Skip to main content

Try Base free for 14 days, no card or M-Pesa needed See pricing

Guide

SEO analysis with MCP: from a question to a report you can verify

Last updated
Page tools
Agent setup

Ask your AI agent one bounded SEO question, ground it in your own data, and challenge every finding before the report is saved and shared.

You connect an AI agent to your SEO data and ask why traffic is down. A minute later you have a confident list of causes. Some are right. You can’t tell which.

A better prompt won’t fix that. A narrower question, named sources, fixed periods and a report someone can check will.

The short version. Ask one bounded question. Give the agent your business context and exact periods. Make it show the number, source and period behind every finding, and keep what it can’t prove in the report. Then a person challenges each conclusion before it is saved or shared.

What MCP changes, and what it doesn’t

MCP is an open standard for connecting AI applications to external systems (MCP). Connected to Ndovu, your agent calls the same project data the app shows, instead of guessing from memory.

It does not make the data complete or the reasoning right. Search Console still omits anonymised queries from its query table, shows dates in Pacific Time, and usually has data 2–3 days late (Google). An agent that reads those numbers inherits those limits. Setting up the connection is covered on the MCP server page and in Connect your AI agent.

Step 1 · Bound the question

Too openBounded
Why is our traffic down?Which landing pages lost the most Google organic clicks between two named periods?
Is our SEO working?Did the pages we updated in October gain clicks against the same weeks before?
What should we fix?Of the pages that lost clicks, which ones led to enquiries before the drop?

Define the words that carry the decision. “Useful” traffic here means Google organic visits to pages that produced enquiries or sales, measured by a named GA4 key event, not every visit.

Step 2 · Give it the business context

An agent without context ranks a popular blog post above a page that sells. In Ndovu, Context holds the business overview, current goal, competitors, key pages and a research log. Your agent reads it with get_project_context before it starts. It can write back with update_project_context, but only facts you have confirmed.

Key pages matter most for this question: they are your list of pages that carry the business. If the list is empty, fill it before you ask.

Step 3 · Choose each source on purpose

Each part of the question needs a different source. Name the tool in the brief so the agent can’t substitute one.

Part of the questionSourceNdovu MCP toolWhat it can’t tell you
Clicks and impressions per page, per periodSearch Consoleget_search_console_performance by page, with start and end datesWhy a page lost clicks. Page-level totals can differ from the chart total (Google).
Queries behind one pageSearch ConsoleThe same tool, by query, filtered to that pageAnonymised queries, which are left out of the table.
Whether organic visits fell overallGA4get_google_analytics_organic_overviewWhich pages; it compares totals with the previous equal period.
Visits and key events per landing pageGA4get_google_analytics_organic_landing_pages and get_google_analytics_key_eventsWhether an enquiry reached the business.
Whether tracking changedGA4 set-upget_google_analytics_measurement_healthWhat the numbers were before a change.
Whether a lost page is still indexedSearch Consoleinspect_urlsLive state: it reads Google’s stored index status, up to 10 pages.

All of these read your own connected data and use no research allowance.

Step 4 · Fix the periods before the agent looks

  • Use two periods of equal length that start on the same weekday.
  • End the recent period at least three days ago, so late Search Console data doesn’t look like a drop.
  • For seasonal businesses, compare with the same weeks last year as well.
  • List known changes in both periods: releases, tracking edits, migrations, campaigns.

The task brief you can copy

Project: [project name]
Question: Which landing pages lost the most Google organic clicks, and which of those mattered to the business?
Periods: [start]–[end] vs [start]–[end] (equal length, ending 3+ days ago)
"Useful" means: pages with at least one [key event name] in the earlier period, or listed as key pages.

Steps:
1. Read get_project_context. Note key pages, current goal and recent research.
2. get_search_console_performance by page for each period. Rank pages by absolute click loss.
3. For the top 10, get_google_analytics_organic_landing_pages and get_google_analytics_key_events for each period.
4. For the top 3 useful losers, get_search_console_performance by query, filtered to the page.
5. Run get_google_analytics_measurement_health if key events change sharply.

Rules:
- Every finding states the number, the source tool and the period.
- Label each cause "observed", "likely" or "untested". Never state an untested cause as fact.
- List what you could not check.
- Do not run anything that uses the research allowance without asking me.
- Call list_reports first; save with save_report only after I approve the draft.

Example: a tour operator loses search traffic

Illustrative example, not real data. An invented Mombasa tour operator, example.co.ke. 28 days to 9 November against the 28 days before. The enquiry key event is generate_lead.

Landing pageClicks beforeClicks afterChangeEnquiries before → afterAgent’s noteEvidence level
/tours/tsavo-east-day-trip/1,420860−56031 → 17Lost clicks on “tsavo day trip from mombasa” queries.Observed
/blog/best-beaches-diani/2,3001,650−6500 → 0Biggest loss, but no enquiries in either period.Observed
/tours/wasini-island-dolphins/640410−23012 → 4Page title changed on 21 October.Likely; not tested
/tours/980940−409 → 2Enquiries fell more than clicks.Untested: check tracking
/kiswahili/safari-tsavo/12015−1051 → 0Not indexed now, per URL inspection.Observed

Google organic clicks per landing page, before and after

  • 28 days before
  • 28 days to 9 Nov
/tours/tsavo-east-day-trip/1,420 · 860
/blog/best-beaches-diani/2,300 · 1,650
/tours/wasini-island-dolphins/640 · 410
/tours/980 · 940
/kiswahili/safari-tsavo/120 · 15

Search Console clicks to each landing page in two equal 28-day periods. The beach guide lost the most clicks.
Source: Illustrative example: invented figures, not real data.

Copy linkDownload SVG
Embed

Enquiries per landing page, before and after

  • 28 days before
  • 28 days to 9 Nov
/tours/tsavo-east-day-trip/31 · 17
/blog/best-beaches-diani/0 · 0
/tours/wasini-island-dolphins/12 · 4
/tours/9 · 2
/kiswahili/safari-tsavo/1 · 0

GA4 generate_lead key events per organic landing page in the same periods. The beach guide had none in either period.
Source: Illustrative example: invented figures, not real data.

Copy linkDownload SVG
Embed

Ranked by click loss, the beach guide comes first. Ranked by useful traffic, the Tsavo and Wasini tour pages matter more, and the /tours/ drop may be a tracking problem rather than lost traffic.

Step 5 · Challenge every conclusion

The reviewer reads the draft before it is saved. These are the conclusions agents most often overreach on:

The agent saysAskEvidence that settles it
“A Google update caused the drop.”Did the drop start on one date, across many pages?Search Console by date and by page. One page falling alone is not an update.
“The page lost rankings.”For which queries, and did impressions fall too?Search Console by query for that page, both periods.
“Enquiries fell because traffic fell.”Did the enquiry event or form change?Measurement health, a test enquiry, and enquiries the business actually received.
“The title change hurt the page.”What else changed in the same weeks?A dated change log; otherwise label it a hypothesis.
“A competitor overtook us.”Where is that in the data?Search Console alone can’t show it; results-page data uses research allowance, so ask first.
“Rewrite the page.”Which finding does that fix?Each action must point to a finding in the report.

If the agent can’t supply the evidence, the finding moves to “Open questions”. It does not stay in the verdict.

Step 6 · Save a report someone can check

When the draft survives review, the agent saves it with save_report: a title naming the subject and date, a short summary with the verdict, top action and key numbers, and a self-contained HTML report. It should call list_reports first and replace an earlier version of the same report, rather than adding a near-duplicate. Replacing keeps no history.

Title: Landing pages that lost useful search traffic: 9 Nov 2026
1. Verdict (two sentences) and the single top action
2. Question, periods and the definition of "useful"
3. Sources: each tool used, property, date range, time zone
4. Evidence table: page, clicks, enquiries, change, evidence level
5. Findings: observed facts only, each with its number and source
6. Likely causes and how to test each one
7. Open questions and what could not be checked
8. Actions, each linked to a finding, with an owner and a check date

In Ndovu, it appears under Reports → Saved reports (agents and monitors), private to the workspace. Share → Public link gives a link that needs no sign-in and is hidden from search engines; Export saves a PDF. Agent briefs hold the instructions your agents follow for a report, so the outline above can become one. See Reports, sharing and Context.

Verify the report before you act on it

  1. Spot-check two numbers in Search Console itself, with the same pages, dates and search type.
  2. Re-run one tool call from the brief. The same inputs should give the same numbers, apart from late data.
  3. Confirm every action points to a finding, and every finding has a source and period.
  4. Set a check date. After four weeks, ask the same bounded question about the pages you changed, against the same period length.

Progress is a smaller list of open questions and actions that were checked, not a longer report.

When you need to…Read
Connect your agentConnect your AI agent
See what the MCP server offersMCP server
Fill in Context and share reportsReports, sharing and Context
Explain why clicks and sessions differSearch Console clicks and GA4 sessions disagree
Check whether enquiries really fellWhy GA4 traffic rises without more enquiries

Sources

Next steps

Was this helpful?