Agent-Ready Website Checklist for 2026 (llms.txt, Schema, AIPREF, UCP, NLWeb)

A site can pass a crawler check and still be useless to an agent that needs to buy, book or quote. Robots.txt says the bot may enter. It does not say the agent can complete a purchase, hold a reservation or trust your AI-training preferences.

That gap is what "agent-ready" means in 2026. Crawl access is the floor. Actionability is the next layer: clear product and service facts, booking or checkout signals agents can parse, and preference files that say what you allow.

This checklist walks through the signals that matter, marks which ones are standards versus informal conventions, and shows where Beket AI's free tools fit. It does not invent a certification score.

Two layers: readable vs actionable

Layer	Question	Examples
Readable	Can AI systems fetch and extract facts from your pages?	robots.txt for search/answer bots, HTML that includes prices and services, basic structured data
Actionable	Can an agent discover how to buy, book or respect your preferences?	Product/Offer schema, ReserveAction, UCP profile, NLWeb/ask endpoints, AIPREF tokens

Most local businesses still stop at layer one. Ecommerce teams racing into Google AI Mode checkout need layer two as well. Our longer breakdown of llms.txt vs robots.txt vs schema markup covers the readable layer in depth. This page is the action checklist.

The 2026 agent-ready checklist

Use this as a working sheet. Status notes reflect primary documentation as of November 8, 2026. Re-check the linked sources before you brief engineering.

#	Item	What "done" looks like	Maturity	Notes
1	AI search crawlers allowed when you want to be found	robots.txt allows the search/answer bots you care about; training bots are a separate decision	Established	Do not confuse training opt-outs with disappearing from answers
2	Important facts in initial HTML	Name, services, prices, locations and hours appear in view-source, not only after JavaScript	Established	Answer-time fetchers often read HTML only
3	Product or LocalBusiness structured data	Valid Product/Offer or LocalBusiness JSON-LD on the right templates	Established	See Google's structured data docs and our LocalBusiness schema guide
4	Booking Action schema (service businesses)	potentialAction / ReserveAction (or related Action types) point at a real booking URL or API	Vocabulary exists; agent uptake varies	Schema advertises capability; it does not guarantee an AI Mode booking handoff
5	llms.txt	/llms.txt published if you want one	Informal convention	llmstxt.org describes the proposal; it is not an IETF or W3C standard. Google Search has said it ignores llms.txt for Search
6	AIPREF AI preference signals	robots.txt / HTTP preference tokens once you adopt the IETF vocabulary	IETF working group	AI Preferences (aipref) is active; treat drafts as evolving
7	UCP / commerce endpoints (merchants)	Interest + technical onboarding; /.well-known/ucp or documented transports; native_commerce where Google requires it for Buy	Live for select merchants in US/CA/AU	See Google's UCP guide and Merchant Center Help
8	NLWeb or MCP conversational endpoint (optional)	Public ask / MCP endpoint if you choose to expose one	Open project / early platform adoption	Microsoft's NLWeb introduction describes sites as natural-language endpoints; every NLWeb instance is also described as an MCP server
9	Edge policy for agents (if you use Cloudflare)	Personal agents / signed agents classified and allowed or blocked on purpose	Product feature on Cloudflare	Separate shopping agents from training crawlers; see Cloudflare's signed-agent and personal-agent materials
10	Measurement split	You track "named in answers" separately from "Buy/book available"	Process	Naming ≠ checkout

Print the table, or copy it into your project tracker. Skip rows that do not apply (a law firm does not need UCP; a pure content publisher may skip ReserveAction).

1. Fix crawl and HTML facts first

If OAI-SearchBot or other answer-path bots are blocked, or your prices only exist after a client-side render, later protocol work will not help. Run a raw HTML check on your homepage, a money page and a contact or booking page: view-source and search for a real sentence, a price and a phone number.

Beket AI's free 9-point /ai-check is built for this technical pass. It looks at crawl and readiness signals such as robots.txt, llms.txt and structured data. It does not, today, verify UCP profiles, NLWeb endpoints, AIPREF tokens or ReserveAction booking flows. Do not treat a green /ai-check as proof you are buyable or bookable inside an AI assistant.

For a wider menu of free checks (answers vs crawlers), use our guide to free tools that show how AI sees your website.

2. Add the schema agents can act on

Product and local facts

Ecommerce pages need Product/Offer facts that match the visible page. Local service pages need consistent LocalBusiness fields. Our LocalBusiness schema markup guide walks through Google's local structured data without padding the page with invented Q&A blocks.

ReserveAction for appointments

schema.org/ReserveAction is a way to declare that a reservation or booking action exists and where it goes. Typical pattern: a potentialAction of type ReserveAction with a target URL your booking system actually serves.

Be precise in what you claim:

• Do keep the target URL working, HTTPS and consistent with the button humans click.
• Do not claim that publishing ReserveAction forces ChatGPT or Google AI Mode to complete the booking for you. Schema is discovery metadata. Google's AI Mode home-services booking links, where they appear, have also used partner platforms; that is a separate product path from your own ReserveAction markup.

3. Treat llms.txt as optional, not magic

The /llms.txt proposal is a curated markdown map of your site for language models. Shipping one can help tools that choose to read it. It is still an informal convention.

If you publish /llms.txt:

• Keep it short, accurate and updated when URLs change.
• Do not remove robots.txt controls because you added llms.txt.
• Do not expect Google Search to honor it as a ranking or citation lever; Google's own AI optimization guidance has treated llms.txt as something Search does not use.

Our comparison of llms.txt, robots.txt and schema is the place for the longer myth-busting version.

4. Watch IETF AIPREF for preference signals

The IETF AI Preferences (aipref) working group is standardizing vocabulary and attachment methods (including robots.txt and HTTP) for AI-related preferences. Major AI companies have also described shared training opt-out work in this direction.

For site owners:

• Prefer documented preference tokens over one-off invented headers.
• Separate training preferences from search/answer access. Blocking training is not the same as leaving ChatGPT search answers.
• Re-read the working group documents before you tell a client the tokens are "final law." They are standards-track work in progress.

5. Ecommerce: UCP when you want in-answer checkout

If you sell products and care about Buy inside Google AI Mode or Gemini, add a commerce track:

1. Read Google's UCP developer guide and Merchant Center Help on UCP checkout.
2. Confirm your market is in the current eligibility set (Help lists the United States, Canada and Australia for participating merchants and partners).
3. Expect select-merchant / early-access onboarding today; Google Canada's Nov 4, 2026 announcement said simplified Merchant Center integration begins next year.
4. Plan for the native_commerce(checkout_eligibility) attribute on listings that should show Buy.

For a Canada-specific snapshot of naming versus checkout partners, see our AI Mode checkout Canada retailer study. Keep discovery checks on the homepage audit while UCP onboarding runs in Merchant Center.

6. Optional: NLWeb or MCP so agents can ask your site

Microsoft's NLWeb project aims to turn a website into a natural-language interface over your own data, using formats many sites already publish (such as Schema.org and RSS). Microsoft describes each NLWeb instance as an MCP server so agents can query it when you allow that.

This row is optional. Most small businesses will get more from HTML facts, schema and (if retail) UCP than from standing up a conversational endpoint on day one. If a platform you already use (for example a major website builder) offers a one-click NLWeb path, evaluate it as an add-on after the readable layer is clean.

7. If you use Cloudflare, decide agent policy on purpose

Training crawlers, search indexers and personal shopping agents are not the same traffic. Cloudflare has been adding controls and analytics that distinguish personal agents and cryptographically signed agents from older bot categories. Allow or block them deliberately on checkout and booking paths instead of applying a single "block AI" rule to everything.

How Beket AI fits (and what it does not do)

Need	Beket AI option	What it covers
Do ChatGPT, Perplexity and Google AI Overview name us?	Free 25-prompt audit on beket.ai	Mentions and citations across those three engines
Are basic crawl/readiness signals OK?	Free 9-point /ai-check	Technical readiness such as robots.txt, llms.txt and structured data signals
Ongoing monitoring	Beket AI Pro ($49/mo)	Product plan for continued visibility work
Integrations today	Webflow, Google Analytics, Cloudflare	Confirmed product integrations

Beket AI does not apply on-page fixes for you from these free checks, does not enroll merchants in UCP, and does not claim /ai-check validates NLWeb, AIPREF or ReserveAction. Use the checklist above with your developer or agency for those rows.

A one-afternoon sequence

1. Run /ai-check and fix blocked search bots or empty HTML shells.
2. Confirm Product or LocalBusiness JSON-LD on key templates.
3. Service businesses: add a honest ReserveAction (or related Action) that hits a live booking URL.
4. Merchants in eligible markets: open Google's UCP Help article and start eligibility, separately from SEO work.
5. Run the homepage audit so you know whether engines name you while the protocol work is in flight.
6. Revisit llms.txt and AIPREF only after the facts and actions above are solid.

Agent-ready is a ladder, not a badge. Get readable first. Then make the next action an agent can take as obvious as the next action a human can take.