Nobody blocking AI agents has measured what an agent visit is worth

Delta Air Lines is stopping AI agents from booking tickets. Consider what an agent booking request actually is. A person has picked a route, has a payment card ready and has sent software to fill in the form for them. The checkout page exists to take that money, and the airline has chosen to refuse it. When I started reading around this, I assumed Delta and the other sites going dark had some internal figure showing agent sessions were a net loss. I have found nothing of the kind published by anyone, and I think the bans rest on a mix of borrowed numbers and nerves.

The cost numbers people reach for are real, but they describe a different animal. In July 2025, Cloudflare data showed Anthropic crawling about 38,000 pages for every visit it sent back to a site, and OpenAI about 1,091. The Wikimedia Foundation said bandwidth for multimedia had risen 50% since January 2024, and bots accounted for 65% of its most expensive datacenter traffic while making up 35% of pageviews. Those figures measure crawlers hoovering up pages for training and indexing. An agent buying a pair of trainers loads a handful of pages for one customer who already intends to pay. Applying a crawl-to-referral ratio to that session is a category error, and even the crawler ratios are moving: a secondary report of Cloudflare Radar data puts OpenAI at 251:1 and Google at 4.7:1 by July 2026, with falls of up to 95%.

The revenue side is thinner, and I want to be straight about it. Adobe found that in October 2025, visits to US retail sites from generative AI sources were up 1,200% year on year, and those visitors were 16% more likely to buy than visitors from paid search, email, organic search, social and affiliates. Those are humans clicking through from a chatbot, though, which is a different thing from an agent completing the purchase. The figure also swung hard: the same traffic was 23% less likely to convert in July 2025 and 9% less likely in August. So I am making a narrow claim about agent traffic. The best proxy available went from poor to good between July and October 2025, which means any retailer that fixed its agent policy in the summer of 2025 fixed it on stale data.

The security argument deserves a proper answer, because it is the one with evidence behind it. Imperva’s 2025 Bad Bot Report found automated traffic passed human traffic in 2024 at 51% of the web, with bad bots alone at 37%, and said AI is helping attackers build evasive bots that exploit business logic. Travel, retail and financial services took the worst of the API attacks, which happens to be the home turf of Delta, Amazon, JPMorgan Chase and Schwab. Chris Canal of EquiStamp told The Verge that agents trip the same safeguards built for scammers. I believe the banks. Software sitting inside a brokerage account with someone’s credentials has an ugly failure mode, and Chase and Schwab citing security sounds reasonable to me.

For a shop or an airline, though, the security problem is one of identification, and the industry already sorts bots by purpose when it suits it. Cloudflare now blocks crawlers that scrape for search indexing and AI training at once, which requires knowing what each crawler is for. Meta has signed Walmart, Gap, Sephora and Best Buy to let its Muse assistant shop, and Shopify has opened its storefronts to agents. Users on Reddit report checkout at Walmart still failing despite the Meta deal, which tells me generic anti-bot tooling is blunt. It says little about whether a named, accountable agent is worth letting in.

Some of the blockers at least have a priced reason. Yelp lets agents in if they pay for its data licensing programme, which is a business decision with a number attached. Perplexity argues that Amazon wants human eyeballs on its ads, and Amazon says agents should respect a provider’s choice not to take part; an appeals court overturned the order against Perplexity in August and Amazon has filed an amended complaint. I may not like Amazon’s position, but ad revenue per human visit is something it can measure. Cars.com and Amtrak, which users say quietly stopped working with their agents, have offered no reason at all.

If you run a storefront, tag sessions from Muse, Dots, Grok and Instinct separately for a few weeks, by user agent or partner token, and record two figures: conversion rate, and server requests per completed order. Put them next to the same two figures for paid search. If agent orders cost more to serve than they bring in, block them and publish the numbers so the rest of us can check. If you block without running that comparison, you are turning away customers on a hunch.

Prompted by Websites are closing their doors to AI agents, The Verge.