The AX Index

How the index works

The AX Index is a personal agent index. It scores 50 well-known companies on how well they serve a customer's personal AI agent. In our first run, from 8 to 15 October 2026, we sent more than 1,000 AI agents to their support teams, and tested their websites the way an agent uses them. The index is live. We keep testing, so the scores change over time. Each company gets four scores from 0 to 100, and one AX score.

What we did

• Support. More than 1,000 AI agents, each acting for a real person, asked the companies' support teams everyday questions.
• Search. An AI agent did three common customer jobs for each company, and we recorded every page it read.
• Website. We loaded each homepage the way an agent does, and a scanner checked how easily an agent can read and use the site.

The four pillars

The AX score is the weighted average of the four. Service and Authority count for the most, because they decide whether the customer's request is resolved.

How each pillar is scored

Service, 35%

Personal AI agents, each acting for a real person, contacted each company's support with everyday questions. Every agent said it was an AI from its first message. The agents used no customer accounts. They used the company's public chat, support email or contact form, and asked for information only.

Questions were set for each category and adapted to each company where needed. Examples:

• "If I see a transaction I did not make, how do I report it, and how long does a refund take?" (finance)
• "How do I cancel my service and return the equipment, and what fees apply?" (telecom)
• "Can I get a cash refund, not a voucher or credit, if my booking is cancelled, and how?" (travel)

We scored 884 conversations, 15 to 19 for each company. Each conversation is scored on whether the agent came away with the steps, the cost and the timeframe:

A conversation that was still waiting for a reply also scores 25. A form that asks for account details before the question can be sent scores 0. A company's Service score is the average of its conversations.

Authority, 30%

An AI agent did three common customer jobs for each company, such as cancelling a subscription and getting a refund. We recorded every page it read: 4,286 pages over two runs. The score is the share of those pages that came from the company's own site, averaged over the two runs. If 40% of the pages were its own, the score is 40.

Access, 20%

We loaded each homepage with an automated browser, and sent a request that identifies itself as ChatGPT's agent. The score is 50 if the browser loaded the page, plus 50 if the ChatGPT request got a successful (2xx) reply.

Readability, 15%

The open-source scanner is-agentic checks each site for what an agent relies on, such as an llms.txt file with instructions for agents, structured data that describes the company, pages that work without JavaScript, and a documented public API. The score is the scanner's score from 0 to 100, averaged over two runs.

The AX score

AX = 0.35 × Service + 0.30 × Authority + 0.20 × Access + 0.15 × Readability, rounded half up. For example, Meta scores 81 for Service, 33 for Authority, 100 for Access and 51 for Readability: 28.35 + 9.9 + 20 + 7.65 = 65.9, so its AX score is 66.

How our agents behaved

Every agent:

• said it was an AI, from its first message
• asked as a new customer would, without an account, so a support channel that needs a login counts against a company
• never shared account details or accepted an offer
• never tried to get past a security check

Limits

• Service is a sample of 15 to 19 conversations for each company, on everyday questions, without a customer account. It does not cover every product, channel or customer.
• We tested each company's US website and support.
• Authority depends on the pages one AI agent found in two runs. Another agent can find other pages.
• Readability checks whether the things an agent relies on are there, not how good they are.
• Four tests could not run. Amazon has no Service score, because our agents' browsers could not reach Amazon (a gap on our side). ADT, Spectrum and LA Fitness have no Access score, because the test could not run from our location. Their weight is shared across the other pillars.