# Claude for SEO analysis: prepare the data, check the result

Page: https://alidemirbas.com.tr/en/blog/claude-for-seo-analysis

2026-11-13 · SEO & GEO · Ali Demirbaş

A Search Console export holds at most 1,000 rows. Using Claude for SEO, give it the chart totals, ask for the rows behind each finding and check the math.

Using Claude for SEO analysis means handing Claude your own Search Console data and asking it to find query and page patterns. Every finding should point back to rows in the file. Claude sees that data only when you upload it or connect it through a connector, so the analysis can't cover more than you give it.

A Search Console export from the interface holds at most 1,000 rows and never shows anonymized queries. If Claude sums those rows to compare two periods, the total change it reports can be off. Giving Claude the chart totals separately and asking it to tie every figure to rows prevents that mistake.

To decide which issue to fix first, use the priority order in the [SEO checklist](https://alidemirbas.com.tr/en/blog/seo-checklist-audit-guide). Claude speeds up the measurement and content checks on that list, but it doesn't make the call for you.

## Which Claude feature fits which SEO analysis

Where your data comes from decides which Claude features you'll use. File uploads cover query and page performance. Reading competitor pages takes web search, and a weekly report needs a setup that persists.

| Analysis | Claude feature | Data needed | Watch for |
| --- | --- | --- | --- |
| Query and page performance | File upload and code execution | Search Console CSV export | 1,000 rows at most |
| Cannibalization candidates | File upload and code execution | Export with query and page dimensions | A candidate list, not proof |
| Reading competitor pages | Web search | Queries and competitor URLs | The page as of that day |
| In-depth topic research | Research | A question and its scope | Paid plans only |
| The same report every week | Projects, a connector or Claude Code | Data in the same format each time | The connector's permissions |

The first two jobs, performance and cannibalization analysis, rely on Claude running calculations on the file. As of November 2026, Anthropic's page on [creating and editing files](https://support.claude.com/en/articles/12111783-create-and-edit-files-with-claude) lists code execution on Free, Pro, Max, Ultra, Team and Enterprise plans. It also says you can upload CSV files for data analysis. When Claude does the math in code, you can see each step and repeat it yourself.

Reading competitor pages takes [web search](https://support.claude.com/en/articles/10684626-enabling-and-using-web-search). On Team and Enterprise accounts, an owner has to turn it on for the whole workspace first. A page Claude finds through search shows how it looked on the day you searched. Note that date when you compare competitor titles and coverage.

For longer research that draws on several sources, there's [Research](https://support.claude.com/en/articles/11088861-using-research-on-claude). As of November 2026, it works only on paid plans.

If you produce the same report every week, the first option is [Projects](https://support.claude.com/en/articles/9517075-what-are-projects). You put site notes, past reports and your analysis instructions in the project knowledge, and Claude uses it in every chat inside the project. Free users can create up to five projects.

The second option is a connector. According to Anthropic, [remote custom connectors built on the Model Context Protocol (MCP)](https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities) are available on all plans, though Free users are limited to one. Each connector's directory page lists what it can read and write. If you set up a tool that connects to Search Console, check those permissions first.

If you want to pull data with a script on a regular basis and rerun the same analysis in code, [Claude Code](https://code.claude.com/docs/en/overview) fits. It runs in the terminal, in code editor extensions, in a desktop app and on the web. Most of those surfaces require a Claude subscription or an Anthropic Console account. Claude Code also reaches external data sources through MCP, which its docs describe as an open standard for connecting AI tools to outside data.

If Claude suggests titles or meta descriptions, measure their length yourself with the [character counter](https://alidemirbas.com.tr/en/calculators/character-counter).

## Preparing Search Console data for Claude

Three limits in Search Console data shape what Claude can conclude from a file. Google's help page says [exported data](https://support.google.com/webmasters/answer/12919797?hl=en) is cut to 1,000 rows of representative examples. On a small site that may be everything, but on a large site it's a small slice.

For more rows, Google's [October 2022 deep dive](https://developers.google.com/search/blog/2022/10/performance-data-deep-dive) puts the Search Analytics API limit at 50,000 rows per day, per site and search type. Google's [API guide](https://developers.google.com/webmaster-tools/v1/how-tos/all-your-data), updated in 2025, explains paging through results 25,000 rows at a time. [Bulk export to BigQuery](https://support.google.com/webmasters/answer/12918484?hl=en) gives you all performance data except anonymized queries.

The second limit is anonymized queries. Google leaves some queries out of Search Console data entirely to protect the privacy of the people searching. The same post says those queries also drop out of the chart total whenever you apply a filter. So clicks for queries that contain a word plus those that don't won't add up to the unfiltered chart total.

The third limit follows from the first two. According to Google, [chart totals](https://support.google.com/webmasters/answer/96568?hl=en) can be higher than the sum of table rows. If you give Claude only the rows, it can't see the chart total and may treat the sum of rows as the site total.

## Zeros, dates and aggregation in the file

Not every zero in your download is a real zero. Google's data page says values shown as a tilde or a dash in the report, meaning not available, become zeros in the downloaded file. If Claude treats those zeros as real values, averages and rates get pulled down.

Dates can trip you up in two ways. The Performance report records daily data in California local time, so daily numbers may not match tools like Google Analytics that use another time zone. The newest data can also be preliminary and change over the next few hours. Google says collected data is normally available within 2-3 days.

Aggregation is the third trap. Google's [Performance report page](https://support.google.com/webmasters/answer/7576553?hl=en) says chart data is aggregated by property, so two results from your site for one query count as one impression. Tables grouped by query, country, device or date are aggregated by property too, while the page table is aggregated by page. Google says gaps between chart and table totals usually come from this difference.

When you hand Claude the file, put these in your first message:

- The dates of both periods and confirmation that they're the same length
- The chart's total clicks and impressions, separate from the rows
- That the file holds at most 1,000 rows and leaves out anonymized queries
- That some zeros may mean "no data"
- That the page table and the query table shouldn't be summed into one total
- That it should use only the data in the file and show the rows behind every finding

The last item is Anthropic's own advice. Its page on [reducing hallucinations](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) recommends instructing Claude to use only information from the documents you provide, not its general knowledge. The same page suggests asking Claude to cite quotes and sources for each claim. In a Search Console analysis, that means the row with the query, page and figure behind each finding.

## Hypothetical example: measuring a drop from the rows

A hypothetical site compares two 28-day periods, with a 1,000-row export from each. Total clicks in the chart and the sum of exported rows fall at different rates:

| Period | Total clicks in the chart | Sum of rows | Row coverage |
| --- | --- | --- | --- |
| First period | 12,400 | 9,920 | 80.0% |
| Second period | 10,540 | 7,650 | 72.6% |

By the chart total, clicks fell 15.0%. If Claude only sums the rows, it finds a 22.9% drop.

The gap comes from the share of clicks the rows cover. In the first period the rows cover 80.0% of total clicks, and in the second only 72.6%. More of the second period's traffic fell outside the exported rows. That missing share could be anonymized queries or low-volume queries that didn't make the 1,000 rows.

That's why the sum of rows shows a bigger drop than the real one. As row coverage in the second period shrinks, the drop measured by rows grows, while the drop in the chart total stays at 15.0%:

| Second-period row coverage | Second-period sum of rows | Drop by rows | Drop by chart total |
| --- | --- | --- | --- |
| 80.0% | 8,432 | 15.0% | 15.0% |
| 72.6% | 7,650 | 22.9% | 15.0% |
| 65.0% | 6,851 | 30.9% | 15.0% |

When coverage holds steady across both periods, both methods agree. When coverage shifts, the sum of rows misleads. So ask Claude to calculate the total change from the chart totals and to use the rows only to find which queries and pages changed. Also ask it to report coverage for each period: if coverage falls, any reading based on the sum of rows is suspect.

The figures in both tables are hypothetical. Clicks are rounded to whole numbers and percentages to one decimal place.

## Checking Claude's results

Tying every claim back to rows in the file is the quickest way to catch Claude's mistakes. If Claude says a page lost clicks, you want to see that page's rows for both periods and work out the difference yourself. With code execution on, Claude does this math in a script, which you can read and repeat in a spreadsheet.

Anthropic's second check is to run the same prompt several times and compare the outputs. According to the page, inconsistencies between outputs could point to hallucinations. If two runs surface different queries, don't treat that finding as settled.

The same page says these techniques significantly reduce hallucinations but don't eliminate them entirely. So always open the figure you'll base a decision on in Search Console too.

A cause Claude suggests, such as "the title change lowered click-through rate," is only a candidate. The file alone can't prove it, since it only shows when and on which page a change began. To check it, filter Search Console by page and date and look for other changes in the same period.

Cannibalization candidates work the same way. When Claude sees more than one page for the same query, it lists them as candidates. You decide which page stays by looking at what each page is for.

## Uploading client data to Claude

Your plan and privacy setting decide whether the data you upload can be used for model training. Anthropic's [model training page](https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training), dated March 2026, covers consumer plans such as Free, Pro and Max. On those plans, your chats are used to improve Claude only if you allow it.

Incognito chats aren't used even when that setting is on. Commercial products such as Claude for Work and the Anthropic API have a separate page.

This part is an inference, not something the page says. Before you upload another company's Search Console data from a personal account, check your setting and agree with the client on which tool you'll use. The query list shows which products and topics the client is visible for, so it's worth discussing before you share it.

## How to set up your first analysis this week

Keep your first Claude SEO analysis small enough to check yourself:

1. In Search Console, pick two equal periods with a custom date range and leave out the newest preliminary days. Export the query and page tables separately and note the chart's total clicks and impressions.
2. Upload the files to Claude and state the dates, totals and file limits in your first message.
3. Ask Claude for the five queries and pages that changed most, each with its rows.
4. Work out one finding's figure yourself and open it in Search Console.

If the numbers match, move the setup into a project and repeat it every week. If they don't, ask Claude where the gap comes from, starting with row coverage.