If you run an online store, you are now the most attacked business on the internet. Akamai’s State of the Internet report published on July 15, 2026, titled Securing the Agentic Storefront: Attacks on Commerce, found that 47.9% of all AI bot traffic across Akamai’s global network between July and December 2025 landed on commerce sites. Not media. Not banking. Retail.
The uncomfortable part is not the volume. It is that a large share of that traffic is trying to buy something on a customer’s behalf, and your security stack cannot tell it apart from the traffic trying to rob you.
Half a percentage point separates a customer from a criminal
HUMAN Security’s 2026 State of AI Traffic and Cyberthreat Benchmark Report, released April 9, 2026 and built on more than one quadrillion interactions processed by its platform during 2025, put a number on the problem. According to the report, across every interaction analysed, only one half of one percent separated the rate of benign automation from the rate of malicious automation.
Read that again. An agent that loads twelve product pages in nine seconds, fills a form without a single mouse jitter and submits a payment might be someone’s shopping assistant. It might be a carding script. The mechanics are identical.
HUMAN also reported that agentic AI traffic grew 7,851% year over year, though the firm notes 2024 volumes started from a very low base, so the multiplier reflects early-stage adoption rather than a mature channel. More useful is the shape of that traffic: product and search pages accounted for 77% of agentic activity, account pages 8.82%, authentication flows 4.95% and checkout pages 2.31%.
Checkout at 2.31% sounds trivial. It is not. Those are transactions completing without a human touching a button, which was largely theoretical two years ago.
Three machines are hitting your store, and they want different things
Treating them as one category is why so many stores get this wrong.
Training crawlers pull your catalogue in bulk to build models. Akamai reported that AI training crawlers account for more than 70% of AI bot triggers in commerce, with OpenAI, ByteDance and Anthropic the top three bots observed. HUMAN found retail and e-commerce made up 62.5% of all training crawler traffic it measured.
Real-time scrapers pull live pricing and stock to answer a shopper’s question right now. HUMAN reported scraper traffic grew 597% from January to December 2025. This is the traffic that feeds the AI answer where your product either appears or does not.
Agents do not read your store. They use it. They log in, apply discounts, add cards and submit orders.
The first two are a bandwidth and content-licensing question. The third is a revenue and fraud question. If you have not yet mapped which is which, our guide on how to prepare your store for AI shopping agents covers the groundwork, and Google’s Universal Cart rollout shows how quickly the agentic checkout side is being formalised.
What the attackers actually did
HUMAN’s Satori threat intelligence team documented a carding-style probe run through an AI browser agent. According to the report, the actor cycled through 11 card-add attempts and 6 payment attempts across two sessions, then pivoted to loyalty-point redemption once the card paths failed. Nothing about that technique is new. The delivery mechanism is.
Two other findings deserve attention. Akamai reported that web attacks targeting APIs rose 9% year over year, and its 2026 API Security Impact Study found 85% of commerce respondents experienced at least one API-related incident in the past year while only 22% know which of their APIs expose sensitive data. Separately, Akamai measured average daily phishing volume across its commerce customers rising from 56,600 in February 2026 to 134,600 in April 2026, feeding account takeover and loyalty point theft.
Loyalty points are the soft target. They sit behind weaker step-up authentication than card payments and they convert quickly.
Blocking everything is now a pricing decision
Akamai’s own data suggests most merchants are not blocking. The report found commerce organisations placed more than 90% of their AI bot activity in the “monitor” category, and allowed three quarters of the remaining activity to pass unrestricted.
That is the worst of both worlds. You are neither capturing the value nor managing the risk. Akamai’s recommendation is to move away from binary allow and block models toward risk-based governance that categorises bots by intent and business value. Patrick Sullivan, Akamai’s Chief Technology Officer of Security Strategy, framed it as needing to “architect sites that welcome legitimate AI while aggressively shutting down malicious bots.”
What to do in the next thirty days
Separate agent traffic in analytics before you make any policy call. You cannot govern what you cannot see. Most standard analytics setups surface almost none of this. Segment by user agent and by behavioural signature, then look at what those sessions actually did.
Stop trusting user-agent strings. HUMAN’s Satori team found a significant portion of requests claiming to be ChatGPT, Mistral and Perplexity bots did not originate from those operators’ infrastructure. Verify against published IP ranges and reverse DNS. An allowlist keyed to a string is an open door.
Inventory your APIs, especially the ones behind login. If you cannot list which endpoints return personal data, pricing or loyalty balances, that is the first job. The Akamai study number that should worry you is the 22%.
Add friction where the money is, not everywhere. Rate-limit card-add and payment-retry attempts per account and per session. Six payment attempts in one session should trip something. Apply step-up authentication to loyalty redemption at the same threshold you apply to card payments. Our e-commerce security checklist for small stores covers the baseline controls, and if you take card details on-site, the PCI DSS script requirements are a separate obligation worth confirming.
Decide your crawler policy deliberately. Training crawlers, real-time scrapers and shopping agents each deserve a different answer. Blocking all three to stop scraping also removes you from AI shopping results.
The takeaway
The old question was whether traffic was human. That question is dead. The new question is whether a given interaction is trustworthy, and answering it takes intent signals rather than a bot flag. Merchants who sort this out in 2026 will collect the agentic revenue. The ones who leave the switch on “monitor” will collect the chargebacks instead.
Sources
- Akamai Technologies, “Akamai Research: Commerce Becomes the Epicenter for AI Bot Attacks and Agentic Fraud in 2026”, press release, July 15, 2026, and the underlying State of the Internet report Securing the Agentic Storefront: Attacks on Commerce.
- Akamai Technologies, 2026 API Security Impact Study.
- HUMAN Security, 2026 State of AI Traffic and Cyberthreat Benchmark Report, published April 9, 2026.
- PPC Land, “AI agents are now buying things and fraud looks identical”, June 2, 2026, reporting on the HUMAN Security findings.
Last reviewed: July 27, 2026
Affiliate disclosure: this site may earn a commission from links in this article, at no extra cost to you. It does not change what we recommend.
