Connect Meta Ads and Google Ads to Claude, with your store
In short
To compare ad spending with store revenue in Claude, connect authorised data sources through an MCP server and check that dates, currencies and metric definitions match. Loyalz exposes connected advertising and store data through one query layer. The cross-source comparison described here is at merchant and time-period level, not per ad.
An MCP connection can expose advertising reports, store records or both, depending on the server. Check the actual tools and permissions rather than inferring coverage from the platform name. This article describes the local Loyalz source reviewed on 24 September 2026; availability can differ by deployment and connected sources.
What an MCP server for marketing data actually is
MCP is a protocol for handing a model a set of tools it can call. An MCP server for marketing data exposes your accounts as functions — fetch this metric, over this date range, broken down this way — so the assistant queries the data instead of you pasting a screenshot of it.
What separates one server from another is not the protocol but the tools behind it. The Loyalz Insights MCP described here has eleven: four read data and dashboards, seven edit dashboards, widgets and calculated metrics.
| Check | Platform-specific connector | Loyalz Insights MCP described here |
|---|---|---|
| Data coverage | Whatever the server and account permissions expose | Connected advertising and store sources |
| Comparison | Requires compatible definitions across sources | Shared metric layer; definitions still need checking |
| Cross-source grain | Depends on the integration | Merchant and day, aggregatable over time |
| Writes | Inspect the tool list and permissions | Dashboards, widgets and calculated metrics |
Why one platform at a time is the wrong shape
A report from an advertising platform and a report from the store can count different things. Differences in attribution, payment status, refunds, time zones and reporting dates need to be reconciled before comparing totals.
Several connectors can support a valid comparison if the join keys and definitions are explicit. A shared layer makes those rules reusable, but does not make the data correct automatically. Ask the assistant to state the source, period, currency and grain with each result.
MCP transfers tool inputs and results between a client and server; it does not keep returned data away from the client or model. Review the data your tools return and the client's handling of it. The MCP documentation describes tool calls and results.
Claude ads reporting: what you can actually ask
Eleven tools cover the Insights surface. Four read, seven write.
| Tool | What it does |
|---|---|
query_metrics | Spend, revenue, orders, ROAS and the rest, over a date range |
list_metric_catalog | What is queryable for this merchant, including custom formulas |
list_dashboards / get_dashboard | What already exists |
create_dashboard / update_dashboard | Author a view |
add_widget / update_widget / remove_widget | Put things on it |
create_calculated_metric / delete_calculated_metric | Define a ratio once, reuse it everywhere |
The read side answers questions. The write side means you can say "build me a view of paid spend against settled revenue by week" and get a dashboard rather than a wall of numbers you then have to rebuild by hand each month.
Two honest limits on the join, both worth knowing before you ask:
- It is account level. Spend and revenue meet at merchant and day, not per campaign
and not per ad. Store and CRM rows carry no campaign dimension, so a cross-source ratio asked for by campaign collapses to zero — the code emits a warning saying exactly that rather than returning a confident nothing.
- Currency conversion depends on rates being loaded. Where a rate is missing the
converter falls back to 1.0 and logs it. On a single-currency account this never matters; on a mixed one, check the rates before trusting a blended total.
AI agent ad account access: what it can change, and what records it
Inspect read and write permissions separately. A dashboard write changes saved reporting; an advertising write could change live spending. They need different controls.
Every write in the Insights surface leaves a row: which merchant, which tool, the arguments it was called with, and when. Seven write tools, seven log calls — dashboard creation, widget changes, calculated metrics, deletions. You can read that log back through the product.
The source review found these limits:
- Logging is incomplete outside Insights. The board and gamification modules contain nine tools in total, of which five write. Those modules do not call the Insights history logger in the reviewed source. Do not read this as a full audit trail for all MCP actions.
- The token model has no expiry or scope fields. Check effective server permissions and revoke tokens that are no longer needed. A model-field review alone does not establish every runtime authorisation rule.
- The described Insights tools do not write to ad campaigns. They query data and edit reporting objects. Check the complete tool list before granting access to another server or product surface.
Grant only what the task needs: a spend-against-revenue question takes query_metrics, not create_dashboard or the other six write tools. Require review for changes with business impact; pausing a live campaign is reversible in the interface but can still lose sales or interrupt a time-sensitive offer.
MCP server for e-commerce data: what it does not do
- It does not fix attribution. It puts the platform's number and the store's number in
one place so the gap is visible, which is a smaller and more honest claim.
- It does not reach per-ad revenue. Joining spend to settled orders at ad level needs
campaign URLs tagged so the ad id survives to the order, and that is a configuration you do, not something the server infers.
- It does not run inside the product. Authoring happens over the MCP server with a
token you mint and your own client; the assistant in the merchant panel is read-only.
- It does not replace a dashboard. It builds them faster and answers the questions
between them.
This sits in the wider question of giving an assistant access to ad and store data without handing it the keys. For how the same layer is used to judge advertising rather than to report it, see ad measurement for e-commerce. For what the numbers mean once you have them in one place, POAS versus ROAS works through the difference between revenue and profit on the same spend.
Common questions
How do I connect Meta Ads and Google Ads to Claude?
Connect the channels to Loyalz, mint an MCP token in the merchant panel, and point your MCP client at the server. The assistant then queries the normalised layer rather than each platform's API.
Is this different from Meta's own MCP server?
Compare the available data and tools, not just the MCP label. A platform-specific connector may expose one source; a multi-source server may expose several. This Loyalz workflow compares connected advertising and store data at account and time-period level.
Can the agent change my ad campaigns?
No. It has no direct access to the ad platforms and cannot pause a campaign or move a budget through this server. It can author dashboards, widgets and calculated metrics.
Is agent activity logged?
The reviewed Insights implementation calls its history logger on seven write paths. Board and gamification modules have five write tools without those log calls. Verify the deployed version and log coverage before relying on a complete audit trail.
What grain does the ad-to-revenue join work at?
Merchant and day, aggregated to week or month. Not per campaign and not per ad — store and CRM rows carry no campaign dimension, and asking for one returns a warning rather than a wrong number.
See it on your data
On a twenty-minute call we connect one ad account and one month of your orders, and work out the gap between what the platform claims and what your shop took. Profit and POAS stay blank until you enter your margin and costs per order.
Book a demo