> ## Documentation Index > Fetch the complete documentation index at: https://docs.kameleoon.com/llms.txt > Use this file to discover all available pages before exploring further. # Run a weekly experiment review > Get a regular report on your running experiments, with health, results, ideas, and takeaways. Experiments can collect unreliable data for days before anyone notices, such as when a tracking tag stops firing after a site release. Your AI assistant runs the health checks on a site's running experiments, which look for tracking gaps and traffic-split drift, and then reads each experiment's results for its primary goal. You can run the review on demand in any assistant you connect to Kameleoon, and schedule it every week as a task in Claude or ChatGPT. In Claude and ChatGPT, a scheduled run can send the summary to your Slack direct messages, so a problem reaches you before you read the results. ## How the weekly review fits into your workflow * **Check running experiments:** The assistant runs the health checks on each running experiment and reports the checks that didn't pass. * **Read results:** The assistant reports the winner status and the conversion figures for the primary goal of each running experiment. * **Review ideas and takeaways:** The assistant lists the ideas PBX Ideate generated for your pages, and the results Kameleoon recorded for experiments paused or stopped in the past 7 days. * **Schedule the review:** Run the same request every week as a scheduled task in Claude or ChatGPT. ## Before you begin * **Setup:** Connect Kameleoon to your AI assistant. For a chat assistant such as Claude or ChatGPT, see [Connect Kameleoon to Claude and ChatGPT](https://docs.kameleoon.com/mcp/connect-ai-assistants.md). For a coding agent such as Claude Code, Codex, or Cursor, see [Set up MCP clients](https://docs.kameleoon.com/mcp/set-up-clients.md). * **Site code:** The code of the site to review. If you don't know it, ask your assistant to list your sites, for example "List your Kameleoon sites." * **PBX Ideate:** Needed for the ideas section only. The assistant can list ideas only for pages that someone has analyzed in PBX Ideate. Without an analysis, the review reports that no page has ideas. * **Delivery:** Needed for the scheduled review only. The on-demand review doesn't send anything. To send the summary to Slack, you need your Slack user ID, which starts with U, for example `U0123ABCDEF`. In Slack, open your profile, select the three-dot menu, and then select **Copy member ID**. ## Run the review on demand Run the review whenever you want an overview of a site's running experiments. The request only reads data, so no experiment or traffic split changes. ### Ask for the review Ask your assistant for the review to get a report that answers three questions: is each running experiment's data trustworthy, what do its results show, and what do the ideas and past takeaways suggest for the next test? Enter the site code in the Value box, then copy the request into your assistant. When the request works, the report lists each running experiment with its overall health status and its primary goal result. It then lists each health check that didn't pass, with its finding and next step. The ideas section names each analyzed page and its top ideas. If no page of the site has an analysis in PBX Ideate, the ideas section says so. The takeaways section can be empty when you haven't paused or stopped an experiment in the past 7 days. If the site has no running experiments, the list is empty, and the report has nothing to check. ### Read the health check statuses Each health check returns one of the statuses below. The overall status of an experiment is the worst status of its checks. * **PASSED:** The check found no problem. * **WARNING:** The check found a possible problem, such as low traffic or a tracking issue. A tracking warning that continues for 72 hours becomes FAILED. * **FAILED:** The check found a problem that needs action. * **PENDING:** Kameleoon hasn't evaluated enough traffic yet. Run the review again later. * **NOT\_STARTED:** The experiment never started, so it has no data to check. Visit and conversion tracking look at the last 24 hours, or at the last hour during an experiment's first day. Sample ratio mismatch and overlapping exposure look at all traffic since the experiment started. For this reason, a status can change from one run to the next. Kameleoon reuses an evaluation for up to 15 minutes, so a second run within that time returns the same statuses. ### Read the results The review reads each experiment's results for its whole run. A few days of data can be noisy, so don't call a winner from a short window. Kameleoon names a winner only for the primary goal, against the reliability threshold. Secondary goals show the same figures as guardrail metrics, without a verdict, so use them to check for side effects rather than to pick a winner. ### Read the ideas and takeaways Ideas come from PBX Ideate analyses of your pages. Each idea has a score, and the highest scores come first. Takeaways are the learnings that Kameleoon records when you pause or stop an experiment. Resuming an experiment removes its takeaway. ## Send the review every week You can schedule the review in Claude or in ChatGPT. A scheduled task in Claude runs in your assistant's cloud, not on the Kameleoon server you set up in your IDE, so your computer can be off, and the task needs the Kameleoon connector attached. Each run sends its digest to your Slack direct messages. ### Create the scheduled task Create a scheduled task that runs the request every Monday. Choose the tab for your assistant. > [tab] Claude > > Before you start, add Kameleoon and Slack to Claude as connectors. Anthropic's [About custom connectors](https://support.claude.com/en/articles/11175166-about-custom-connectors) article describes how to add one, and [Connect Kameleoon to Claude and ChatGPT](https://docs.kameleoon.com/mcp/connect-ai-assistants.md) covers the Kameleoon connector. > > Create a scheduled task that runs weekly on Monday, at the time you want the summary to arrive. Anthropic's [Schedule recurring tasks](https://support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-cowork) article describes how to create one. Fill in your site code, and choose your Slack user ID in the Variable list and enter it in the Value box. Then click **Copy prompt**, and paste the request as the task's instructions. To find your user ID, open your profile in Slack, select the three-dot menu, and then select **Copy member ID**. > > Check that the task lists both the Kameleoon and Slack connectors. A task without them can't read your experiments or send the Slack message. > > To check a run, open the task in Claude. Each run has its own session, and when a run finishes, the summary arrives in your Slack direct messages. If no message arrives, open the task and check its latest run for an error. > > Each run uses tokens for the data it reads. The [token usage section](https://docs.kameleoon.com/mcp/mcp-server.md#token-usage-and-plan-considerations) on the MCP server page describes what drives usage. > [tab] ChatGPT > > Before you start, connect Kameleoon to ChatGPT as described in [Connect Kameleoon to Claude and ChatGPT](https://docs.kameleoon.com/mcp/connect-ai-assistants.md). To send the summary to Slack, also connect the [Slack app in ChatGPT](https://help.openai.com/en/articles/12525822-using-slack-in-chatgpt). > > Fill in your site code, and choose your Slack user ID in the Variable list and enter it in the Value box. Then click **Copy prompt**. To find your user ID, open your profile in Slack, select the three-dot menu, and then select **Copy member ID**. Then paste the copied prompt into a new chat. > > Check that the new task appears in your scheduled tasks. ChatGPT's [scheduled tasks](https://learn.chatgpt.com/docs/automations) article describes how to view and change them. ## What the summary contains In Claude, the Slack message starts with a summary line, then Problems, Results, Ideas, and Takeaways sections. An Errors section appears only when a tool call fails. ```text theme={null} 3 running experiments checked. 1 PASSED overall, 2 WARNING or FAILED. Problems - Homepage hero banner (id 100234), FAILED. Traffic split: the variation received 42% of visitors, against an allocation of 50%. Next step: check the traffic allocation and any targeting rule that changes who enters the experiment. - Checkout progress bar (id 100871), WARNING. Visit tracking: visitor traffic is low, and no visitor reached the experiment in the last 24 hours. Next step: check that the experiment is targeted correctly and that visitors can reach the page. Results - Homepage hero banner: no clear winner on "Add to cart click". Original 4.2% (210/5,000), variation 4.5% (221/4,910). - Checkout progress bar: no clear winner on "Purchase". Original 12.0% (36/300), variation 12.7% (38/299). - Product page layout: no clear winner on "Add to cart click". Original 3.3% (82/2,485), variation 3.5% (87/2,486). Ideas Homepage: Show the free-returns message near the hero call to action (score 8.4). Takeaways Pricing page layout: LOSS, uplift -3.1%. Tested a single-column plan layout. ``` Tool parameters are in the [Experiment tools](https://docs.kameleoon.com/mcp/tools-experiments.md) and [Ideas and learnings tools](https://docs.kameleoon.com/mcp/tools-ideas-learnings.md) references.