Claude Penland

By Claude Penland - marketing and business strategy for companies that are good at what they do and hard to find.

Agentic Buying Is Coming and Your Site Has No Door

What AI buying agents need from your website, what stops them cold, and the small machine-readable layer to build now.

Picture a buyer who reads your whole site in four seconds, has zero patience for a “Talk to Sales” button, and cannot be reached by your nurture emails. That buyer is software, and it already builds the shortlist. Here are the questions founders ask most, with numbers.

Q1. Is agentic buying real, or just a conference slide with a robot on it?

1. The research stage has already moved. G2 surveyed 1,076 B2B software buyers in March 2026 and found 51% now start research in an AI chatbot more often than in Google, up from 29% a year earlier. Chatbots are the top influence on shortlists at 54%. Vendor websites sit at 36%.

2. The shortlist is where deals get decided. 6sense reports that buyers pick one of the four vendors on their day-one shortlist 95% of the time, and they wait until 61% of the journey is over before talking to a seller.

3. The transacting stage is next. Gartner predicts: “By 2028, 90% of B2B buying will be AI agent intermediated, pushing over $15 trillion of B2B spend through AI agent exchanges.” McKinsey sizes US consumer agentic commerce at $900 billion to $1 trillion by 2030.

Table 1. The scoreboard so far

What was measuredNumberSource
Agent requests vs. human pageviews on one dev-tools site, May to July 2026268,000 vs. 107,000Evil Martians
AI traffic to US retail sites, 2025 holiday season, year over yearUp 693.4%Adobe
Conversion of AI referrals compared with all other traffic31% higherAdobe
Growth in bots fetching pages for a live user prompt, 2025More than 15xCloudflare Radar
Buyers who switched vendors because of a chatbot recommendation69%G2

Q2. What do the search patents and agent protocols say is coming?

1. Google has been patenting this since 2018. Patent US11487832B2, “Analyzing web pages to facilitate automatic navigation” (granted November 1, 2022), describes an assistant that studies existing websites so it can complete tasks on them. Patent US11775254B2 (granted October 3, 2023) covers reading a screen and filling in the input fields. Olaf Kopp tracks this family under “agentic commerce optimization” on his patent database.

2. The protocols turn those patents into plumbing. Table 2 shows what each expects you to publish.

3. A reality check. OpenAI launched Instant Checkout on September 29, 2025, then moved away from it in March 2026. Search Engine Land reported that Walmart’s in-chat purchases converted at about one-third the rate of click-throughs to walmart.com. The checkout wobbled while the product feed survived, so bet on clean data over any single checkout.

Table 2. Agent protocols and what each one wants from you

ProtocolBackers and dateWhat your site exposes
ACPOpenAI and Stripe, September 2025A product feed (id, title, price, availability) plus five checkout endpoints
UCPGoogle with Shopify, Etsy, Wayfair, Target, Walmart; January 2026A business profile at /.well-known/ucp listing services and payment handlers
MCPAnthropic; donated to the Linux Foundation December 2025Tools an assistant can call; 10,000+ public servers, 97M+ monthly SDK downloads
WebMCPW3C community group draft; Chrome origin trialNamed, typed actions registered on the page itself
Web Bot AuthCloudflare, Visa, Mastercard; October 2025Signed requests so a legitimate agent can prove who it is

Q3. What does an agent need to find on a page before it can act?

1. A complete offer. Jono Alderson puts it best: “To recommend the booking, it needs enough information to evaluate the actual offer.” That means what it is, the total cost, availability, the terms, and a way to proceed.

2. In the raw HTML. The Vercel and MERJ crawler study concluded that “none of the major AI crawlers currently render JavaScript,” on a network where GPTBot made 569 million fetches a month. Passionfruit calls Googlebot “the only major crawler with full JavaScript rendering support.” Alderson, quoted by Sitebulb: “if your initial state is just a loading skeleton thing, then the actual content is invisible.”

3. Near the top, and brief. Dan Petrovic at DEJAN measured 7,060 queries and found a grounding budget of roughly 2,000 words per query. The top source gets about 531 of them (28%). Pages under 1,000 words had about 61% of their content used; pages over 3,000 words, about 13%.

4. The same everywhere. Alderson again: “Making three contradictory prices machine-readable gives the agent a better-documented problem.” Gartner found 69% of B2B buyers already see mismatches between the website and what sellers say.

Table 3. The five questions an agent asks, and where the answers live

Agent’s questionSchema.org markupCommon failure
What is this?Product or Service: name, description, brandSlogan instead of a category
What does it cost, all in?Offer: price, priceCurrency, shippingDetails“Contact us for pricing”
Can I get it today?Offer: availabilityStock shown only in a script
What if it goes wrong?hasMerchantReturnPolicyTerms buried in a PDF
How do I proceed?url to signup, checkout, or bookingA form and a two-day wait

Q4. Why do gated PDFs and contact forms fail?

1. Gated means invisible. ZipTie is blunt: “Gated content is invisible to AI search engines.” Crawlers do not trade an email for a white paper, and the humans who do are rarely shopping: NetLine data reported by Sword and the Script shows 66% of people who register for gated content are more than a year from buying.

2. Forms stop agents at the last step. The 2026 ClawBench study ran 153 everyday tasks on 144 live websites. The best model finished 33.3% of them, and 44.4% of the tasks beat every model tested. In a sample of 26 failures, the top causes were CAPTCHA and anti-bot walls (8), quitting before the final submit (7), and login or signup screens (5).

3. Forms barely work on people. Baymard puts average cart abandonment at 70.22% across 50 studies, with 18% of abandoners leaving because the site demanded an account.

4. “Contact sales” answers a question with a calendar invite. In Microsoft Research‘s Magentic Marketplace simulation, agents chose the first proposal they received 60% to 100% of the time, a speed advantage of 10 to 30 times over quality. HockeyStack studied 80 B2B SaaS sites and found hidden pricing wins more form fills (4.6% vs. 2.8%), yet leads from transparent pricing become pipeline 1.7 times as often (17.50% vs. 10.31%).

5. An example. A buyer tells her assistant, “Find a payroll API under $500 a month with a sandbox I can test today.” Vendor A’s pricing page says “$399 per month” in plain HTML, with Offer markup and a signup link. Vendor B has a 14 MB brochure behind a form. Vendor A makes the shortlist before Vendor B’s sales rep finishes a first coffee.

Q5. Is llms.txt the fix?

1. For now, mostly no. Otterly logged 62,100+ AI bot visits over 90 days. Only 84 touched /llms.txt, about 0.1%. Their verdict: “llms.txt is almost invisible.”

2. The biggest reader wants HTML. Evil Martians saw ChatGPT-User make 196,973 requests (73% of agent traffic) and ask for Markdown just 272 times. “Keep your HTML clean, because your biggest AI reader only reads HTML.”

3. Lean pages still pay. Mike King at iPullRank measured one page at 16,180 tokens as HTML and 3,150 as Markdown, an 80% cut. Publish the file if you like, then go fix the pricing page.

Q6. What is the minimum machine-readable layer to build now?

1. Seven layers, in build order. Table 4 has the list. None of it needs permission from a platform.

2. Layers 1 through 4 are about a week of work for most startups and require no new protocol at all.

3. Wait on the fancy parts until a customer asks. chudi.dev estimates full ACP readiness (catalog, payment token, signed receipt, audit trail) at roughly 1,000 lines of code. Evil Martians advises fixing your API before shipping an MCP server. Simon Willison notes that GitHub’s MCP server alone adds 93 tools and about 55,000 tokens to a prompt.

Table 4. The minimum door: seven layers, in build order

LayerBuild thisEvidence
1. Readable HTMLServer-render price, plan limits, terms and specsAI crawlers skip JavaScript (Vercel)
2. Public pricingReal numbers, or “from $X” for enterprise“$99/month flat” beats “affordable pricing” (Evil Martians)
3. JSON-LDOrganization, Product or Service, Offer, return policy, FAQPageGoogle requires name, image, price, currency
4. One truthSame price on the page, in markup, in feeds, and from salesAlderson on contradictory prices
5. Feed or APIA product or pricing feed, public docs, an OpenAPI fileACP and UCP both begin with a catalog
6. An actual doorSelf-serve signup, checkout link, or booking URL“Let agents work first, humans claim later.” (Evil Martians)
7. An open gateConfirm robots.txt and your CDN allow ChatGPT-User and peersSome hosts block AI bots by default (Otterly)

Q7. What can go wrong, and do humans still matter?

1. Agents can be fooled. Simon Willison’s “lethal trifecta” names the dangerous mix: private data, untrusted content, and the ability to communicate externally. In the Microsoft simulation, prompt injection redirected payments.

2. Buyers still check the work. Gartner found 69% of B2B buyers prefer to validate AI-generated insights with a sales rep. Your reps keep their jobs; the agent decides which ones get the call.

3. The stakes. Jason Barnard of Kalicube: “if you are not the solution the AI autonomously selects, you will not be part of the transaction.”

Q8. What should I do Monday morning?

1. Open your pricing page with JavaScript turned off. Is there a number on it?

2. Ask two AI assistants what you cost and how to cancel. Fact-check them.

3. Search your server logs for ChatGPT-User. Metehan Yesilyurt watched GPTBot hit a new site 29,000+ times in 12 hours while his analytics showed nothing.

4. Count the steps from “interested” to “bought” that need a human. Remove one.

Agents are polite guests. They leave quietly when there is no way in. Build the door.

Sources

From the independent source list: Evil Martians (also agent rules, MCP); Jono Alderson; Olaf Kopp; DEJAN; Sitebulb; Passionfruit; ZipTie; Sword and the Script; Otterly (also bot access); iPullRank; chudi.dev (also WebMCP); Simon Willison; Kalicube; Metehan.ai.

Other sources: G2; 6sense; Gartner (predictions, 2025 survey, 2026 survey); McKinsey; Adobe; Cloudflare (Radar, payments); Google (patent 1, patent 2, UCP, markup rules); OpenAI (checkout, feed spec); Anthropic; Search Engine Land; Vercel and MERJ; ClawBench; Baymard; Microsoft Research; HockeyStack.


Discover more from 1000 Startups

Subscribe to get the latest posts sent to your email.

Claude Penland

Claude Penland builds the marketing and business strategy for companies that are good at what they do and hard to find. Thirty years operating, one exit, eight of them as a practicing casualty actuary.

The free two-page read is genuinely free. Email claude@1000startups.com and I'll send back what I can see from the outside. Or see the work samples and how to work with me.

Leave a Reply