How to Analyze Search Console With AI: Export, Prompts and Checks

How to analyze Search Console with AI, in Claude and other tools: export the Performance report, what to paste and what not to paste, size limits, prompts for queries and pages, and how to check the answer against the data.
Written with AI, reviewed by humans

In this article

See how your brand shows up in AI answers

Book a demo and find out what to fix first, with a GEO specialist.

Book a demo

Reading time: 8 minutes

To analyze Search Console with AI, start by exporting the data as a CSV. Then use Search Console prompts that ask for the result in steps. The method works for Search Console with Claude, ChatGPT or Gemini. Afterward, check the totals against the file, because AI can get the math wrong.

In short

  • The Performance report exports as a CSV, but the table shows at most 1,000 rows. For more data, Google offers the API and the bulk export to BigQuery.
  • Before you paste client data, check the tool’s data policy, which changes with the plan and the settings.
  • Ask for each analysis in steps, with the math done in code and the source rows shown, and treat the result as a draft.
  • Check rows, totals, CTR and position against the CSV before you take any number to the client.

How do you export Search Console data to analyze it with AI?

To export Search Console data, open the Performance report and set the date range, the search type and the filters. Then click Export. To analyze Search Console with AI, choose the CSV, because it opens in any chat. Claude, for example, accepts XLSX only with code execution and file creation turned on, according to Anthropic’s help.

First, export one tab per question. Google’s help explains that the table groups the data by the dimension of the tab. So Queries lets you read the search text, and Pages lets you read the URLs. In other words, each tab becomes one file.

Two precautions matter before you send the file:

  • False zeros: Google warns that values shown as ~ or – in the report become zeros in the downloaded data. So a zero in the CSV can mean the data is not available.
  • Row limit: the table shows at most 1,000 rows, and anonymized queries are left out, according to the help page on data discrepancies. So count the rows in your file.

To export Search Console data beyond 1,000 rows, there are two official paths. The Search Analytics API accepts 1 to 25,000 rows per request, but it does not guarantee all rows. The bulk export, on the other hand, sends the performance data to BigQuery every day, except anonymized queries.

Before you export Search Console data, it helps to filter. The guide to regex in Search Console shows how to isolate long questions. You can also connect Search Console with Claude Code through MCP, but the servers we found come from third parties. The guide to Claude Code for SEO explains the risks.

Analyze Search Console with AI: queries table read by an assistant, with check badges on the rows

What should you paste, and not paste, when you analyze Search Console with AI?

As a good practice, when you analyze Search Console with AI, paste only aggregated site data: queries, pages, clicks, impressions, CTR and position. Also, do not paste passwords, API tokens or CRM and GA4 exports with personal data. Before you use client data, confirm the client’s permission and the plan’s data policy.

The policy changes with the tool and the plan. The official pages below cover only training and retention:

  • Claude (Free, Pro and Max): you decide whether your chats help train the models. Incognito chats are not used, but chats flagged for safety review may be, according to the Anthropic privacy center. For the commercial products (Claude for Work and the API), Anthropic’s Commercial Terms say it may not train models on customer content from the services.
  • ChatGPT (Free, Plus and Pro): as of October 2026, OpenAI says it may use your content for training unless you turn the option off. Business, Enterprise, Edu and the API do not train by default.
  • Gemini: a subset of chats is reviewed by people, and reviewed chats are kept for up to three years. The Privacy Hub asks you not to enter confidential data while Keep Activity is on. Even with it off, Google says it still uses chats for protection, including with human reviewers.

Still, aggregated data needs care. A query is free text, so it can carry a person’s name. Also, it comes from strangers. Anthropic’s help on creating and editing files warns that hidden instructions in external files can trick Claude when code execution is on, and it recommends that you monitor what it does.

So, when you use Search Console with Claude, treat the CSV as external content. Never ask the AI to obey the text in the rows.

What size limits apply to the file and the chat?

Limits vary by tool and plan, and uploading a file does not guarantee a full analysis. Anthropic’s help pages give 500 MB per file and 20 files in a Claude chat, and 30 MB per file with file creation and code execution. In ChatGPT, spreadsheets and CSV files reach about 50 MB, according to the upload help (as of October 2026).

ToolPublished limitNote
Claude500 MB and 20 files per chatXLSX only with code execution and file creation
Claude (file creation and code execution)30 MB per fileCreates Python scripts
ChatGPTAbout 50 MB for CSVText files: 2 million tokens

The upload limit is only part of the problem. Claude’s help warns that extra token limits may apply, depending on the length of the extracted text. In ChatGPT, the data analysis help says a file can upload without an error and still be too large for a full analysis. In that case, it suggests asking for specific parts or splitting the file.

So shrink the CSV before you send it. Keep only the columns the question needs, cut the long tail and split by period. Then always ask for the count of rows read, in any tool.

Which Search Console prompts work for analyzing queries with AI?

Good Search Console prompts ask for the analysis in steps and for the math in code. With Search Console with Claude, the Anthropic prompting documentation recommends putting long data first and the question last. It also suggests asking for quotes from the material. So the first step is to check rows and totals.

First, the prompt below finds queries with many impressions and few clicks. Replace the values in brackets with yours.

I attached "queries.csv", exported from the Search Console Performance report (period: [dates]; columns: query, clicks, impressions, CTR, position).

Before you analyze, calculate in code and show the code: the number of rows, the sum of clicks and the sum of impressions.

Then list up to 20 queries with more than [X] impressions and CTR below [Y]%, in a table, with the source row number in the file. Describe only the pattern. Do not suggest causes.

Next, this prompt groups the queries by topic and intent, without letting the AI rewrite the text:

Group the queries in "queries.csv" into at most eight groups by topic and intent (informational, comparison, transactional, brand).

For each group, give the name, the number of queries, the sum of clicks and the sum of impressions, calculated in code, and 3 examples copied from the file without changing the text.

List separately the queries that did not fit in any group.

For long questions that look like prompts, the filter is in the guide to regex in Search Console. The same guide shows how to turn those queries into monitoring prompts.

How do you use Search Console prompts to analyze pages?

To analyze pages, compare two exports of the Pages tab, with the same number of days and the same filters. That way, the AI calculates the change for each URL in code instead of estimating it. Then ask it to say whether the drop came from impressions or from CTR, without proposing the cause.

I attached "pages-a.csv" (previous period) and "pages-b.csv" (current period), with the same number of days and the same filter.

Join the files by URL and calculate in code the change in clicks, impressions, CTR and position for each page.

List the 15 pages with the biggest drop in clicks, with the values of both periods side by side. For each one, say whether the drop came from impressions or from CTR. Do not suggest causes.

To find pages that compete for the same search, you need the query and the page on the same row. The API returns several dimensions per row, so that file calls for another prompt:

I attached "query-page.csv" (one row per query and per page, with clicks and impressions).

List the queries where 2 or more pages of the same site appear. For each one, show the URL and the impressions of each page. Sort by the sum of impressions and show the first 10.

If a drop is your problem, the guide to organic traffic drops shows how to investigate the cause. Google’s AI data has its own rules, explained in AI data in Search Console.

How do you check the AI analysis of Search Console with Claude and other tools?

Recalculate the numbers in the CSV itself. AI can get the math wrong or leave rows out, and OpenAI says to review the code, the outputs and the assumptions before you rely on them. The guide to AI hallucination explains why models make mistakes.

Run five checks, in this order:

  1. Rows and totals: compare the row count and the sums of clicks and impressions with the file. The chart total in Search Console can differ, because the table cuts at 1,000 rows and omits anonymized queries.
  2. CTR: according to Google, CTR is clicks divided by impressions. For a group, divide the sum of clicks by the sum of impressions. The average of the row CTRs distorts the result.
  3. Position: the table’s average position is already an average per row, and the chart’s is the topmost result of the whole site. So the simple average of the rows may not match the chart.
  4. Sample: open five rows the AI cited in the CSV and check each value.
  5. Cause: finally, treat any explanation as a hypothesis. Search Console shows what changed, not why.

For example, a hypothetical case shows the CTR error. One page has 1,000 impressions and 100 clicks (10%), and another has 10 impressions and 5 clicks (50%). The simple average gives 30%, but the CTR of the group is 105 ÷ 1,010, or 10.4%.

To check without relying on the AI, run a short Python snippet. Adjust the file name and the column names to your CSV:

import pandas as pd

df = pd.read_csv("queries.csv")
clicks, impressions = "Clicks", "Impressions"

print("rows:", len(df))
print("clicks:", df[clicks].sum())
print("impressions:", df[impressions].sum())
print("Overall CTR: {:.2%}".format(df[clicks].sum() / df[impressions].sum()))

So if a number from the AI does not match the output, ask it to redo the math and show the code. For the broader flow of AI in SEO work, see the guide to AI for SEO and GEO.

Want to know how your brand shows up in AI answers today?

Book a demo and see how your brand shows up in AI answers, with support from a GEO specialist.

Main sources

About this articleWritten with AI and reviewed by humans, with verified sources. Found something out of date? Write to the NotHuman team.

Get new NotHuman articles by email

By subscribing, you agree to receive blog articles by email and to the Privacy Policy. Unsubscribe any time.

Keep reading

Demo

Want to know how your brand shows up in AI answers today?

Book a demo and see how your brand shows up in AI answers, with support from a GEO specialist.

Book a demo

Your privacy

We use necessary cookies to make the site work and Google Analytics and Microsoft Clarity statistics cookies to improve our content. You can decline the statistics ones; in the EU, UK and Switzerland they are only turned on after you accept. Cookie Policy.

Cookie preferences

Cookie Policy