What machines read when they visit an AI tool directory

The most interesting part of our weekly report is where the traffic went. Automated clients retrieved comparisons, execution sandboxes and decision APIs. Browser traffic brought voice cloning and the practical details of using these tools into view ^[1].

We label browser traffic by the browser a request claims to come from. That category can include automation, so it is not a verified measure of human readership.

The more useful story is what was requested. The charts below use our weekly session report, with the traffic classes kept separate ^[1].

Comparisons drew the automated traffic

Head-to-head pages drew more requests than individual listings from both agents and crawlers. That is where this week's story starts ^[1].

A request cannot tell us whether a client was choosing a product, collecting material for a future answer or building an index. There are also more possible comparisons than individual listings, so the totals do not establish a preference per page. They show where the retrieval happened.

The leading comparison in agent-classified traffic was Serper versus You.com ^[1]. The products approach access to the web differently.

Serper exposes Google search results through an API, including news, images and maps ^[2]. You.com has a hosted MCP server for search, content extraction and research. It supports a free search profile without an API key, alongside authenticated and machine-payment routes ^[3].

The choice involves how much of the search-and-read workflow to assemble, how an agent gets access and where a person must intervene. Those details belong in a comparison just as much as a feature list.

The busiest days tell different stories

Browser traffic rose towards the end of the reporting window. Agent traffic was less dramatic, while crawler activity peaked earlier. Combining them into one line would hide those differences ^[1].

We are treating this as a baseline. The preceding week's logs are incomplete, so there is no traffic growth claim here. A busier day is something to investigate before calling it a larger audience ^[1].

New listings brought execution and identity into view

The catalogue grew during the week, but an addition to Anchor Terminal is not necessarily a newly launched product. Among those additions, Microsoft Execution Containers led the browser category ^[1].

Microsoft Execution Containers runs untrusted code, including model-generated code, inside a configurable sandbox. An application embeds the SDK and specifies the workload and its containment policy. Filesystem, network and UI controls depend on the selected backend ^[4].

Its first stable SDK release arrived during the reporting window, giving developers a versioned surface across Rust, .NET and Node.js ^[5].

AgentCore Identity addresses another part of the same problem. It manages identities and credentials so an agent can access services on a user's behalf ^[6].

These are the operational questions that appear when an agent can take actions. Where can its code run? Which credentials can it use? Whose authority does it carry? The traffic gives us a reason to examine these products, rather than evidence of an industry-wide shift.

Agents also read about making decisions

OpenAI Decisions led the new-listing table for agents. Sheets, messaging and marketing APIs appeared alongside it ^[1].

The Decisions endpoint evaluates text and images and returns a probability, a choice from supplied options or a score against ordered levels. Those outputs fit jobs such as routing a request, checking a condition or assigning severity ^[7].

An application choosing which department should receive a ticket needs a bounded answer it can act on. The shape of that answer becomes part of the product.

Voice cloning has a workflow question

Cartesia versus ElevenLabs voice cloning led the browser comparison table ^[1].

Cartesia's instant cloning route starts from a short recording, while its professional route trains on substantially more audio ^[8]. ElevenLabs' Professional Voice Cloning adds an ownership constraint. The speaker must create and verify their own clone, then share it if someone else needs access. Consent alone does not let another person create that professional clone in their own account ^[9].

These are different cloning modes, rather than a direct quality comparison. But they illustrate why the implementation details matter. A proposed workflow can fail before anyone compares the generated audio.

Markdown and JSON are being read

Our head-to-head pages were retrieved in several formats. HTML remained important, and the machine-readable versions accounted for a substantial share of agent-classified requests ^[1].

The rounded shares do not add up exactly. The practical observation is that each representation gets used. Maintaining readable HTML alongside Markdown and JSON serves the traffic we can already see.

The blog had a quieter week

Decision models also led the blog's browser traffic, followed by why we built the directory and our reading of vendors' data terms. The scale is much smaller than the directory traffic, and deserves its own chart ^[1].

These are subjects we can keep investigating in the listings and explain more fully on the blog. Our next measurement question is whether reading a comparison leads to a successful connection and a useful completed task. A fetch alone cannot answer that.

How to read these charts

The figures are frozen to the reporting window, rather than live counters. Views include successful responses and redirects across HTML, Markdown, JSON and other formats. They are requests, not unique readers or confirmed product usage. Each chart panel starts at zero and uses its own scale ^[1].

Browser, agents and crawlers are inferred from User-Agent strings. Agent traffic includes scripts and some search fetchers. Listing additions use the catalogue's listed date; undated entries are excluded. New-listing views count only from the listing date within the week. Previous-period logs are incomplete ^[1].

References

Product sources read on 13 October 2026. Traffic comes from the supplied weekly report; raw logs are not published.

1. [1]Anchor Terminal weekly session report, 6–12 October 2026, not published. Source for every chart, traffic observation, catalogue addition count and reporting limitation in this post.
2. [2]Serper, product page. Source for Google search results and supported search surfaces. serper.dev
3. [3]You.com, MCP server documentation. Source for search, extraction, research and access methods. you.com
4. [4]Microsoft, MXC repository. Source for sandboxed execution, embedded SDKs and policy controls. github.com
5. [5]Microsoft, MXC SDK v1.0.0 release notes, October 2026. Source for the stable SDK release and supported languages. github.com
6. [6]AWS, AgentCore Identity documentation. Source for agent identity and credential management. docs.aws.amazon.com
7. [7]OpenAI, Decisions guide. Source for inputs, answer types and uses of the Decisions endpoint. developers.openai.com
8. [8]Cartesia, Instant Voice Clone guide. Source for instant and professional cloning workflows. docs.cartesia.ai
9. [9]ElevenLabs, Professional Voice Cloning guide. Source for ownership verification and sharing requirements. elevenlabs.io