Actum
Agent commerce

How AI agents decide which vendor to recommend — and why most sites never get shortlisted

A growing share of buying — B2B and B2C — starts inside an AI assistant. Here's how the agent actually chooses, and the two stages where sellers lose the deal before a human ever shows up.

12 Jun 2026 · 4 min read

A growing share of buying now starts inside an AI assistant. The buyer asks ChatGPT or Perplexity to find and compare options, and an agent does the legwork — arriving at your site already half-decided, or never arriving at all. The uncomfortable truth is that you're being filtered out before a human ever sees you, and you can't see it happening.

So how does an agent actually choose? Cutting through the noise, four factors decide it — and the funnel narrows fast.

Discover Read what you do Assess fit Transact most B2B sites never reach here
Discovery gets you in the door. The deal is won — or silently lost — at the last two stages.

1. Can it read what you do, fast. Agents don't browse; they parse. If your offering, who it's for, and your pricing aren't in machine-readable text — structured data, a clear /llms.txt, plain prose rather than JS-rendered visuals — you're illegible. And illegible means skipped.

2. Can it assess fit. This is where it dies quietest. If pricing sits behind a “contact sales” wall, or your specs and terms aren't stated, the agent can't judge budget or fit — so it recommends a competitor who made it easy.

3. Citation history. Increasingly, agents weight sources that have been consistently referenced over time over a freshly-optimised page. Authority compounds — which is exactly why publishing genuinely useful content now, rather than later, matters.

4. Can it finish the job. Discoverability gets you shortlisted; completability wins the deal. If the next step — buy, book a demo, start a trial — needs a human, the agent abandons at the last step.

Most readiness advice stops at #1. But being found is table stakes; the deals are lost at #2 and #4 — the stages you've never been able to measure. That's the gap Actum closes: we send an autonomous AI buyer through your live site, show you the exact stage it drops out, and keep you transactable as the models and standards keep moving.

The bar isn't static. ACP, MCP and AP2 are all under a year old; every model release shifts what “agent-ready” means. Treat it as something you stay, not something you finish.
Sources: OpenAI, McKinsey, Adobe, Mirakl (2026).

Can AI agents actually buy from you?

Free scan in ~30 seconds. See exactly where an AI buyer drops out.

Scan your site free →