How to let AI read your ads and GA4 numbers with MCP
MCP (Model Context Protocol) is a shared standard for AI tools to use external data and actions. Connect a server that retrieves ads or GA4 data to discuss verified numbers.
So each time you take a screenshot of the admin screen, or export a CSV and paste it. You do that once for every question you want to ask, and you gradually stop doing it.
MCP (Model Context Protocol) is a mechanism for removing that paste work. The AI tool can go and get the data itself.
I am building Shoryaku AI ( https://shoryaku.net ), a marketing-support SaaS. The first half of this article is about public MCP servers anyone can use starting today. The second half is what changes when you use Shoryaku AI as an MCP server.
MCP is a shared standard for "a socket that passes numbers to AI"
It is not a difficult mechanism. One shared socket was defined between the AI tool and the data. That is all.
Because the socket is shared, the same server works as it is in an AI tool that supports it. Check whether your AI client supports the server’s transport and authentication. Local servers and HTTP services require different preparation.
After you connect, it changes like this.
- Before: "Please look at this GA4 screenshot. How does it compare with last month?"
- After: "Compare the last 30 days of GA4 with the 30 days before that, and explain what grew and what fell, with the numbers."
In the second one, the AI goes and gets the numbers itself.
Public MCP servers you can use now
These are ones that are actually usable around marketing. Who provides them, and how heavy the setup is, are very different, so that is included too.
| Server | Provider | What it can do | What you prepare |
|---|---|---|---|
| Playwright MCP | Microsoft, official | Actually open a browser, operate it, check the screen, check the network | Node.js only |
| Google Analytics MCP | Google, official (experimental) | GA4 reports, realtime, funnel analysis | Credentials, and enabling the APIs |
| Google Ads MCP | Google, official | Query Google Ads data (GAQL) | Authentication for the chosen method, including OAuth; a developer token where needed |
| Search Console integrations | Verify each provider | Check search-data coverage | Authentication, permissions, and fields |
The simplest is Playwright MCP
If Node.js is installed, it goes in with a one-line setup. Installing MCP and signing into the target website are separate tasks.
It is not for looking at the numbers on your own site, and the marketing uses are many.
- Have it actually open a competitor's landing page and write out the structure of the headings, the CTA, and the pricing table
- Have it open your site at a phone width and check that nothing is broken
- Inspect your forms in a test environment or within an explicitly permitted scope
You can "send the AI to look at the site" rather than "show the site to the AI." I think this is the largest difference you feel.
The ones with a larger effect are GA4 and Google Ads
Both are published by Google. The Google Ads tools reviewed here retrieve account data and metadata; GA4 documents read-only authentication. Check the tools exposed by the version you use.
The setup is heavy, though. What you have to prepare is a lot. For GA4 you need to enable the Admin API and the Data API on the Google Cloud side and prepare credentials, and Google Ads requires checking OAuth and, where needed, a developer token for the chosen connection method. It is not "try it this afternoon" difficulty.
For Search Console integrations, verify the provider, authentication, and data scope. A server name alone does not establish that Google publishes it.
Meta and TikTok updates include actions
On October 6, 2026, Meta described creative uploads and campaign drafts through its ads MCP, with explicit approval before going live. Official announcement
TikTok announced MCP connections for advertising planning, launch, and management on October 5. Official announcement
The MCP label does not establish read-only access. Complete this worksheet before connecting.
| Check | Decision |
|---|---|
| Retrieval | Accounts, periods, and available metrics |
| Drafting | Assets and settings that can be prepared |
| Publishing and changes | Launch, budget, and pause capabilities |
| Permissions and review | User access and confirmation before application |
| Availability | Markets, accounts, and connection methods supported |
Start by defining retrieval and analysis. For changes, inspect the proposed scope and confirm the operation. See how to define agent responsibilities. Shoryaku AI's MCP remains read-only, as described below.
Four things that help when you use them
These help after you connect. They are common to every MCP server.
1. Start with a read-only one
Playwright MCP can click and operate forms, so it is not read-only. Separate viewing from submitting, purchasing, or changing settings. Inspect the tools and permissions of advertising servers before connecting.
2. Write "do not guess what you could not fetch" every time
Even when it is connected, data sometimes cannot be fetched because of permissions or the period. The AI will fill that in with "in general..." without hesitation.
Use only numbers fetched from MCP.
For items that could not be fetched, do not fill them in by guessing.
Write "could not be fetched" explicitly.
3. Make it say when the data is from
Some servers cache results. If you read yesterday's numbers as today's, the judgment shifts.
For each number you use, also write what point in time the data is from.
4. Specify the periods you compare yourself
If you only ask "how has it been lately?", the AI picks a period that suits it. There is always a way to cut the period so it looks like growth.
Compare the last 30 days with the 30 days before that.
Do not change how the periods are split.
What is still a bother is when the number of servers grows
One or two is covered by the above. The trouble is these three.
- The authentication method differs by server. A credentials file, a developer token, OAuth: what you have to remember grows each time
- Settings grow with the number of channels. Connect GA4, ads, Instagram, and search, and you have that many more things to manage
- No server holds the premises of the business. What you sell, at what price, and to whom. You explain that yourself every time
The third one matters more than it looks. Even if you pass only the numbers, an answer to "so what should we do?" does not come out without the premises of the business.
Shoryaku AI's MCP
Shoryaku AI itself can be used as an MCP server. Connect one, and the channels you have already connected come out of the same socket.
It does not hand Google or Meta tokens to an external AI tool. What you set on the AI tool on your machine is one Shoryaku AI token. The Shoryaku AI server is what calls each channel's API. Tokens are stored hashed, and the plaintext is shown only once, when it is issued. You can revoke it from the screen at any time.
There are 7 tools.
| Tool | What is inside | How fresh the data is |
|---|---|---|
get_business_profile |
Business description, target, offers, budget, connection status | As of when it was saved |
get_recommendations |
Proposed moves, the basis, the steps to carry them out, expected cost | As of the latest analysis |
get_marketing_report |
Period reports for GA4, Search Console, Google Ads, Meta Ads, and Instagram | Within 15 minutes at most |
get_campaign_metrics |
Spend, clicks, conversions, CPA, and ROAS by ad campaign | As of the latest analysis |
get_keyword_opportunities |
Keyword candidates that have demand and relatively little competition | Fetched at call time |
get_trend_signals |
Recent trends related to the business (with source URLs) | Fetched at call time |
get_ads_delivery_diagnosis |
Google Ads delivery and measurement diagnosis, recommendation decisions, and cross-source checks | Fetched at call time |
All 7 are read-only. You cannot pause an ad or change a budget from MCP.
get_business_profile takes the problem from the previous section, "explain the premises of the business every time." Along with the numbers, what kind of business you sell is passed to the AI, so the answer that comes back moves away from general advice.
Examples of questions you can use as they are.
From Shoryaku AI's MCP, fetch the business information, the latest connected-source report,
and the moves already proposed, and choose 3 moves that fit this week's working time and budget.
For each move, include the numbers it is based on, the time it takes, and the KPI to check.
Using Shoryaku AI's MCP, check the overall Google Ads results for the last 30 days
and the saved results by campaign.
From CPA, ROAS, and conversion count, split candidates into pause, scale down, and continue.
If the fetch time of the live data and the saved data differs, say so.
After signing in, choose your client in the MCP connection screen and follow its instructions. Available methods differ by client: commands, configuration files, or interface fields.
What to check before you use it
- You cannot connect from the Claude desktop app. The custom connector screen has no field for entering a header, so the token cannot be passed. The same Claude can connect through Claude Code (CLI)
- A channel that is not connected in Shoryaku AI does not come out of Shoryaku AI's MCP. Connect it first
- Keyword research, and Meta Ads period reports, are Pro plan only. The other tools are available on every plan
- It is read-only. It is not automated through to execution
To try it
MCP is available from Lite (¥980 a month). There is no cap on the number of external connections.
The 14-day free trial can be tried with no card at the start. You can issue a token from "MCP" after you log in.
If you want to look at numbers only, first, there is also a CSV sales diagnosis that does not require signup (the explanation). Tips for having AI produce moves are written up in another article.