it arrives before Thursday. What happens if it's returned opened. Why your best salesperson steers marathon runners to the other model.
Now check how much of that list exists as structured, machine-readable data anywhere in your systems. For most companies the honest answer is around 20%. The other 80% is tribal knowledge: it lives in marketing copy, PDFs, support macros, and the heads of veteran employees. Humans navigate that gap by asking questions and inferring. Agents don't.
Agents skip what they can't verify
This is the vagueness problem, and it fails silently. Research from delivery-platform nShift on agent-mediated shopping found that when delivery windows, shipping costs, or return terms are unclear, the agent simply skips the offer — no human ever sees that you were excluded, and nothing in your analytics records the loss. A person tolerates "ships in 2–3 business days" and infers the details. An agent comparing you against a competitor whose terms are explicit doesn't deliberate. It takes the offer it can verify.
The same mechanics apply in B2B. When a buyer tells their agent "I need something proven at 10,000 customers," that claim either exists as data the agent can check, or your product isn't in the consideration set. The agents mediating purchases are built to probe vague human intent for the reality underneath — and they can only probe what's encoded.
The highest-leverage move nobody's making
Everyone doing agent-readiness work starts with the easy layer: Product schema, price, availability. Necessary — that's the 20%. The differentiating move is encoding the 80%:
- Compatibility logic. The matrix your support team knows by heart: what works with what, and the exceptions.
- Provenance and higher-order attributes. "Same ball used in the tournament." "Sourced from a farm that funds a local school." These are what buyers actually ask their agents for, and they're all expressible as data.
- Edge-case policies. Returns on opened items, delivery to that awkward region, what "business days" means. The FAQ answers, as fields.
- Decision trees. How your best rep converts "what's the right option for me?" into a specific answer. If it's structured, an agent runs it automatically. If it's in someone's head, the transaction goes to whoever structured it first.
The forcing function
Here's the uncomfortable, useful part: you can't fake this with markup. Encoding tribal knowledge forces you to reconcile the catalog, the fulfillment system, and the FAQ that have quietly disagreed for years — disagreements your human-facing website papered over. Agents take data literally, surface contradictions, or skip you. Agent-readability turns out to be the most effective data-quality forcing function most organizations will ever encounter, precisely because it's non-optional.
And the payoff isn't only agents. Once the knowledge is structured, better human experiences fall out of it almost for free.
There's a token-economics footnote worth knowing too: when Cloudflare shipped automatic markdown conversion for agent requests, its example page dropped from 16,180 tokens as HTML to 3,150 as markdown — an ~80% reduction. Every wasted token is patience your content spends before the agent gets to the substance. Serve the clean version — an llms.txt is the natural front door — and put the encoded knowledge behind it in schema that earns citations.
Start by finding out what agents can currently see of you — run a free Legible report. The gap between your score and what your best salesperson knows is your tribal knowledge backlog, in priority order.