A Model Context Protocol (MCP) server is a live, read-only connection between an AI assistant and a business system. Instead of exporting a report, uploading it, and asking questions about a file that froze the moment it downloaded, the AI queries the system directly and works from what is true right now.
Akorn runs five MCP servers in production across client accounts: Google Ads, Google Merchant Center, Google Tag Manager, Google Analytics 4, and Google Search Console. All five are read-only. Nothing in a client account can be changed through them. They exist to answer one question faster and more reliably than exports ever could: what is actually true in this account right now?
MCP is not an Akorn invention. Anthropic introduced it as an open standard in November 2024, and by December 2025 it had been donated to the Agentic AI Foundation and adopted well beyond Anthropic. What is specific to Akorn is what got built on top of it, and what that has caught.
The standard way to analyze an ad account looks like this: export a CSV, upload it somewhere, ask questions, find something odd, export again to check. Every answer describes the account as it was at export time. Every rename, budget change, or fixed disapproval after that moment is invisible.
We keep one old export around as a reminder. It is a full-year campaign report where several campaign names in the file no longer exist in the account. They were renamed or restructured after the download. The file was accurate for exactly one moment.
A live query does not have that problem. Ask the same question tomorrow and it describes tomorrow's account.
Each connector answers a different class of question:
Google Ads. Campaign and ad group performance, search terms, asset-level data, conversion configuration, change history. The questions that used to require an export or twenty minutes of UI navigation.
Merchant Center. Feed health, disapprovals with reasons, product-level performance. The Shopping side of the funnel, queryable row by row.
Tag Manager. Which tags exist, which triggers fire them, under what conditions. Whether tracking is configured the way everyone assumes it is.
Analytics 4. Traffic, conversions by event, source breakdowns, and a real-time view of the last 30 minutes, which is how a tag change gets verified minutes after it ships instead of days later.
Search Console. Organic query and page performance, sitemap status, and real-time index checks on individual URLs.
Two properties hold across all five. Everything is read-only, enforced at the system level, not as a habit. And every query is logged against the person who ran it. Anything that would change a live account goes through a human, deliberately. That is a governance decision, not a technical limitation.
A live connection alone produces fast wrong answers. What makes the numbers hold is a maintained set of operating rules for how each one gets calculated. Three examples from ours:
Filter by ID, not by name. Campaigns get renamed. A query filtered on a campaign name silently returns zero rows from the rename forward, and nothing warns you. Standing rule: filter by campaign ID or by what the campaign actually is (its channel type), never by what it happens to be called this month.
Know which conversions column you are reading. Performance Max cost-per-acquisition has to come from one metric and Search from another. Mix them up and PMax cost per booking reads roughly 3x better than it is. The rule specifies the column per campaign type, so the same query produces the same correct number regardless of who runs it.
Halt on unexpected conversion actions. Before any analysis starts, a sanity check confirms the conversion actions in the account match the expected list. If something new shows up, the analysis stops rather than quietly blending a contaminated number into a report.
None of those rules live in the software. They are operating discipline sitting on top of it, written down, versioned, and applied every time. That pairing, live connection plus maintained rulebook, is the whole system.
A few closed examples, details on the linked pages:
On one account, reported conversions were inflated 6.5x because the platform's headline metric silently summed every conversion action together, sign-ups blended with purchases into one number. A live query segmented by conversion action showed it in minutes. Read: 7 things live queries catch →
A task checklist said video conversion goals were configured on June 26. A live pull on July 1 showed zero such conversion actions existed anywhere in the account, while campaigns depending on them were already spending. "Assigned" and "done" are different claims, and only a live query tells you which one you are looking at.
The platform's own reporting view claimed roughly 8,291 products in one catalog. Querying Merchant Center directly showed 4,076 actually active. The platform view retains historical inactive products, and every analysis built on the bigger number was reasoning about inventory that does not exist.
It does not replace judgment. It surfaces the discrepancy and the trade-off; a person still decides the target, the budget, the fix. It is only as good as the underlying systems: if a feed or tracking setup is a mess, this finds the mess faster, but someone still has to fix it. And it does not manufacture history. A live query answers what is true now; longitudinal trending still needs data that was deliberately kept.
Is it safe to connect an AI to my ad account?
The connections here are read-only, enforced at the API-permission level where the platform supports it and at the tool level where it does not. Nothing can be edited, paused, or spent through them. Every query is logged. Any change to a live account is made by a person.
Do I need one of these to work with Akorn?
No. This is infrastructure Akorn built and runs on client accounts as part of the engagement. There is nothing for a client to install or maintain.
Is this the same as Google's AI features inside the platform?
No. Platform AI optimizes inside Google's system toward the signals it is given. This layer sits outside it and checks whether those signals, and the numbers everyone reports from them, are actually correct. The two do different jobs. Read: MCP vs pipeline vs dashboard →
Can this work for systems other than Google?
The pattern applies to any system with a real API, which is most modern business software. What transfers is not the tool, it is the pairing: live connection plus a maintained rulebook for how each number gets calculated.
Your own stack looked at this way — a structured read of ads, feed, tracking, and analytics, with every finding traceable to a query.
Get in touch