Agentic Commerce · Agent Traffic
Meta's Muse Made 59% of AI Agent Requests on HUMAN's Customer Sites. OpenAI Made 30%.
Andreessen Horowitz partners say sites built to block bots must now compete for agents' business.
· 2 min read
- META
Meta’s Muse leads the AI agents reaching the customer sites of cybersecurity company HUMAN Security, with 59% of October’s requests to OpenAI’s 30%, according to a published report on HUMAN’s data.
On HUMAN’s customer sites, two companies’ software sends nearly all agent requests, so a policy on agent access is mostly a policy toward Meta and OpenAI. Andreessen Horowitz partners Olivia Moore and Justine Moore cite the data in an essay on how agents will choose where to shop and subscribe, along with HUMAN’s finding that agent traffic grew 8x between August and October.
Who sends the requests?
| Agent | Share of October agent requests |
|---|---|
| Meta’s Muse | 59% |
| ChatGPT Agent | 16% |
| Codex Browser | 8.8% |
| Codex Chrome extension | 5.2% |
| Perplexity’s Comet | 3.6% |
| Claude Chrome extension | 3.5% |
The three OpenAI products add up to the 30%, which puts Meta and OpenAI together at about 89% of requests. Every other agent is in single digits.
What does the 8x count?
The 8x counts requests that sites allowed and blocked alike, the report said, so it reflects agents trying to reach sites rather than completed visits. HUMAN’s October data put 85% of agent requests on product and search pages, and e-commerce sites drew 60% of agent traffic.
Are retailers really blocking agents?
The essay says most enterprises still block agent traffic over fraud and chargeback worries, but it offers no count of the sites that do. DataDome, a bot-defense company, offers one measure of how many sites could stop automated visitors at all. It sent ten kinds of simulated bots at 21,491 popular sites, and 65.3% failed to detect any of them. Only 2.4% stopped or challenged every type. The test covered homepages, not the login, account and payment pages where some protections may run.
A single request may not reveal a visitor’s purpose, DataDome said: an assistant checking an order can at first resemble a bot trying to get into a customer’s account. Telling them apart will be critical, said Jerome Segura, vice president of threat research at DataDome.
HUMAN reportedly plans a report on how agents behaved over Black Friday and Cyber Monday 2026.
