Skip to main content
Methodology

How we score a site

Every number on a scorecard comes from something an agent tried on your site. We check what agents can reach, then send them in to do what your customers would.
Inside every category score
Agents trying real tasks
60%
Quick checks
40%
The same split applies to all 8 categories, wherever a site has both.

01

The short version

Quick checks, triage, task generation, dispatch, review and aggregate, then benchmark calibration.

  1. We check the basics

    Before any agent visits, we make a handful of quick requests to your site. Can an agent reach it? Does it say what agents may do? Does it publish the files agents look for? Does the search agents use turn it up?

  2. Agents try real tasks

    Claude, ChatGPT Agent, Gemini, Kimi, Llama and Grok open your site in a real browser and try what a customer would: find a product, read the refund policy, get to checkout.

  3. We combine the two

    Both results feed one score. What the agents managed counts for more, because that is what your customers get.

  4. We calibrate against the benchmark

    The combined score is calibrated and adjusted to better reflect how the site actually performs against every other site we score, so the published number reads the same way in every sector.

02

What the quick checks look for

They run every time a site is scored, and each one keeps its evidence, so a scorecard shows what passed, what failed and why.

  • Can agents get in?

    We read what your robots.txt allows each agent, request your homepage as ourselves, and look for bot challenges that stop an agent at the door.

    robots.txt rules per providerchallenge headers

  • Do agents find you when they search?

    Most agents search before they visit. We ask the search index agents use for your name and for what you do, and check that it holds readable pages from your site.

    Exa or Tavily searchyour namewhat you doindexed pages

  • Can they find their way around?

    We follow the paths a customer would take: pricing, signup, cart and contact, plus your sitemap.

    sitemap.xmlkey task pages

  • Do you tell agents what you offer?

    We look for the files written for agents, and whether your pages can be read as plain text rather than a wall of markup.

    llms.txtmarkdown negotiationstructured data

  • Can an agent connect and sign in?

    We check whether you publish the details an agent needs to use your service on someone’s behalf.

    MCP server cardOAuth and OIDC discoveryA2A agent card

  • Can an agent buy from you?

    For shops, we check whether you advertise a way for agents to check out.

    Universal Commerce Protocol discovery

03

What the agents do

Each sector has its own set of tasks, chosen from what people there ask agents to do.

Each provider runs four tasks on your site: research who you are, your prices and policies; go from your homepage through to signing up or checking out; find how an agent could act for a customer; and find how to reach you. We also ask each model, with no browser, how you compare to competitors, to test what it already knows.

The agent uses a real browser and the same tools on every provider, clicking, typing and scrolling for itself. Every step is recorded, and a separate grader marks each session against a fixed rubric, so no agent marks its own work. Each category averages the providers below.

The top 50 sites get a monthly run of the research, the journey and the competitor question, which cover most of the score. Other sites get the quick checks. A category with no new session keeps its last result. Verified owners can rescan every task and watch the replays.

  • Claude
    claude-opus-5-5
  • ChatGPT Agent
    gpt-6-luna
  • Gemini
    gemini-3.8-flash
  • Kimi
    kimi-k2.6hosted onCloudflare
  • Llama
    llama-4-scout-17b-16e-instructhosted onCloudflare
  • Grok
    grok-4.7hosted onCloudflare

04

What the score is made of

Eight questions, each scored on its own and weighted by how much it matters.

Where the score comes from8 categories · agents 60% · quick checks 40%
Agents trying real tasks60%Quick checks40%Brand awareness10%Discovery15%Information retrieval15%Market ranking10%Accuracy15%Task completion20%Delegated access10%Contact & communication5%

  1. Can agents recognize you exist?

    Brand awareness

    10%

  2. Will they pick you?

    Discovery

    15%

  3. Can they read your site?

    Information retrieval

    15%

  4. Where do you sit in the lineup?

    Market ranking

    10%

  5. Do they tell the truth about you?

    Accuracy

    15%

  6. Can an agent complete a task on behalf of a user?

    Task completion

    20%

  7. Do you let agents in, safely?

    Delegated access

    10%

  8. Can an agent reach a human?

    Contact & communication

    5%

The overall score is the average of the eight, weighted as shown. That raw score is then calibrated against the benchmark: every published score, overall and per category, is adjusted to better reflect actual performance across the sites we score. Calibration never changes the order of sites, only how far apart they read.

Some sites turn our quick checks away (a 401, 403 or 429 on the homepage) or do not answer at all. Unless an agent session got through, we mark those sites not scored: a number would measure our requests, not the site. They keep a scorecard page but stay out of rankings, averages and charts until a check or an agent reaches them.

05

Value at risk

Unclaimed scorecards for sites that sell something show an estimate of the revenue a low score could cost. Free sites (public services, non-profits, open-source projects) don’t get one, because there is no revenue to lose. To tell them apart, Jev, a decision model by TypeSafe AI, reads the site’s own name and description and decides whether it sells anything. When it isn’t confident, a general model makes the call, and a site we can’t place gets no estimate.

It is an estimate, not a measurement. We take the share of sector traffic expected to come through agents by 2027 (drawn from Gartner’s agentic AI forecasts and Adobe Analytics’ data on AI-referred retail traffic), apply it to an illustrative revenue figure for the sector, and scale it by how far the site is from a perfect score. Sources are cited wherever the figure appears.

Owners who claim their domain can swap in their own numbers or hide the estimate.

06 / 06

How our agents behave on your site

If our agent got stuck on your site, an agent working for a real customer would too.

  • They say who they are

    Every request and every agent carries StuntDoubleIndex in its user agent, so you can see us in your logs and name us in robots.txt.

  • They respect your opt-out

    If robots.txt or your headers say no, we stay out.

  • They never pay or sign up for real

    No payments go through and no real registration details are ever entered.

  • We keep no personal information

    Sessions record what the agent saw and did, not who you are.

  • We visit gently

    We cap how often we request pages from each site.

Scoring methodology

v2026.10

Since October 1, 2026

When we improve how we score, or agents gain new or updated technology, we release a new version of the methodology. Every score is stamped with the version it was calculated with, so results stay accurate, comparable and open to scrutiny.