Most attribution vendors now offer a connector for Claude and ChatGPT, built on MCP, and most brands that turn one on use it for about a week. They ask a few questions, get some interesting answers, and drift back to the dashboard. The connector was never the problem. Treating it like a novelty search box is. The value shows up when you build it into specific, recurring jobs that used to mean exporting a CSV and spending an hour in a spreadsheet.
This is the workflow I would set up. It applies to any attribution tool with an official connector. Which vendors have one, and on which plans, is covered in attribution tools with AI assistant connectors.
Set it up once, properly
Connect through the vendor’s sign-in flow rather than a pasted API key wherever both are offered, because a sign-in is scoped and revocable. Check what your assistant plan needs, since custom connectors generally require a paid plan and team workspaces often need an admin to add them first. Then do one thing most people skip. Start a project or saved instructions in Claude or ChatGPT that tells the assistant your business basics: which attribution model you treat as the source of truth, what counts as a new customer, your target CAC, and your margin. Every answer gets better when it starts from those. The setup steps for one tool are in how to connect ThoughtMetric to Claude.
The Monday review
The single best use. Every Monday, ask for last week against the prior four weeks: MER, spend by channel, cost per new customer by channel, and new-customer share, with anything that moved more than 15 percent called out. That replaces the weekly export most teams do by hand, and because the assistant pulls the same metrics the same way every week, it is more consistent than a person copying numbers into a sheet. Save the prompt so it is identical each time.
The why did this move question
When a top-line number changes, a connector is much faster than clicking through filters. If MER dropped, ask which channel’s spend rose without a matching rise in attributed revenue, then which campaigns inside it, then which ads. Three follow-up questions get you from the symptom to the ad set in a couple of minutes. This is where connectors that expose detailed dimensions, like ad, product, SKU, and discount code, earn their keep.
Cross-cutting questions dashboards make hard
Dashboards are built around one dimension at a time. Assistants are good at combining them. Which products do new customers from Meta buy first, compared with new customers from Google? Which discount codes bring in new customers rather than repeat ones? Which countries have the best new-customer ROAS on TikTok? These used to be analyst requests. I collected the ones I actually use in the attribution questions I actually ask through MCP.
Using Claude and ChatGPT side by side
If your team uses both, there is no need to pick one. Connect the same attribution tool to each and let people work where they already are. The data comes from the same place, so the numbers match. What differs is the work around the numbers. I tend to use one assistant for longer analysis and written summaries for leadership, and the other for quick questions during the day, but that is preference, not a rule. What matters is that both answer from the same connector rather than from files someone uploaded last month.
Checking the answers
Three habits keep this trustworthy. Ask the assistant to state the date range, attribution model, and metric definitions it used, because a wrong date range is the most common error. Spot-check one number per week against the dashboard. And be suspicious of any figure the assistant calculated itself rather than retrieved, especially ratios built from two separate queries. Connectors that answer from defined metrics reduce this risk a lot, which is one reason to prefer them. Security questions are covered in is it safe to connect your analytics to an AI assistant?.
For a concrete example, ThoughtMetric, which sponsors this blog, offers AI Connectors for Claude and ChatGPT on every plan, answering from its catalog of defined metrics like MER, new-customer ROAS, and cost per new-customer order, with dimensions down to ad, SKU, and discount code.
The routine
- Connect through a revocable sign-in, and save your business basics as project instructions.
- Run the same saved Monday review prompt every week.
- Use follow-up questions to trace any top-line change down to the ad.
- Ask the cross-cutting questions your dashboard makes hard.
- Make the assistant state its date range and definitions, and spot-check one number weekly.
Leave a Reply