Search Console answers every question you ask it and volunteers nothing. That is why most people open it, look at the line going up or down, and close it again. With the property connected to Claude or ChatGPT, you can ask in sentences instead of filters. Here are the five questions we ask about our own site every Monday, with our own numbers, including the unflattering ones.

Before the questions: two settings

The model needs the search terms, and search terms are what real people typed. In Loopthink the query dimension is therefore off until you switch it on for the connection. For SEO work you want it on. The second setting is in your head: Search Console data lags about three days, so "last week" always means "the week before the last few days". All figures below are for the 28 days up to 10 November 2026.

One habit makes every answer better: ask for the table first and the advice second. A model that has to show its rows cannot round a guess into a finding.

1. Which queries are almost there?

Show me queries with more than 50 impressions and an average position between 8 and 20, sorted by impressions. For each one, name the page that ranks.

This is the classic, and it is the classic for a reason: a page on position 14 has already convinced Google that it is relevant. It is cheaper to move it to position 6 than to rank a new page from nothing.

Our own answer was humbling. The query connect google search console to claude showed us 320 times at an average position of 18 and brought one click. The same sentence with ChatGPT: 195 impressions, position 11, no click at all. That is demand for the exact thing we sell, parked on page two.

2. Where does the click rate not fit the position?

List queries where we rank in the top 5 with at least 100 impressions and a click rate under 1 percent. What do the titles and descriptions of those pages say?

A top position without clicks has two possible causes. Either the snippet promises the wrong thing, or the searcher wanted something else. We have one of each. A comparison page of ours sits on position 3.5 for an "alternatives" query with 189 impressions and zero clicks, which smells of a title that does not answer the question. And our connector page for a project tool collected 399 impressions for that tool's name plus "login". Those people want to log in, not to read about connectors. No title fixes that, and the right action is none.

Telling those two cases apart is the work. The model is good at proposing which is which, and you should still check five of them by typing the query yourself.

3. Which queries land on two of my pages?

Group by query and page. Show me queries where two or more of our URLs each have more than 50 impressions.

In the Search Console interface this takes a filter per query. In a chat it is one question, and it found our most useful result. For the Claude query from above, two of our pages take turns: the overview page for Search Console at position 20 with 275 impressions, and the page written specifically for Claude at position 31 with 227. The more specific page loses to the general one, and together they hold each other on page two. That is a title and internal linking problem, not a content problem, and we would not have looked for it.

4. What is slipping?

Compare the last 28 days with the 28 days before, by page. Show pages that lost more than 20 percent of their clicks, with the position in both periods.

The interface can compare two periods, but it cannot explain them. Ask the follow-up straight away: is it the position that dropped, or the impressions at the same position? The first means a competitor or an update. The second means fewer people searched, and no edit to the page brings them back.

5. What is new?

Which queries had impressions in the last 28 days and none in the 28 days before? Group them by topic.

New queries are the cheapest keyword research there is, because Google is telling you what it already considers you relevant for. In our list this month: the "alternatives" query from question 2, which did not exist for us four weeks earlier and now has 189 impressions. So the page is new to Google, and a weak snippet on a new page is worth fixing early. The same question, asked the other way round, shows what keeps coming back: Spanish versions of the Claude query have appeared in both periods. We have no Spanish page. Whether we want one is a business decision, but the question would not have come up from a dashboard.

What the model cannot know

  • Search volume. Search Console knows impressions, which depend on your position. A query with 20 impressions on position 40 can be large.
  • Competitors. You see your own rows only. Who holds positions 1 to 5 is something the model would have to look up, or invent. Tell it not to invent.
  • Rare queries. Google withholds queries that few people typed, so totals by query are always lower than totals by page.
  • Cause. The data shows that a page dropped. Why it dropped is a hypothesis, however confident the paragraph sounds.

Why a connector and not an export

You can paste a CSV into a chat and get the first answer. The second question needs a different export, and the third needs query and page together, which the interface caps at a thousand rows per download. With a connector the model fetches what the question needs, and you stay in the conversation.

For a team there is a second reason. With Loopthink the Google sign-in happens once, with the read-only scope, and every colleague then uses their own login in Claude or ChatGPT. Nobody gets Owner rights on the property to see a chart, and the audit log shows who asked what. If you are comparing ways to connect it, we looked at seven of them in Search Console MCP servers compared.

Search Console in Claude →
Search Console in ChatGPT →