A/B testing for agencies with many client websites

What an agency needs from a testing tool

1. Separation: one client's people, data and settings never visible to another client.
2. Per-client rules: each brand's voice, words it must never use, pages that must never be tested.
3. Approvals that match the relationship: the client signs off, or the agency does within agreed limits.
4. Pricing that scales with the number of clients, not with every client's traffic combined.
5. Labour that doesn't grow with each new client: research, design, QA and reporting are where the hours go.

How Duelpilot fits an agency

Plans for several sites

Prices are flat per site with no visitor meter; fair use on Team and Code is 10 million pageviews per site per month. Pro and up include unlimited tests, and there are no AI credits to budget per client: the AI runs in your or your client's own agent. Organisations with procurement or security requirements can ask about Enterprise.

Workspaces per client

In Duelpilot, an organisation holds sites, members and one plan. Members see every site in their organisation, and a role applies to all of them. That gives an agency two set-ups:

• One organisation per client: the client's own people, plan and billing, fully separate from your other clients. You are a member of each and switch between organisations in the top bar. Each organisation's plan starts at its own first launch.
• One agency organisation with a site per client (Team or Code): one bill and one place for your team. Suits clients who don't log in themselves, because anyone you invite sees every client site.

Roles

• Owner: everything, including settings, autopilot and custom JavaScript.
• Approver: starts design, approves, pauses and decides tests.
• Viewer: reads the board and results, and comments.

A common split: the client's marketing lead as approver, the agency as owner, stakeholders as viewers.

Brand and guardrails per site

Each site has its own brand kit, measured from the live site in desktop and phone browsers (colours by role, fonts, buttons, spacing), and its own uploaded brand guidelines and confirmed brand rules. Words or claims a client must never use block a variant from being submitted, approved or launched by autopilot. The site profile marks sensitive pages, such as checkout and account, per client.

Autopilot is a switch per site, with its own guardrails: how many autopilot tests may run at once, the maximum share of traffic, and URL patterns it must never touch. One client can approve every test while another runs on autopilot.

The agent per client

The research, variant design, site checks and monitoring are done by an AI agent over MCP: the client's own Claude, ChatGPT or Cursor, or yours. An agent token can be limited to specific sites, and an OAuth connection asks which organisation, sites and permissions it gets. The three routines (Analyst, Designer, Monitor) carry the site's id, so you schedule them per client site. Duelpilot shows per site whether each agent actually runs. See running A/B tests with Claude.

What Duelpilot doesn't have (yet)

• No cross-client dashboard: you open each site's board, and switch organisations in the top bar. There is no single view of every client's live tests.
• No per-site roles inside one organisation: a member sees every site of that organisation.
• No agency or reseller pricing, and no consolidated bill across client organisations.
• No white-label board. Team and Code can embed the board in your own tool with single sign-on, but it is Duelpilot's board. The partner and referral programme isn't open yet; you can leave your details there.
• No visual editor for building variants by hand, and no heatmaps or session recordings.

Other tools agencies use

• Convert offers A/B, multivariate, split-URL and full-stack testing with a visual editor (features) and published prices by tested users (pricing). Choose it if your team builds variants by hand.
• Intelligems lists an Agency plan at a custom price (pricing) and tests Shopify prices, shipping and offers, which a browser script can't.
• Compare more options by who they suit: best A/B testing tools and all comparisons.

Try Duelpilot with your own agent