Vlad Ilchenko

Product builder, Ukrainian, Rock'n'rolla.

[Vlad Ilchenko @Advalor]

@patrickc This is the gap I keep hitting on the commerce side: agents show up at the edge before the policy layer does. I measured published rules vs actual access on 83 US retailers — most sites had neither.

[Vlad Ilchenko @Advalor]

@Cloudflare Borrowed logins are how a lot of agent pilots start, and the audit trail falls apart the next morning. Separate identity for the agent feels like the minimum before you let it touch customer data.

[Vlad Ilchenko @Advalor]

@jeff_weinstein This is painfully true. I only trusted the agent-access numbers after watching the same check daily for weeks — the one-off run lied to me twice.

[Vlad Ilchenko @Advalor]

@AravSrinivas The interesting part for me is what happens when the agent hits a storefront that never published an access policy. Most of the big US retailers still haven't.

[Vlad Ilchenko @Advalor]

@modernretail Curious how many of those holiday channel plans include a path for agent traffic, not just paid and organic. From what I can see, most retailers still have no published rules for it.

[Vlad Ilchenko @Advalor]

@tobi Bring-your-kids day producing better product questions than half the roadmap reviews. "Why the hell not" is an underrated filter.

[Vlad Ilchenko @Advalor]

@AravSrinivas Batch decisions with grounded search is the shape a lot of commerce agents actually need. The hard part is still getting past the storefront's access layer once the decision is made.

[Vlad Ilchenko @Advalor]

https://t.co/XYFytiuzyx

[Vlad Ilchenko @Advalor]

2. One labeled bot. One pass. Top US retailers by traffic.

Methods and limits are on the page — including what this does not measure.
I'll re-run this monthly starting Dec 1. Happy to take method notes if something looks off.

[Vlad Ilchenko @Advalor]

84% of top US retailers publish no rules for AI agents.

I ran a labeled bot against 83 sites. Short thread on what showed up.
35% refused the labeled bot — same share with or without published rules.
17% had an llms.txt.
5% exposed a UCP endpoint.

Most of the rest had no machine-readable policy at all.

[Vlad Ilchenko @Advalor]

@Cloudflare Useful. When I checked 83 retail sites, 70 of them (84%) had no rules for AI agents at all. Most sites haven't decided anything yet, so seeing which agents actually show up seems like the right place to start before writing policy.

[Vlad Ilchenko @Advalor]

@jeff_weinstein @stripe I'd add disputes to that list. When an agent orders the wrong size or the package never shows up, who files the claim, and what proves what the person actually asked for? I spent years on marketplace disputes and that part gets messy fast.

[Vlad Ilchenko @Advalor]

@harleyf @Shopify This is the part merchants underrate. Every new surface wants its own catalog feed, checkout rules and returns flow. If the platform handles that once, the merchant can stay on the product instead of maintaining ten integrations.

[Vlad Ilchenko @Advalor]

@harleyf @johncoogan @tbpn Bringing order history along matters a lot for agents. Sizes and past purchases are probably the most useful thing an agent can know before it buys for you. Does return history come with it too?

[Vlad Ilchenko @Advalor]

@Akamai Labeling the agent up front feels much cleaner than guessing from behavior, especially for merchants who want good agents in and scrapers out. The part I'd want to understand is who issues the tokens and how fast they can be revoked.

[Vlad Ilchenko @Advalor]

@AravSrinivas @Shopify Interesting that store owners are adopting agents for operations first, not shopping. Audits, campaigns and ad spend have clear inputs and a result you can check. Buying on someone's behalf is still a lot harder to verify.

[Vlad Ilchenko @Advalor]

@modernretail Narrow and deep makes sense for a styling business. Fewer SKUs means more fit data per item, which should help the recommendations and the return rate at the same time.

[Vlad Ilchenko @Advalor]

@tobi Agree from the merchant side. Browser agents break every time checkout gets redesigned; a declared tool doesn't. Curious whether stores end up rate-limiting tool calls differently from page loads.

[Vlad Ilchenko @Advalor]

@tobi 0.016% of weights touched and math up to 6x, with no training run. If that holds outside math benchmarks, a lot of small-model fine-tuning budgets look very different next year.

[Vlad Ilchenko @Advalor]

@jeff_weinstein Clever design. Same cost to the company as a bigger payout, but the researcher picks the cause. And $130K for a single-user prompt injection says a lot about where the risk is moving.