What is agent conversion?

Agent conversion, defined

Agent conversion is the share of AI-agent sessions that end in the agent choosing and transacting with a business. An agent session starts when a person asks an assistant to find, compare or buy something and the assistant acts on their behalf: fetching pages, reading offers, comparing sellers and, increasingly, completing the purchase.

It is not the same thing as AI visibility, which counts how often a business is mentioned or cited in AI answers. Visibility is a prerequisite and a different measurement.

A business can be cited in every "best running shoes" answer and still lose every "buy me waterproof running shoes under $150, delivered by Friday" request, because the agent could not confirm its delivery time or got a 403 at the cart.

The difference matters because the sale is decided at the second kind of prompt, and almost nothing measures it. The visibility category is well served and well funded. The conversion half is where most businesses are losing without knowing.

The agent funnel

An agent buys in stages, and each stage is a gate that can be measured and that leaks, like any funnel:

1. Landed. The agent reached the page without a 403, a rate limit, a "checking your browser" challenge or a consent wall that stops it.
2. Found product. It located the item that matches the shopper's brief.
3. Confirmed offer. It read the price, delivery time and returns terms correctly, as checked against the true values.
4. Added to cart. It picked the right variant and the cart updated.
5. Reached checkout. A guest path was available, with no login wall or captcha before payment.
6. Ready to pay. It reached the payment step with the correct total.

Drop-off at stage one is the most common and the least visible, because the business never sees the session. Drop-off at stage three is the most expensive, because the agent does not retry; it moves to a competitor whose offer it could read.

What agents actually read

Most agent fetches do not run JavaScript. They read the server response: the HTML, and inside it the structured data. In practice that means:

• Product and Offer JSON-LD with a numeric price, a currency, availability and an identifier (GTIN, MPN or SKU). ChatGPT, Perplexity, Claude and Google's AI Mode all parse JSON-LD, and schema markup has been confirmed as infrastructure for AI features since 2025.
• Server-rendered name and price. A price that only appears after a script runs is invisible to the fetch.
• Policy pages in plain HTML. Delivery times, shipping costs and returns terms, linked from product pages, in text rather than images.
• Discovery documents. A UCP business profile at /.well-known/ucp, an ACP document, a product feed, and an MCP endpoint that lets the agent call the business rather than scrape it. llms.txt is cheap to publish but still largely ignored, so it carries little weight.
• Consistency. One business name across schema, page title and social metadata; ratings with counts; a verifiable identity.

Agents treat ambiguity as risk. Two different prices on one page, a promotion expressed only as a badge, or a price behind a region selector all read as "cannot verify", and an agent comparing offers picks the one it can verify.

Agent blocking, mostly by accident

Since 15 September 2026, Cloudflare, which fronts roughly a fifth of the web, blocks the Agent and Training bot categories by default on ad-bearing pages for new domains, while allowing Search bots. Cloudflare's own data shows 403 responses to agents roughly doubling since the summer, with Amazon leading the blocking and winning an injunction against Perplexity's shopping agent. The verified-identity route is Web Bot Auth, a signing scheme now at the IETF.

The default is aimed at publishers, but it reaches stores through security apps, user-agent blocklists written in 2024, and WAF rules nobody has reviewed. One store's 53% block rate turned out to be blocking Bing. The effect is the same for an agent sent by a shopper: a dead end, and a competitor one prompt away.

Meanwhile, about 132 million US adults have bought a retail product with AI's help, and only 23% of merchants can identify that traffic and the sales it produces. The traffic merchants cannot see converts at three to five times the rate of traditional search, by Cloudflare's figures. That gap, invisible and high-intent, is the business case.

Agent readiness

Agent readiness is the set of conditions a site must meet for agents to work with it: open access for the named agent identities, structured and server-rendered offers, discoverable policies, guest checkout, and the discovery documents above. It is a property of the site, measurable without any agent traffic, which is why it is the first thing to fix and the first thing we score.

Shopify and Adobe Commerce switch most of the discovery layer on for their merchants. Everyone else, and every store whose theme or security app undoes the platform's defaults, has to check.

The Agent Conversion Score

The Agent Conversion Score is ConversionVector's 0 to 100 measure of agent readiness. It runs 49 checks and reports two sub-scores, because being recommended and being bought are different jobs:

• Agent Choice: can an agent find, read, trust and recommend you? It combines Access (can named agent identities reach your pages), Legibility (is the offer in structured data and server HTML), Offer clarity (one verifiable price, with delivery and promotion terms where an agent looks) and Trust (can an agent verify who it is buying from).
• Agent Checkout: can an agent complete the purchase? Guest checkout, a cart and checkout agents can reach, no captcha before payment, and the discovery documents agent checkout uses: UCP, ACP, MCP and a product feed.

The headline works like a funnel rather than an average: Agent Choice × (0.7 + 0.3 × Agent Checkout). A closed checkout limits the score instead of being averaged away, but a site that agents can recommend still sells through the handoff to a person, so it is not zeroed. As agent-completed purchases become common, the checkout share rises in a versioned update.

Three checks cap the score when they fail, because each one stops a sale outright: agents refused on product pages (cap 40), checkout requiring login (cap 60), no machine-readable price (cap 50). The cap is the headline a business should remember, and it is always the first fix.

The full check list, weights and scoring method are published under a Creative Commons licence on the specification page, with a machine-readable version at /spec/agent-conversion-score.json. Anyone can implement it. We publish it so that "agent-ready" means one thing.

How to start

1. Get your free score. It names your blocker, which is usually one of three things and usually fixable in an afternoon.
2. Fix the cap first, then the fixes ranked by points per hour. The full report carries the exact change for each.
3. Once the score is above 70, run the Simulated Agent Buyer to see where real agents stall in your flow.
4. Track your share of recommendation on buy-intent prompts with Recommendation Monitoring, so you know whether the fixes moved the sale.

Sources

• Cloudflare's July 2026 crawler controls and the 15 September default, reported by Help Net Security.
• Agent-assisted buying and merchant visibility figures from PYMNTS Intelligence, September 2026.
• 403 responses to agents and the Amazon injunction, a16z via KuCoin, October 2026.
• The accidental Bing block, Predicta Digital, 2026.
• The protocol stack (UCP, ACP, AP2, MCP), Agentic Commerce Atlas, updated 1 October 2026.
• LLM browser agents as synthetic buyers, validated against human outcomes: Shopify's SimGym, arXiv, February 2026.
• Which engines read JSON-LD and how, SEOJuice and We The Flywheel, 2026.