TL;DR
| Question | Answer |
|---|---|
| What does each source give you? | Search Console: pre-click. GA4: post-click. |
| What do you join on? | Landing page URL, after normalizing. |
| Why do the numbers differ? | Clicks and sessions measure different events. |
| Biggest gotcha | Search Console lags a couple of days. GA4 does not. |
| Best first question | Which pages get impressions but no engagement? |
| Do you need BigQuery? | No, if the agent can call both APIs and join in memory. |
Search Console knows which queries brought someone to your page. GA4 knows what they did once they arrived. Neither knows the other half.
That gap is why SEO reporting involves so much copy and paste. You export one, export the other, match rows in a spreadsheet, and rebuild it next month.
An AI agent with read access to both can do the join at query time. This guide covers how the join works, the four things that break it, and what to ask once it works.
What joining Search Console and GA4 means
Joining these two sources means pulling a row set from each API and matching them on a shared key, so one answer contains both pre-click and post-click metrics for the same page.
Search Console returns queries, impressions, clicks, click-through rate, and average position. GA4 returns sessions, engagement rate, average engagement time, and conversions. The shared key is the landing page URL.
Done well, you get a single table where every row tells the whole story of a page. Done badly, you get two tables side by side with mismatched rows and no way to tell which mismatch is real.
Why this pairing is the one worth building first
Of all the connections marketers ask for, this is the cheapest useful one. Both APIs are free. Both cover data you already own. Neither requires a warehouse.
It is also the most requested. In a Reddit thread where someone shipped an MCP server for Google Analytics, the immediate reply was "Please do search console!" The author's own answer explained the motivation for building any of it: "all the default views in GA's web UI are absolutely useless these days."
There is a supply gap worth knowing about. Google publishes an experimental MCP server for Google Analytics, and one for Google Ads. There is no official Search Console equivalent. The community filled that void with a large number of independent implementations, which is convenient until you need one that is still maintained next quarter.
The four things that break the join
Most failed joins fail for one of these reasons. All four are fixable once you know to look.
URL format mismatch
Search Console returns full URLs, including protocol and domain. GA4 returns page paths. https://example.com/pricing/ and /pricing are the same page and will not match on a naive string comparison.
Normalize one side before joining. Strip the protocol and domain from Search Console, then decide a rule for trailing slashes and query strings and apply it to both.
Date lag
Search Console data settles a couple of days behind real time. GA4 is close to live. Ask both for "yesterday" and Search Console returns little or nothing, which reads as a traffic collapse rather than a reporting delay.
Pin both queries to the same settled window. Absolute dates beat relative ones here, for the same reason they do everywhere else in AI analytics.
Clicks are not sessions
These will never reconcile exactly, and chasing that is wasted effort.
A Search Console click is Google recording that someone went from search to your site. A GA4 session requires your tag to fire, which needs the page to load and consent to pass. Ad blockers, fast bounces, and consent mode all remove sessions that Google still counted as clicks.
Expect Search Console clicks to run higher. Treat a stable ratio as healthy and a sudden change in the ratio as the actual signal.
API limits
Both APIs cap what a single request returns. Search Console enforces per-site quotas and a row limit per request, documented in Google's Search Console API limits. GA4 applies its own quotas and can sample once a query crosses its processing thresholds.
The failure mode here is silent. A capped result set looks like a complete one to an agent summarizing it. We covered why that matters in why your GA4 MCP server returns confident, wrong numbers.
Two ways to do it
You have a choice between modelling the join once or doing it per question.
| Warehouse approach | Agent approach | |
|---|---|---|
| Setup | Export both to BigQuery, model the join | Connect both sources, ask |
| Time to first answer | Days | Minutes |
| Historical depth | Full retention you pay to store | Whatever the APIs return |
| Cost | Storage plus query | Per query |
| Changing the question | Edit the model | Ask differently |
| Best for | Fixed recurring reporting | Investigation |
Neither is wrong. They solve different problems, and most teams eventually run both. If the same report ships every Monday, a pipeline is more reliable than a chat session someone has to remember to open. If you are chasing a drop that started on Tuesday, the agent wins outright.
This guide covers the agent approach.
Setting it up
The steps are the same regardless of which AI tool you use.
- Connect Search Console. Authenticate against the property you want. Read scope is enough. If you manage client properties, check whether your tool supports multiple accounts without re-authenticating each time.
- Connect GA4. Authenticate against the property that corresponds to the same site. Confirm the property ID matches the domain in Search Console, because mismatched pairs produce joins that run cleanly and mean nothing.
- Confirm the connection is read-only. Neither source should need write access for reporting. Some community Search Console servers expose destructive operations behind a flag. You want the ones that do not.
- Set your URL normalization rule once. Decide how you handle trailing slashes and query parameters, and state it to the agent so it applies consistently.
- Ask a question you already know the answer to. Verify against both UIs before trusting anything new.
With Sequel, both connectors live in one workspace, so this is two authentications rather than two servers with separate credential handling. Step-by-step setup guides are available for connecting Search Console to Claude Code and connecting Google Analytics to Claude Code, plus ChatGPT and Claude Desktop variants.
claude mcp add --transport http sequel https://api.sequel.sh/mcp \
--header "Authorization: Bearer sql_your_api_key"Five questions worth asking
These are the ones that justify the setup, because none can be answered from either source alone.
Which pages get impressions but no engagement? High impressions with low engagement rate usually means the query intent does not match what the page delivers. This is the single most actionable output of the join, and it is invisible in either tool by itself.
Which pages convert well but rank poorly? Strong engagement and conversions on a page with low impressions means you have a proven asset nobody can find. That is an internal linking and coverage problem, not a content problem.
Where did the click-through rate change without a position change? If position held and CTR dropped, something changed in the result presentation. A competitor's title, an AI overview, a new rich result. Position alone would have told you nothing was wrong.
Which queries drive sessions that actually complete a goal? Search Console ranks queries by clicks. Joined with GA4 conversions, you can rank them by outcome instead, which frequently reorders the list.
Did a traffic drop start before or after the click? An impressions drop is a ranking or demand problem. A sessions drop with flat impressions and clicks is a site problem. Distinguishing those in one query saves a day of investigation.
Verify the join before you trust it
The join is easy to get subtly wrong, and a subtly wrong join is worse than none.
- Pick five pages and check both numbers by hand against each UI.
- Count your rows. If the joined output has far fewer rows than either source, your URL normalization is dropping matches.
- Check the tail, not the top. Your top pages will match under almost any rule. The tail is where format mismatches hide.
- Ask for the query. Any tool that will not show you which date range, filters, and matching logic produced a number cannot be verified.
That last point matters more than the rest combined. Practitioners in these threads consistently ask for auditability over autonomy, and it is the difference between a report you can defend and one you can only hope about. Our guide to securely connecting data to AI agents covers the access side of the same principle.
How Sequel does the join
Sequel connects to Search Console and Google Analytics in the same workspace and joins their results at query time. You ask one question in plain English, and it calls both APIs, normalizes the keys, and returns a single table.
Because the sources live behind one authenticated workspace, adding a second AI tool is a permission grant rather than another round of OAuth per source. That addresses the multi-account setup problem that agencies run into hardest.
Every query is shown before you act on it, including the date ranges and matching logic used, so a bad join surfaces as visible logic rather than a wrong number. Access is read-only, and results return as conversation context.
You can extend the same join further. Add Stripe to see which queries drive paying customers, or HubSpot to see which drive qualified leads. That is the pattern we describe in MCP for databases and argue for in what 30 Reddit threads reveal about connecting marketing data to AI.
For the broader category, see AI agents for marketing analytics.
Start with one page
The fastest way to trust this is to run it on a page you know well. Pick one, check what Search Console says brought people to it, check what GA4 says they did, and see whether the joined answer matches your intuition.
Get started free and ask which of your pages get impressions but no engagement.