What changedAI visibility now has a server-side evidence layer on Vercel. Teams can see which pages bots request, but the gap between retrieval and business value remains.
Vercel AI traffic insights separate crawlers from assistant retrievals
Vercel announced AI traffic insights on October 27, 2026. The new Observability view reports requests by verified AI bot, the most-crawled pages, assistant retrievals, Markdown requests and errors.
The distinction matters. A crawler may fetch a page for indexing or training, while an assistant retrieval is a page request made after a user action or search query. Both are server requests, but neither is a human page view.
For example, a pricing comparison may receive many crawler requests while a technical setup guide receives more assistant retrievals. That difference can guide delivery checks and content review without pretending either page generated revenue.
The new view adds access evidence, not attribution
Vercel AI traffic insights answer whether known AI systems requested a page and whether the request succeeded. They do not reveal the model’s final answer, the passage used, a citation shown to the user or the user’s next action.
A page can be retrieved because an assistant is comparing sources and still never appear in the response. It can also influence an answer without producing a referral that a traditional analytics tool records.
Treat the new data as the first two layers in a longer measurement chain. Keep a separate website analytics implementation for human sessions, useful actions and conversions.
| Layer | What it shows | What it does not show |
|---|---|---|
| Crawler request | A verified AI crawler fetched a path | Indexing, use in an answer or quality |
| Assistant retrieval | An assistant fetched a path after a user action or search | Citation, answer content or user trust |
| Human referral | A person arrived from an identifiable source | Whether the visit created useful value |
| Business outcome | A defined action happened after the visit | Why it happened without further research |
Failed requests are the most actionable starting point
A failed AI request is a concrete delivery problem. The path may return a server error, redirect unexpectedly, block a verified bot or serve a format the requester cannot use.
Start with pages that matter to a buyer, such as pricing, product documentation or comparison content. If an assistant repeatedly requests `/docs/setup` and receives errors, fix that response before rewriting the page for hypothetical answer-engine preferences.
The Vercel Observability documentation lists top crawler paths, top assistant paths, failed AI requests and Markdown requests in the AI Traffic view. Pair those paths with a technical SEO audit when canonical, status, rendering or crawl rules are also uncertain.
Markdown requests test delivery, not content quality
The dashboard also shows pages requested as `text/markdown`. That reveals format demand from AI bots, but a Markdown response is useful only if it preserves the headings, links, tables and facts that make the page understandable.
Consider a comparison table whose HTML is clear but whose Markdown output loses row labels. A successful request would hide a content failure. Test important pages in both formats and compare the information, not just the status code.
Use a website content audit to decide whether the page answers one intent clearly before optimizing an alternate representation. A clean Markdown response cannot rescue vague or unsupported content.
CLI metrics make repeatable checks possible with Observability Plus
Vercel exposes five AI traffic metrics through `vercel metrics`: total AI requests, crawler requests, assistant requests, Markdown requests and failed requests. The official CLI documentation says metrics beyond Web Analytics and Speed Insights require Observability Plus.
A team can query `vercel.ai_traffic.assistant.count` for production, group results by an available dimension and export JSON for a recurring review. Vercel advises running `vercel metrics schema` first because dimensions and aggregations depend on the metric and account.
For a weekly check, compare the top assistant-requested paths with failed requests and recent publishing changes. Do not turn a small count into a growth percentage with strategic meaning unless the baseline, bot coverage and time window are stable.
What to do now with Vercel AI traffic insights
The useful first review is operational and editorial. It should end with a short list of fixes and questions, not a new vanity dashboard.
- Check failed AI requests on pricing, product, documentation and comparison paths.
- Compare crawler-heavy pages with assistant-heavy pages instead of combining them into one number.
- Inspect Markdown responses for lost headings, links, table labels and source context.
- Record the date and page changes before comparing request trends.
- Keep human referrals, meaningful events and conversions in a separate measurement layer.
- Treat retrieval as a content review signal, not proof of citation or commercial impact.
Use the data to choose the next test, not declare an AI traffic win
Vercel’s release makes AI bot access easier to inspect on sites already running on the platform. Its strongest use is finding delivery failures and identifying pages worth a closer content review.
The next useful step is to examine the top assistant-requested and failed paths, then connect any change to separate human and business evidence. Seuil can help turn that chain into a measurement plan without pretending a server request is a customer.