Analyze Google Search Console Data with ChatGPT & Claude: 10 Copy-Paste Prompts

Google Search Console knows exactly which queries bring people to your site, which pages are bleeding clicks, and which keywords are one position away from page one. The problem is that all of that sits in a dashboard built for reporting, not for action. You can sort and filter, but you can't ask it "which of my titles are underperforming and how should I rewrite them?" That is precisely the kind of question a large language model is good at — if you hand it clean data and a precise prompt.

This guide is a working library of prompts for analyzing your GSC data with ChatGPT or Claude. The core idea is simple: export your Performance report as CSV or JSON, paste it into the model with a clear structure, and use a tightly scoped prompt that tells the model exactly what columns it is looking at and what decision you are trying to make. No magic, no live API connection — just disciplined prompting against data you already own.

One honest caveat up front, because it shapes everything below: the model only analyzes what you paste. It cannot fetch your live Search Console account, it does not know your traffic unless the numbers are in the text, and it will happily invent a metric if your prompt invites it to. Treat the LLM as a fast analyst with no memory and no database access. Give it the data, constrain its job, and verify the output. Do that, and these prompts will save you hours.

Step 1: Export your GSC data the right way

Before any prompt works, you need the data in a format the model can read. In Search Console, open the Performance report, set your date range, and make sure all four metrics (clicks, impressions, CTR, average position) are toggled on. Then use the Export button to pull the Queries and Pages tabs.

CSV is the most universal format and pastes cleanly into both ChatGPT and Claude. JSON is better when you want the model to reason about nested structure or you are scripting the export through the API. If you are not sure which to grab or how to get a clean export without the UI mangling it, we walk through both in detail in how to export Google Search Console data as JSON or CSV.

A few things to do before you paste:

• Trim to what matters. You rarely need 5,000 rows. Sort by impressions or clicks and keep the top 100–300 rows for the analysis at hand. This respects token limits and keeps the model focused.
• Keep the header row. The column names (query, page, clicks, impressions, ctr, position) are what let the model interpret the numbers correctly.
• Strip anything sensitive. GSC query data occasionally contains personal information people typed into search (names, emails, order numbers). Remove those rows. Never paste PII into a third-party model.

Step 2: Structure the paste so the model can't misread it

The single biggest cause of bad output is ambiguous input. The model needs to know what each column means. Always prepend a one-line schema before your data, like this:

You are an SEO analyst. Below is Google Search Console Performance data.
Columns are: query | page | clicks | impressions | ctr | position
- clicks: total clicks in the period
- impressions: total times shown in search
- ctr: click-through rate as a decimal (0.034 = 3.4%)
- position: average ranking position (lower is better)
Do not invent or estimate any numbers not present in the data. If something
cannot be determined from the data, say so.

DATA:
query,page,clicks,impressions,ctr,position
best running shoes,/blog/best-running-shoes,420,18900,0.022,8.3
...

That "do not invent numbers" line is doing real work — it is your main defense against hallucinated metrics. Paste your CSV rows directly after DATA:. Now you can run any of the prompts below by appending them to (or sending them after) that block.

The prompt library

Each prompt assumes you have already pasted the schema and data as shown above. Adjust the thresholds (position ranges, CTR cutoffs) to fit your site.

1. Find striking-distance keywords

Striking-distance keywords are queries ranking on the edge of page one — positions roughly 8 to 20 — where a modest optimization can produce an outsized traffic gain. This prompt surfaces them and ranks by opportunity.

From the data above, list every query with an average position between 8 and 20
AND at least 100 impressions. For each, output a table with: query, page,
position, impressions, ctr. Sort by impressions descending. Then add a column
"opportunity" estimating relative upside as impressions × (1 - ctr), and flag
the top 10 rows I should prioritize. Only use numbers present in the data.

If you want the full methodology behind this, including why position 8–20 is the sweet spot, see our guide to finding striking-distance keywords in Search Console.

2. Rewrite low-CTR titles

When a query gets thousands of impressions but a click-through rate well below its position would predict, the title and meta description are usually to blame. This prompt finds those pages and proposes new titles.

Identify queries with more than 500 impressions and a CTR below 0.02 (2%) that
rank in position 10 or better — these underperform their ranking. For each one,
output: query, current position, impressions, ctr. Then write THREE alternative
title tags (max 60 characters each) targeting that query. Make them specific and
benefit-driven, not clickbait. Note which emotional or informational angle each
title uses.

Title rewrites are one of the fastest SEO wins available. For a deeper playbook on diagnosing and fixing this, read how to fix low CTR in Google Search Console.

3. Cluster queries to find content gaps

Hundreds of individual queries often collapse into a handful of topics. Clustering reveals topics where you rank for fringe terms but have no dedicated page — your content gaps.

Group the queries above into 8-15 semantic topic clusters based on search intent
and subject, not just shared words. For each cluster, output: a cluster name,
the member queries, total clicks, total impressions, and the page(s) currently
ranking. Then flag any cluster where impressions are high but clicks are low or
where multiple unrelated pages rank — these are likely content gaps or pages
that should exist but don't.

4. Detect keyword cannibalization

Cannibalization happens when several of your own pages compete for the same query, splitting authority and confusing Google about which to rank. This prompt finds those collisions.

Find cases of possible keyword cannibalization: queries (or near-identical query
variants) where MORE THAN ONE page from my site receives impressions. For each
case, output the query, every competing page, and each page's clicks,
impressions, and position. Recommend which single page should be the canonical
target and a one-line reason. Do not flag cases where only one page ranks.

When you spot a cluster, you can confirm it inside Search Console using pattern matching — our guide to Google Search Console regex filters shows how to isolate all variants of a query in the live report.

5. Classify search intent

Knowing whether a query is informational, navigational, commercial, or transactional tells you what kind of page and CTA it deserves. This prompt labels every query at once.

For each query in the data, classify the search intent as one of: Informational,
Commercial Investigation, Transactional, or Navigational. Output a table:
query, intent, position, clicks. Then summarize: what percentage of my
impressions come from each intent type, and where the biggest mismatch exists
(e.g., transactional queries landing on blog posts instead of product pages).

6. Build a content brief from a single query

Once you have a target query, the model can draft a full brief — headings, entities to cover, and the questions a page must answer. Replace the bracketed query with your own.

Act as a content strategist. My target query is "[TARGET QUERY]". From the GSC
data above, list every related query my site already gets impressions for, and
group them as subtopics. Then produce a content brief: a working title, a meta
description, a recommended H2/H3 outline, the key entities and terms to include,
and 5 questions the article must answer. Base subtopics only on queries present
in the data; clearly label any general SEO suggestions as your own additions.

7. Spot decaying pages by comparing two date ranges

Content decay is gradual traffic loss as a page ages or competitors catch up. To catch it, export the same report for two periods (for example, the last 90 days and the prior 90 days), label each, and paste both.

Below are TWO Google Search Console exports for the same site: block A is the
recent period, block B is the prior period (same length). Same columns in each.
Compare them and list pages where clicks dropped by more than 20% from B to A.
For each, output: page, clicks_before, clicks_after, % change, position_before,
position_after. Indicate whether the decline came from lost ranking (position
got worse) or lost CTR (position held but clicks fell). Use only the numbers
given.

PERIOD A (recent):
[paste recent CSV]

PERIOD B (prior):
[paste prior CSV]

Separating ranking loss from CTR loss matters because the fix is different: one calls for content and links, the other for a title and snippet refresh.

8. Generate FAQ schema questions from real queries

Question-style queries in your GSC data are real People-Also-Ask-style demand. Turning them into an FAQ section (and FAQ schema) targets that demand directly.

From the queries above, extract every query phrased as a question or that implies
one (who/what/why/how/when/can/does/is). Group near-duplicates. For each distinct
question, write a concise 2-3 sentence answer suitable for an FAQ section. Output
the result as valid FAQPage JSON-LD schema. Use only questions derived from the
data; do not fabricate questions that aren't represented.

9. Find quick wins across the whole account

When you just want a prioritized to-do list, this prompt scans everything and ranks actions by effort versus impact.

Analyze all the data above as an SEO consultant. Produce a prioritized action
list of the 10 highest-impact, lowest-effort opportunities. For each: state the
action, the specific query/page it applies to, the supporting metrics from the
data, and the expected outcome. Categorize each as Title/Meta, Content,
Cannibalization, or Internal Linking. Rank by impact-to-effort ratio. Do not
recommend anything the data doesn't support.

10. Map queries to the right page

Sometimes the issue isn't the content — it's that Google is ranking the wrong URL for a query. This prompt audits relevance.

For each query above, judge whether the ranking page is the most appropriate URL
on my site to serve that query's intent, based on the query text and the page
URL. Output: query, current page, intent match (Good / Weak / Mismatch), and a
short reason. For every Mismatch, suggest what type of page should rank instead.
Note that you can only judge from the URL and query, not page content.

For a broader view of how GSC feeds your whole keyword strategy, see how to use Google Search Console for keyword research.

Pitfalls to avoid

These prompts are powerful, but LLMs fail in specific, predictable ways. Know them.

• Hallucinated metrics. If you ask "what's my CTR for X?" and X isn't in the paste, the model may invent a plausible-looking number. Always include the "use only numbers present in the data" instruction, and spot-check any figure before acting on it.
• Token limits. Very large pastes get truncated, sometimes silently. The model might analyze only the first portion and present it as the whole dataset. Trim to your top rows, or split the analysis into batches and combine the results yourself.
• No live access. The model cannot connect to Search Console, refresh your data, or "check the latest numbers." Everything it knows is in your paste. If the data is a week old, the analysis is a week old.
• PII and privacy. Search query data can contain personal information. Strip it before pasting, and check your organization's policy on sending data to third-party AI tools.
• Math drift on big tables. When summing clicks or computing percentages across hundreds of rows, models occasionally miscalculate. Verify any totals that drive a real decision — or have the model show its arithmetic so you can check it.

The honest summary: an LLM is an excellent reasoning layer on top of your data, not a source of truth about your site. It reads what you give it and thinks clearly about it. That is genuinely valuable — but the value comes from your export and your prompt, not from the model knowing anything about your website.

Let the analysis run itself

Copy-pasting CSVs and tuning prompts works, but it gets tedious when you do it weekly across multiple properties. Search Console Tools connects to your Google Search Console with read-only OAuth and runs every analysis in this guide automatically — striking-distance detection, low-CTR title rewrites, cannibalization checks, decay comparisons, and AI-generated content briefs — without any pasting. It is free to use. If you want a starting point for your toolkit, we also maintain a roundup of the best Google Search Console tools for 2026.

Frequently Asked Questions

Can ChatGPT or Claude connect directly to Google Search Console?

No. Neither model can log into your Search Console account or fetch live data on its own. They only analyze the text you paste into the chat. To work with your data, you must export it from GSC first (as CSV or JSON) and include it in your prompt, or use a dedicated tool that has an authorized API connection.

What's the best format to paste GSC data into an LLM?

CSV is the most reliable for direct pasting because both ChatGPT and Claude parse comma-separated tables well and it is compact. JSON works when you need structured or nested data. In either case, include a header row and a short schema line explaining each column so the model interprets clicks, impressions, CTR, and position correctly.

How do I stop the AI from making up metrics?

Add an explicit instruction such as "use only numbers present in the data; if a value isn't given, say so" to every prompt, and ask the model to cite the rows it used. Models hallucinate figures most often when a question references data you didn't actually paste, so keep prompts scoped to the data in front of it and verify any number before acting on it.

How much data can I paste at once?

It depends on the model's context window, but practically you should keep exports to the top 100–300 rows for a focused analysis. Very large pastes can be truncated, sometimes without warning, causing the model to analyze only part of your data. If you need everything, split it into batches or use a tool that processes the full dataset programmatically.

Is it safe to paste my Search Console data into ChatGPT?

The metrics themselves (clicks, impressions, positions) are generally low-risk, but query data can occasionally contain personal information people typed into search. Remove any rows with names, emails, or other PII before pasting, and confirm your company's policy on sending data to third-party AI services. When in doubt, use a tool with a clear data-handling policy and an authorized connection instead.

Which is better for SEO analysis, ChatGPT or Claude?

Both handle these prompts well; the differences are marginal for this use case. Claude tends to handle larger pastes and longer tables comfortably, while ChatGPT's data-analysis mode can run actual calculations on uploaded files rather than estimating. Try the same prompt in both and use whichever gives you cleaner, more verifiable output for your data.