Vercel's "Is Agentic" score: what it measures, and what we shipped

On the afternoon of August 21, 2026, Vercel announced "Is Agentic", a free public tool at is-agentic.com that grades a website on how easily AI agents can work with it. The scanning and scoring behind it belong to Ora, an agent-experience research company whose fuller ranker scores four separate layers, from discovery and access to usability and payments. Reports are readable in the browser, through a command-line tool, through a read-only JSON API, and through an MCP server.

What the score measures, in plain English

The scoring page describes its subject as "the public parts of a website that AI agents can discover, retrieve, understand, and use." Two tiers of checks produce the number:

• Essential checks are where the bulk of the points live. They cover what the page calls "fundamentals such as server-rendered content, correct HTTP behavior, clear document structure, recoverable errors, and usable controls."
• Recommended checks switch on only when the scan finds evidence of "an API, OAuth flow, GraphQL endpoint, MCP server, developer portal, or commerce surface." A check that does not apply is dropped from the calculation rather than marked failed, so a content site loses no points for lacking a checkout or a developer portal.

Reports also attach what Is Agentic calls an "observed agent journey," a recorded attempt by a single agent to use the site, along with the friction it ran into. That recording sits outside the number, since a single task cannot stand in for every agent or every goal, and it is often the most legible evidence in the report.

The reason this class of measurement exists at all: assistants and coding agents fetch pages the way crawlers do. They read the raw HTML response, and most do not execute JavaScript. A page whose content only appears after scripts run, or whose error pages only render client-side, is partly invisible to them regardless of how it looks in a browser.

What we shipped

The announcement landed on Friday afternoon. By Sunday morning, the following changes were live on one of our partner sites. Each of these is measured, current work, not a plan.

Markdown from the same URL. Public pages now return Markdown when a client sends the Accept: text/markdown request header, at the same URL that serves HTML. The implementation follows the Accept: text/markdown convention: a Vary: Accept response header so shared caches keep the two representations separate, a 406 response when a client accepts neither format, and quality-value handling per RFC 9110's content-negotiation rules. Pages also advertise a .md sibling through <link rel="alternate" type="text/markdown">, because some agents follow an advertised link rather than negotiate. Is Agentic's own report pages serve Markdown at the same canonical URL, which is a reasonable signal of where the convention is heading.

Real 404 documents. A dead URL returned a not-found page whose visible text, before JavaScript ran, was 27 characters. A complete prerendered 404 document existed the whole time; the routing configuration never let an unmatched path reach it. Dead URLs now serve that full document, with recovery links an agent can read in the initial HTML response. The status code was always correct. The document behind it was not.

An llms.txt file with boundaries. We added a when-to-use section listing the work the business takes on, and, as deliberately, what it does not do, since a stated boundary routes an agent as usefully as a stated capability. Two honest labels belong on this item: llms.txt is a proposed convention, not a standard, and Google states it does not use llms.txt for its AI features. We ship it as low-cost signage for the agents that do read it, and claim nothing more.

One identity node, not two. An automated recommendation from a scan suggested adding an Organization block to a site whose existing JSON-LD already declared a LocalBusiness node. On the schema.org type hierarchy, LocalBusiness is a subtype of Organization: the node already is one. Adding a sibling would have created two competing identity nodes for the same business. We kept one node and recorded why, then reported the check itself: on August 21 we sent Vercel the report URL, the affected endpoint, the expected behavior and a curl reproduction, which is what its contact page asks for. An automated reply told us that published address is not monitored for product reports, so we cannot say the correction reached anyone. This is the pattern worth generalizing: a scanner's finding is a claim to verify against the primary source, not an instruction to follow.

And the score moved. The site scored 74 out of 100 when we started on 21 August, and 87 when we checked again on 24 August.

Two things belong beside any number a scanner produces. It measures what the scanner can observe about a site, which is not the same as whether an agent succeeds at what a visitor actually wanted. And the rubric is young enough that a future scoring change could move this number without the site changing at all — which is why the dashboard below stores the scoring-model version beside every score, rather than the score on its own.

Putting a third party's score on a client dashboard

We maintain a platform for our partner sites that we will be announcing in the coming weeks, and it keeps a standing checklist for every site on it: robots.txt resolved per crawler token, sitemap health, structured data read rather than counted, Markdown availability. This week that checklist gained a monthly Is Agentic scan, with the score displayed on the same page each partner already reviews.

Three decisions made a third party's number safe to display, and they generalize to any external score:

1. Store the evidence, not a summary. Every check in the report is stored verbatim beside the score. A recomputed summary can disagree with itself month to month; the stored report cannot.
2. Version the rubric, and refuse false comparisons. Scoring models change. We record the scoring-model version with every run, and when two runs were scored under different models, the dashboard says the model changed instead of displaying a month-over-month difference that would compare two different rubrics.
3. Render the unknown visibly. A check status our code does not recognize renders in neutral grey, never in green. A new status vocabulary should arrive visible, not silently pass.

One operational note from building it, measured on August 22: the service keeps its most recent finished report on hand and serves that, and its own guidance tells you to rescan once a fix ships. A monthly job therefore has to request a fresh scan rather than re-read the stored result. A dashboard number that never moves reads as "nothing regressed," which is worse than no number.

The benefit, without overstating it

Is Agentic states its own limit plainly, telling readers to "use the report as a prioritized technical review, not as a guarantee that every model or browser automation system will complete every task." We would add the equivalent restraint on outcomes. Nothing in this article claims that a higher agent-readiness score produces rankings, traffic, citations, or inclusion in AI answers. Google's guidance for AI features directs site owners back to ordinary search fundamentals, and that matches what we see in practice.

Our own server logs support the same caution about Markdown specifically. Most assistant fetches we observe request HTML, and the convention's agent matrix currently lists the major consumer assistants as HTML-only, with the Markdown-negotiating agents concentrated in developer tools. We shipped same-URL Markdown because the marginal cost was near zero, since the content already exists as Markdown, and because the interfaces reading it are multiplying. That is optionality, honestly priced, not a traffic projection.

Review your site's Is Agentic report with someone from the Veysey Software Solutions team

If you have any questions or would like to review your site's Is Agentic report with someone from the Veysey Software Solutions team, please email us or book some time on our calendar.