Agent Experience Design
AI agents are starting to research, compare and complete tasks for your customers. The Gamma AXD Index measures whether your website lets them. Scan your site against a 100-point rubric and we will email you the evidence.
Scan your website Read the method
Rubric version 1.0. Automated scan, about 30 seconds.
Why now
Websites have been designed on a simple assumption: the person visiting is the one doing the clicking. That assumption is weakening. Three changes are converging, and together they move agent readiness from an interesting experiment to an operational question for digital leaders.
Customers increasingly start with an AI assistant rather than a search results page. The assistant reads, summarises and recommends on their behalf. Generative and answer engine optimisation help organisations get found, but visibility is only the first step.
Browser-using agents can now navigate pages, operate forms and move through multi-step journeys. The question is no longer only whether an agent can find your information, but whether it can request a quote, check a service, book, apply or buy.
Organisations need to decide which agents they allow, what those agents can do, where a person must approve, and how outcomes are verified. Without a deliberate position, they risk locking out a growing channel or exposing actions they cannot audit.
GEO and AEO help organisations get found by AI. AXD makes sure an agent can actually complete the job once it arrives.
Many of the failure points for agents are familiar: unlabelled form fields, state that is only visible, inconsistent product and pricing data, and error messages written for a human glance. The Index turns these into measurable, fixable issues.
The Index
The Gamma AXD Index is a 100-point methodology for scoring how well a website or application supports authorised AI agents. It is built to be evidence-based, repeatable and explainable to both technical and executive audiences. Every point in your scan result links back to a check, the evidence we observed, and what to change.
20 points
Whether agents can reach and render content: crawl accessibility, rendering without fragile client-side dependencies, bot and access policies, and discovery paths.
Example evidence Fetch and render logs, robots and access policy review, sitemap coverage.
20 points
Whether agents can understand what they find: semantic structure, structured data, and clarity and consistency of product, pricing and service information.
Example evidence Schema validation, extraction accuracy tests against source of truth.
20 points
Whether agents can operate the interface: labelled forms and controls, predictable state, accessible names and interpretable errors.
Example evidence Accessibility tree inspection, interaction traces.
25 points
Whether agents can finish defined, authorised journeys accurately, and recover when something goes wrong.
Example evidence Completion rates across repeated runs, step counts, failure analysis.
Indicative in the automated scan. The scan checks that each journey has a findable, readable, operable path. It does not complete the journey.
15 points
Permissions, authentication patterns, human approval points, safe handling of actions, and verifiable outcomes.
Example evidence Policy review, approval flow tests, confirmation evidence.
Indicative in the automated scan. Approval flows and confirmation evidence need a full assessment to test properly.
Readiness bands
Scores translate into four bands. The thresholds are proposed and will be confirmed once the first cohort has been tested.
0 to 39
Significant barriers to access, comprehension or interaction.
40 to 59
Agents can find and partly understand content but struggle to complete tasks.
60 to 79
Core content and journeys work, with specific interaction or governance gaps.
80 to 100
Agents can reliably discover, understand and complete key tasks, with appropriate governance in place.
Testing principles
The credibility of the Index depends on how it is run. The automated scan applies principles 03 and 04 directly: it never signs in or submits anything, and it stays within your robots.txt and sensible rate limits.
Every score is supported by captured evidence that can be shown to the organisation assessed.
Each task is run multiple times across more than one agent platform, with agent versions and test dates recorded.
Controlled test accounts only, no unapproved real-world transactions, and journeys stop before payment or submission unless a sandbox is provided.
Testing stays within published terms and reasonable rate limits, and avoids load that could affect live services.
The rubric is published as version 1.0 and re-run on a regular cycle, so results can be compared over time.
Scores are peer reviewed internally before release, with external review considered to strengthen credibility.
Task library
Each sector is tested against a consistent set of tasks so results are comparable. When you scan, you choose a sector and the report shows which of its tasks have a findable entry point on your site.
| Sector | Representative agent tasks |
|---|---|
| Retail | Find a product matching stated criteria, confirm price and availability, add to cart and reach checkout. |
| Banking and insurance | Locate product features and fees, compare options, start a quote or application. |
| Utilities | Check an outage or service status, find plan information, begin a connection or move-in request. |
| Travel | Search availability for set dates, compare fares or rates, proceed to the point before payment. |
| Government and services | Find eligibility information, locate the correct form, begin an application. |
The crawl
The scan is a single, polite pass over your public pages. It reads what an agent would read, then scores the evidence against the rubric. Nothing is signed in to, submitted or purchased.
Once as the Gamma scanner and once as a standard browser. If the scanner is blocked or challenged while the browser is not, that is recorded as a finding, and the scan stops there rather than working around your bot management.
robots.txt is parsed per agent. Six AI search and assistant agents are tested against it. Blocking training crawlers is treated as a legitimate choice and is not penalised.
XML sitemap, llms.txt, security.txt, and standard locations for OpenAPI descriptions, MCP manifests and agent cards.
Pages are picked from your homepage links by how likely they are to hold journeys: contact, search, pricing, quote, apply, book, checkout, sign in, and your terms. Pages disallowed for our scanner are skipped.
For each page we read the served HTML without running scripts: headings, landmarks, structured data, link and button names, form fields and labels, search, and challenge widgets.
Each check earns points with the evidence beside it. Anything the crawl cannot observe earns neutral credit and is marked as not observed, so a gap in evidence never reads as a pass.
It does not run an agent, so it cannot measure whether a journey completes. It cannot see pages behind a sign-in, content that only exists after scripts run, or the accuracy of your product and pricing data. A full assessment tests defined journeys repeatedly across more than one agent platform and captures evidence for each run.
Requests identify themselves with this user agent. To opt out, disallow it in your robots.txt and the scan will stop at a single page load.
GammaAXDIndexBot/1.0
Requests are sequential, spaced apart, size-limited, and capped at ten seconds each.
Run the scan
Tell us where to send the results. You will see your score here and receive the full report by email, with the evidence behind every point and the five changes that would move your score most.
Scores are indicative until validated through full assessment. No benchmark figures are published and no individual site is named without permission.
Scanning
Gamma AXD Index
An Index score is the baseline. A readiness assessment turns it into a prioritised roadmap, and delivery work in your platform addresses the priority changes. Each step stands on its own.
