The Law of Least Token Resistance

There's a question underneath every "how do I show up in ChatGPT" conversation that almost nobody states plainly: why would an AI agent pick one store over another? Not in the SEO sense of ranking signals — in the literal sense of what an agent does, token by token, when it tries to help a user buy something.

The answer is an incentive, and it's worth naming because it predicts almost everything else.

Call it the Law of Least Token Resistance: an AI agent completes a user's request at the highest success rate and the lowest token-and-time cost it can manage, so it gravitates toward the path of least resistance. A store it can read and act on in roughly one pass is that path. A store it has to scrape, guess at, and retry is not.

This isn't a favor an agent does for "agent-friendly" businesses. It's arithmetic — and the arithmetic runs the same way every time.

---

The 30-second version

	Store without agent files	Store with the kit
What the agent does	Fetches raw HTML, infers structure, probes paths, guesses form fields, retries failures	Reads agents.json → typed actions with exact locations and parameters
Token + time cost	High — thousands of tokens spent scraping and recovering from dead ends	Low — one structured read resolves the intent
Success rate	Low — stalls at the cart, the booking modal, or a mis-mapped form	High — the flow completes
The agent's revealed preference	Give up, or fall back to a competitor it can finish on	Use it, finish the task, surface it in the answer

The two columns aren't a quality judgment about the store. They're a cost judgment about the path. The agent doesn't know or care that your products are better. It knows one path costs 4,000 tokens and ends in a guess, and another costs 400 and ends in a typed action.

---

Why would an AI agent prefer one store over another?

Because agents are optimized to finish the user's task at the highest success rate and lowest cost — and an agent-readable store is dramatically cheaper to resolve than one the agent has to scrape and infer.

Every production agent — ChatGPT's browsing and Operator modes, Claude's computer use, Perplexity's shopping flows, Gemini's commerce surfaces — runs against the same three pressures: a finite context budget, a latency ceiling the user will tolerate, and a success metric it's graded on. Those pressures don't reward the agent for being thorough on a hostile site. They reward it for being done.

So when an agent has a choice — and on any "best X for Y" query it has many — the store that resolves in one structured read is simply the cheaper bet. Not the more deserving store. The cheaper one.

What does "expensive" actually mean for an agent?

On a site with no agent-readiness layer, completing a transaction looks like this: fetch the HTML, parse a DOM that was built for human eyes, infer which <div> is the add-to-cart control, guess the field names on the checkout form, submit, hit a validation error or a JavaScript-gated step, back up, and try again — burning tokens and wall-clock time on every loop, with a real chance of failing at the end anyway.

agents.json collapses that. It's the acting layer: a manifest of the transactional intents a site actually exposes — checkout, booking, signup, search, contact — each with a location, typed parameters, and a success_indicator. The agent reads what it can do and exactly how, instead of reverse-engineering it from markup. The guessing, the retries, and most of the failure modes disappear in a single read.

That's the whole mechanism. Cheaper to read means cheaper to act, which means more likely to be the path the agent finishes on.

Does being cheaper to resolve actually make an AI recommend you?

Not as a guarantee — as a tendency. When an agent assembles a shortlist or completes a purchase, the stores it can act on are the ones that survive to the answer; the ones it can't resolve get quietly dropped before reputation ever enters the picture.

It's worth being honest about the limit here, because the overclaim is everywhere and it's wrong. Shipping three files does not vault a small store above a famous one on a query where reputation dominates. An agent asked for "the best-known running shoe" will still name the household brands.

But most agent journeys aren't that query. They're "find me an everyday olive oil and add it to my cart," "book a table for four on Friday," "sign me up for the trial." On those, the agent is assembling a working set of stores it can actually transact on — and an unreadable store isn't ranked low, it's excluded, because including it costs more than it's worth. You can watch this happen: our recommendation check asks an AI which stores it would recommend in a category and reports whether a given store made the list. The stores that get named are, with boring consistency, the ones an agent can resolve.

Unreadability doesn't lose you the comparison. It removes you from the comparison.

Where's the evidence?

Two places, one historical and one you can run yourself.

The historical signal: across 11 sites that shipped the full readiness stack between February and April 2026, the internal agent-readiness score averaged a 17 → 72 jump, with the citation-lift signal correlated to score. Sites that shipped llms.txt alone in the same window averaged 17 → 28 — the partial-deployment result the 2026 skeptics correctly measured. Same files, different completeness, different outcome.

The live signal: send an AI agent at a store that hasn't shipped the files and watch where it stalls — it can read the catalogue, then hits the checkout and can't proceed, because nothing tells it how. The agent's-eye view narrates exactly that, first person, on a real store. Run it against a fully-equipped site and the same five-step flow — understand, find, cart, policy, checkout — closes instead of breaking. The difference isn't the products. It's the resistance.

What do I actually ship to lower my token cost?

The five-layer stack — the same one that turns the skeptics' flat line into measurable lift:

Layer	The file	What it removes for the agent
Reading	llms.txt	"Which of your URLs is worth my limited budget?"
Acting	agents.json	"What can I do here, and how — without guessing?"
Behaving	agent-instructions.md	"What's allowed, what's the brand voice, what's the canonical answer?"
Discovery	<link rel="alternate"> tags	"Where is each of the above? Don't make me probe paths."
Structure	Schema.org JSON-LD per page	"What is this specific page, in machine-friendly form?"

Every layer you ship is a chunk of guesswork you delete from the agent's path — which is to say, a reason for it to finish on you instead of bailing to a competitor. This rides the open standards the whole agent ecosystem is converging on; it is not a proprietary tag you have to bet on. (If you're untangling how agents.json sits beside MCP, A2A, and Google's UCP, the standards map lays out which one actually matters for a public site.)

---

The law, restated

Agents don't reward virtue. They reward the path of least token resistance. Visibility, citations, recommendations — these aren't things you win by being good; they're things that fall out of being cheap to resolve and reliable to finish on. The store that an agent can read in one pass and act on without guessing is the store that survives to the answer.

Make yourself the path of least resistance, and being recommended stops being something you chase. It becomes a side effect of the math.

---

Related

• agents.json vs WebMCP vs llms.txt — the three-layer mental model — the reading/acting/behaving stack this post is the incentive story behind
• llms.txt is a dud, the skeptics say — they're half right — why one layer alone collapses to a flat line, and the 17→72 vs 17→28 data
• UCP, agents.json, MCP — three protocols, three jobs — where the commerce-transaction layer is heading and what to ship today
• MCP vs A2A vs agents.json vs WebMCP — the standards map for site owners — which standard actually matters for a public website
• Run the free readiness audit → — score your store's resistance in five seconds