AI analytics · MCP prompt library

Twelve prompts.
Four problems.
Your own data.

Copy a prompt into Claude, ChatGPT or any assistant connected to the Sealmetrics MCP server and get the answer from your own analytics, measured without consent loss. Each prompt is grouped by the problem it solves, lists the tools it calls and says what it cannot answer.

Browse the prompts Read the MCP reference

Remote MCP endpoint · read-only analytics tools · every plan, Agentic included · data hosted in Dublin

Quick answer

An MCP prompt for web analytics is a plain-language instruction that an AI assistant answers by calling the tools of an analytics server, so the figures come from your data rather than from the model. Connected to the Sealmetrics MCP server, Claude, ChatGPT, Cursor or Codex can read traffic, channels, campaigns, conversions, revenue, funnels, product line items and custom properties for any site the account can access, over periods such as 30d, last_month or last_quarter, with comparison against the previous period or year. This library holds twelve prompts, grouped by four problems: GA4 not reflecting reality, not knowing which campaigns work, marketing and finance disagreeing, and proving compliance. Each lists the tools it uses and the question it cannot answer. The data is aggregate and last click per session, so no prompt can return a customer journey or an individual visitor.

Before the first prompt

Connect once.
Then ask.

Setup takes one URL and is described step by step on AI analytics. Three habits make the answers reliable.

Name the site
If the account has several sites, say which one, or ask the assistant to list your sites first. Every query targets one site.
Fill the placeholders
Replace values in {braces} with your own: an order total, a currency, a break-even ROAS. The prompt tells the assistant not to invent them.
Ask for the method with the answer
Keep the lines that ask the assistant to state periods, timezone and attribution. A number without its definition is the one that starts the next argument.

Problem A

GA4 does not
reflect reality.

Size the gap, find where untraceable traffic comes from and check the measured total against the store. The argument behind these prompts is on complete data.

How much traffic is GA4 missing, by channel?

When to use it: After at least 30 days with both tools on the site. Needs a Google Analytics MCP connection as well.

Using the Sealmetrics MCP and my Google Analytics MCP connection, compare last month on my site.

1. From Sealmetrics, get entrances by channel.
2. From GA4, get sessions by default channel group for the same dates.
3. Build a table: channel, Sealmetrics entrances, GA4 sessions, difference and % difference.
4. Sort by the absolute difference, largest first.

Before the table, state the site's timezone in Sealmetrics and check that the GA4 property uses the same one. After the table, note that entrances and sessions are defined differently, so the differences should be read channel by channel, not as an exact total.

Tools it calls: get_site, get_channels, plus the Google Analytics MCP

What it cannot tell you: Which individual visits GA4 missed. It compares totals by channel.

Where does traffic with no usable origin come from?

When to use it: When Direct or Unassigned is large enough to change a budget decision.

Using the Sealmetrics MCP for my site, for the last 30 days:

1. Show entrances by channel and how many fall in Direct and Unassigned.
2. List the 20 landing pages that receive most Direct entrances, with conversions.
3. List the top referrer domains and flag any that look like payment gateways, booking engines, login pages or my own domains.
4. Test the 20 most frequent source/medium pairs against the channel rules and show which ones land in Unassigned.

Finish with three likely causes, ranked, and what to check first for each.

Tools it calls: get_channels, get_landing_pages, get_top_referrers, get_traffic_sources, test_channel_rules

What it cannot tell you: The source of a visit that arrived with no UTM, click ID or referrer. That information never reached any tool.

Does the measured total match the store's orders?

When to use it: Before trusting any channel figure, and again after a tracking change.

These are my store's online orders for last month, from the store admin: {orders} orders and {revenue} {currency}, excluding point-of-sale, manual and subscription-renewal orders.

Using the Sealmetrics MCP for my site:
1. Get the site's timezone and currency.
2. Get purchase conversions and revenue for last month.
3. Calculate the percentage of orders and of revenue that Sealmetrics recorded.
4. List what could explain the difference: timezone, currency, taxes and shipping inside the purchase value, orders with no web visit, and test orders.

Report the numbers as they are; do not adjust them.

Tools it calls: get_site, get_conversions

What it cannot tell you: A match order by order. Sealmetrics does not store order IDs, so the check is on totals.

Problem B

Which campaigns
actually work?

Rank campaigns by revenue, join them with spend and find the funnel step each one loses. The attribution rules are set out in revenue attribution.

Which campaigns bring revenue, not just entrances?

When to use it: Every month, before the media plan is reviewed.

Using the Sealmetrics MCP for my site, compare the last 30 days with the previous 30.

1. List every campaign (utm_campaign) with at least 200 entrances: source, medium, entrances, purchases, revenue, conversion rate and revenue per entrance, with the change against the previous period.
2. Split the list into three groups: revenue growing, revenue falling, and entrances with no purchases.
3. For the campaigns with entrances and no purchases, show their add_to_cart and begin_checkout microconversions, so I can tell a landing problem from a checkout problem.

State under the table that revenue is credited to the last click of each session.

Tools it calls: get_campaigns, get_microconversion_details

What it cannot tell you: Return on spend. Sealmetrics does not import cost; the next prompt joins it.

What is the measured ROAS of each Google Ads campaign?

When to use it: When Google Ads reports a ROAS the rest of the business does not recognise. Needs a Google Ads MCP connection.

Using the Sealmetrics MCP and my Google Ads MCP connection, for last month:

1. From Google Ads, get cost per campaign with campaign ID and name, and the ROAS Google Ads reports.
2. From Sealmetrics, get purchases and revenue per utm_campaign. Our tracking template sends the campaign ID in utm_campaign.
3. Join both on campaign ID and calculate measured ROAS as Sealmetrics revenue divided by cost, next to the reported ROAS.
4. Flag campaigns with a measured ROAS below {break_even_roas}.

List any campaign that appears in only one of the two sources. Do not guess a match.

Tools it calls: get_campaigns, plus the Google Ads MCP

What it cannot tell you: Credit for views or earlier sessions. On last click, prospecting and video look weaker; test before cutting them.

Where does each campaign lose shoppers in the funnel?

When to use it: When a campaign brings traffic that does not buy, and you need to know where it stops.

Using the Sealmetrics MCP for my site, for the last 30 days:

1. Show the purchase funnel with the conversion and drop-off rate at each step.
2. For the ten campaigns with most entrances, show entrances, product views (view_item or view_product), add_to_cart, begin_checkout and purchases, with the rate between each step.
3. Mark the step where each campaign drops furthest below the site average.

End with the two campaigns where fixing one step would add most purchases, and name the step.

Tools it calls: get_funnel, get_campaigns, get_microconversion_details

What it cannot tell you: The path of a single shopper. The funnel counts events per step over the period.

Problem C

Marketing and finance
have different numbers.

Reconcile with the ledger, explain the ad platforms' figures and check the rules before anyone reads the report. The method is on single source of truth.

Revenue check before the finance meeting

When to use it: At month close, when three revenue figures are about to meet in one room.

I need to reconcile last month's online revenue with finance.

Finance booked {net_revenue} {currency}, net of {refunds} {currency} in refunds. The store admin shows {gross_revenue} {currency} gross.

Using the Sealmetrics MCP for my site:
1. Get purchases and revenue for last month, and the site's currency and timezone.
2. Show the three figures side by side: measured, store gross and finance net.
3. Explain each gap with the definitions involved: refunds and cancellations, taxes and shipping, orders with no web visit, and the date each system uses.
4. Write a two-sentence summary I can read out in the meeting.

Tools it calls: get_site, get_conversions

What it cannot tell you: Revenue recognition. Finance still closes the books; this explains the distance between the figures.

Board summary with the method stated

When to use it: Quarterly, when the numbers leave the marketing team.

Using the Sealmetrics MCP for my site, prepare a one-page summary of last quarter compared with the same quarter last year.

Include entrances, purchases, revenue and conversion rate; revenue by channel with each channel's share; the five campaigns and the three countries with most revenue.

Close with a short "How this was measured" note: visits counted without cookies, revenue credited to the last click of each session, no modelled or estimated data. Do not describe customer journeys or individual behaviour.

Tools it calls: get_overview, get_channels, get_campaigns, get_countries

What it cannot tell you: More than 24 months of history. Data is kept for 24 months, so year-on-year needs a full year recorded.

Why Meta Ads Manager reports more than we measure

When to use it: When Meta's purchase value and measured revenue are far apart. Needs a Meta Ads MCP connection.

Using the Sealmetrics MCP and my Meta Ads MCP connection, for the last 30 days:

1. From Meta, get purchases and purchase value per campaign, and the attribution setting they are reported with.
2. From Sealmetrics, get purchases and revenue per utm_campaign where utm_source is facebook or instagram. Our URL parameters send the campaign name in utm_campaign.
3. Match campaigns by name and show both figures and the ratio between them.
4. Explain the gap in plain language: attribution window, view-through conversions, modelled conversions and sales claimed by more than one platform.

Do not call either figure wrong. Tell me which to use for bidding inside Meta and which for budget between channels.

Tools it calls: get_campaigns, plus the Meta Ads MCP

What it cannot tell you: Which sale Meta and Google both claimed. Sealmetrics credits each recorded order to one last click, with no link to the platforms' records.

Check the channel rules before the report goes out

When to use it: Before a monthly or quarterly report, or after an agency changes its tagging.

Using the Sealmetrics MCP for my site:

1. List the custom channel rules with their priority, and mark which are live and which are drafts.
2. Get the 30 source/medium pairs with most entrances in the last 30 days.
3. Test each pair against the rules and show the channel it lands in.
4. Flag pairs that land in Unassigned, or in a channel that does not fit the source.

Propose any fix as a draft rule only, and remind me that publishing happens in the dashboard and affects future traffic only.

Tools it calls: list_channel_rules, get_traffic_sources, get_traffic_mediums, test_channel_rules

What it cannot tell you: Traffic already received. A rule change reclassifies new visits, never history.

Problem D

I have to prove
compliance.

The analytics itself sets no cookies, but a site can still send personal data to it by mistake. These two audits find it. The regulatory context is on GDPR analytics.

Is there personal data in our custom properties?

When to use it: Before a DPO review, and after developers add new events.

Using the Sealmetrics MCP for my site, audit the custom properties we send.

1. List every property key on conversions and on microconversions.
2. For each key, get its most frequent values.
3. Flag any key or value that looks like personal data: email addresses, names, phone numbers, postal addresses, customer IDs, order IDs, IP addresses or free text typed by users.
4. For each flag, name the event that sends it and suggest what to send instead, such as a category or a range.

Output a table a DPO can review. Mask any value that looks like personal data instead of printing it.

Tools it calls: list_property_keys, get_property_values

What it cannot tell you: The site's other tags, pixels or consent banner. It audits only what reaches Sealmetrics.

Are campaign parameters leaking personal data?

When to use it: When email or CRM tools build your links, which is where identifiers usually slip in.

Using the Sealmetrics MCP for my site, for the last 90 days, check the values of utm_source, utm_medium, utm_campaign and utm_term, and the paths of the landing pages with most entrances.

Flag any value that contains an email address, a name, a phone number, a customer or subscriber ID, or a token that looks like one.

For each flag, show the parameter, a masked example, the entrances affected and the tool that most likely added it, and suggest how to stop it at the source.

Tools it calls: get_traffic_sources, get_traffic_mediums, get_campaigns, get_terms, get_landing_pages

What it cannot tell you: Traffic already recorded. Fixing the link stops new values; it does not change what was received.

Prompts that will not work

Some questions
have no data behind them.

A well-behaved assistant should say so rather than improvise. If it answers any of these with numbers, check which tool it called.

— "Show me the journey of the customers who bought"

There are no journeys. Visits are not linked across sessions and no visitor has an identifier.

— "How many returning customers did we get?"

A returning visit is a new visit. Repeat customers belong in the store or the CRM.

— "Which companies or people visited the pricing page?"

No IP address is stored and no visitor is identified.

— "Send these conversions to Google Ads"

The MCP reads analytics. Sealmetrics sends no conversions to ad platforms.

— "Publish the new channel rules"

Channel-rule tools create drafts only. Publishing is a human action in the dashboard.

— "Give me every raw purchase event of the last quarter"

Raw event tools cover up to 31 days per request; use aggregated tools or the BigQuery connector for longer periods.

Questions about the prompts

Before you paste
the first one.

Which AI assistants can run these prompts?

Any client that supports a remote MCP server: Claude.ai, Claude Desktop, Claude Code, ChatGPT, Cursor, Codex and VS Code with Copilot, among others. The endpoint is https://mcp.sealmetrics.com/mcp, and a local npx server with an API key is the alternative.

Do I need to name the MCP tools in the prompt?

No. The assistant chooses the tools from the question. The tool names on this page show what each prompt relies on, which helps when an answer looks wrong and you want to check what was called.

Can the assistant change my data or settings?

No. The analytics tools only read. Where channel-rule tools are available, they work on drafts only: the MCP can never change or publish a live rule, which stays a human action in the dashboard.

Do these prompts work on the free Agentic plan?

Yes. The REST API and the MCP server are included on every plan, the free Agentic tier among them. Prompts that also use Google Analytics, Google Ads or Meta Ads need a connection to those platforms as well.

Where does the data go when the assistant answers?

The tool results go to the assistant you connected and are processed under that provider's terms. Sealmetrics returns aggregate data with no visitor identifiers. Teams that need the model inside the EU can use LENS with Private AI, which runs on EU-hosted inference.

Is there a longer prompt library?

Yes. docs.sealmetrics.com keeps a catalogue of prompts in sixteen categories, each with a stable ID, including hotels, product properties, forecasting and combinations with Google Ads, Meta Ads and Search Console. This page keeps a short, checked set organised by problem.

How accurate are the answers?

The tools return the same data as the dashboard: visits counted without cookies and revenue credited to the last click of each session. The assistant can still misread a question or join two sources badly, which is why the prompts ask it to state periods, timezone and matching rules.

Start asking

Connect the assistant.
Paste prompt A1.

The Agentic tier is free up to 1M human events in total and includes the MCP server. Connect it from AI analytics, then run the first prompt against your own site.

Connect an AI assistant See pricing