Vercel AI Traffic Observability: Track Crawlers, Assistant Retrieval and Errors

Vercel added an AI Traffic view to Observability on October 27. It shows verified AI bot requests, separates crawling from assistant retrieval, and surfaces requested pages, Markdown requests and errors. For teams working on AI discovery, it answers a basic question: can these systems actually reach the content?

That is useful evidence, but it needs the right interpretation. A request to your website is not the same thing as a citation in an AI answer, let alone a new customer.

What Vercel's AI Traffic view shows

The release announcement puts the new view under Observability, then AI Traffic, in the project dashboard. It brings together the pages AI systems request and the requests that fail.

Vercel distinguishes crawlers gathering content for indexing or training from assistants retrieving pages in response to a user action or search. The view also tracks requests for Markdown, which can help teams see how these systems are asking to consume their pages.

The distinction matters because the two kinds of traffic answer different questions. A crawl tells you that a system visited a page. Assistant retrieval is closer to an active user workflow, although it still does not reveal the full question, final answer or business outcome.

Start with failed requests and useful pages

The Observability documentation describes both traffic over time and the most-requested paths for crawlers and assistants. It also includes error paths and Markdown paths.

For a developer product, I would start with pricing, API documentation and migration guides. If assistants frequently request a migration guide but encounter errors, fixing that access problem is more concrete than rewriting the page to chase an assumed ranking signal.

The same reasoning applies after a site change. A team could compare request patterns before and after moving documentation, then investigate whether important pages disappeared from the traffic view or started returning errors. This is an operational check, not proof that the change caused more or fewer AI recommendations.

The strongest immediate use is finding broken access to valuable content. A higher bot-request count, on its own, is a much weaker success metric.

AI traffic metrics through the Vercel CLI

Vercel also exposes these signals through its metrics CLI. AI traffic queries require Observability Plus; the CLI documentation lists these metrics for Pro and Enterprise plans with that add-on.

The catalog includes separate counts for AI bot requests, crawlers, assistants, Markdown requests and failed requests. The aggregate metric is vercel.ai_traffic.count. The release gives grouping by botName as an example.

For a reporting workflow, inspect the available fields with vercel metrics schema before building a query. The CLI supports time ranges, project selection and structured JSON output, so a team can bring these observations into an existing report rather than manually copying dashboard numbers.

Keep that report narrow at first. A weekly check of failed assistant requests to documentation is easier to act on than a large chart mixing every crawler and every page.

What this does not measure

There are at least three separate questions in AI discovery: did a system access the page, did it use or cite the information, and did that interaction produce a useful business result?

The new view provides evidence about requests. It should not be presented as a citation or conversion report. Neither a Markdown request nor a crawler visit proves that a model understood the page, retained it or recommended the product.

For founders investing in technical content, that is a helpful boundary. Use these observations to make information reliably accessible, then evaluate citations and customer outcomes separately. Better measurement starts by being precise about what each number can actually tell you.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help founders turn technical product knowledge into clear content and owned distribution, with a useful distinction between publishing activity and evidence that it is working.

Book a 30-minute call.