AIO Copilot

How to Use Claude Code for SEO: Seven Real Jobs, What Each Costs, and Where a Person Must Check

Anthropic's coding agent can audit a crawl or check a negative keyword list for a dollar or two. It does that best when it writes the code that does the counting, and worst when it is trusted to count.

Updated 3 November 202615 min read

How to use Claude Code for SEO: the short answer

Claude Code is Anthropic's coding agent. It runs in a terminal or a desktop app, reads the files in a folder, and can run commands and edit code there. To use it for SEO, put a client's exports in one folder (a crawl, a Search Console download, a Google Ads search terms report), describe the job, and have Claude Code write and run the scripts that answer it. Point it at the site's repository and it can change templates, structured data and internal links, with you approving each diff. It comes with Claude Pro at $20 a month, Max at $100, Ultra at $200, or pay-per-token API billing.

Anthropic's cost documentation puts the average at about $13 per developer per active day across enterprise deployments, with 90% of users under $30. The bill is rarely what goes wrong. What goes wrong is a model that sums a column of an Ads export in its head, counts the report's own Total row as another search term, and reports the doubled figure in the same assured prose it uses for a correct one.

This guide covers what Claude Code costs now that Anthropic has replaced all three of its model tiers since 22 September, then works through seven SEO jobs on a sample client, Harbour & Finch LLP, a Toronto law firm that exists only as demo data. The argument running through it is that Claude Code should write the code that does the counting, while a person reads the result and keeps the apply button.

What is Claude Code, and what does it cost in November 2026?

Claude Code began as a command-line tool and still works best there, though Anthropic also ships a desktop app and editor extensions. Its setup documentation says it needs a Pro, Max, Ultra, Team, Enterprise or Console (API) account. Prices below come from Anthropic's pricing page and its help articles on the Max and Ultra plans, as of November 2026.

PlanPriceWhat it means for SEO work
Pro$20 a month, or $17 a month billed annuallyEnough to learn on and run occasional audits
Max$100 a monthFive times Pro's usage, every five hours and every week
Ultra (formerly Max 20x)$200 a monthTen times Pro's weekly usage; no five-hour limit once Anthropic verifies the account
TeamStandard seat $25 a month ($20 billed annually); Premium seat $125 ($100 annually)Shared administration and commercial data terms
API keyPer token, at list priceNo allowance to run out of; every session is billed

Every paid plan and the API start Claude Code on Claude Opus 5.5, according to its model configuration docs. Anthropic released Opus 5.5 on 22 September at $4 per million input tokens and $20 per million output tokens. Sonnet 5.5 followed on 28 September at $2 and $10; Anthropic says that in its testing it costs up to 30% less per task than Sonnet 5. Haiku 5.5 arrived on 7 October at $0.10 and $0.50 for prompts up to 100,000 tokens, rising to $0.50 and $2.50 above that.

In Claude Code the aliases opus, sonnet and haiku point at those three when you sign in with an Anthropic account; switch with /model sonnet mid-session or claude --model haiku at launch. On Amazon Bedrock and Google Cloud, sonnet still resolves to Sonnet 4.5, so pin full model IDs there. All three default to medium effort, and the cost page notes that thinking cannot be switched off on any of them and is billed as output. Our analysis of what Claude Opus 5.5 costs on SEO and Google Ads jobs works through effort settings on a 500-page audit, and our comparison of Claude Code with Codex, Cursor and Antigravity covers the alternatives.

One line in Anthropic's terms matters more to agencies than any price. Under its August 2025 consumer terms update, Claude Code sessions on Free, Pro and Max accounts can be used for training if the user allows it, with data kept for five years in that case and 30 days otherwise. Team, Enterprise and API use fall under commercial terms the change does not cover. Before client data goes through Claude Code, check that setting or use a Team seat or an API key.

How do you set up Claude Code for SEO work?

Installation is one command. On macOS, Linux or WSL:

curl -fsSL https://claude.ai/install.sh | bash

In Windows PowerShell:

irm https://claude.ai/install.ps1 | iex

Then open a terminal in the folder you want it to work in and run claude; the first run walks you through signing in from the browser. On native Windows, the setup docs recommend installing Git for Windows too, so Claude Code can use Bash rather than falling back to PowerShell.

The arrangement that pays off is one folder per client, with exports kept apart from anything Claude Code writes:

harbour-finch/            # sample client, demo data
  CLAUDE.md
  exports/
    crawl/internal_all.csv
    gsc/query_page_2026-09.csv
    ads/search_terms_2026-09.csv
  scripts/
  reports/

CLAUDE.md is the file Claude Code reads at the start of every session. Anthropic's memory documentation says it treats the file "as context, not enforced configuration" and recommends keeping it under 200 lines. For SEO, the most useful lines are rules about data, because those cover the mistakes a model makes when nobody tells it otherwise:

# Harbour & Finch LLP (sample client, demo data)
Canonical host: https://www.harbourfinch.example (https, www, no trailing slash)
Practice areas: employment law, wills and estates, residential real estate

- Never edit files in exports/. Write outputs to reports/.
- Compute every number with a script in scripts/ and print the command used.
- Blank, "-" and "--" cells mean missing, not zero.
- Drop rows whose first cell starts with "Total" before summing an Ads export.
- Do not state search volume, difficulty or rankings unless a file here has them.

Before a large job, press Shift+Tab to cycle into plan mode, so Claude Code proposes an approach before touching anything. Run /clear between unrelated tasks, because the whole conversation is resent with every step, and /usage to see what a session has cost.

How to run a technical SEO audit from a crawl export

Start with the crawl, because every other job needs a list of URLs and their status codes. Screaming Frog's SEO Spider is free up to 500 URLs and £199 a year beyond that, says its pricing page, and its user guide documents a headless mode Claude Code can run:

screamingfrogseospider --crawl https://www.harbourfinch.example --headless \
  --save-crawl --output-folder exports/crawl \
  --export-tabs "Internal:All,Response Codes:Client Error (4xx)" --export-format csv

That is the Linux command; on Windows the executable is ScreamingFrogSEOSpiderCli.exe. Then give Claude Code the job, including how you want it done:

Read only the header row of exports/crawl/internal_all.csv and list the
columns you will use. Then write scripts/audit_crawl.py to find, for
indexable HTML pages: missing or duplicate titles, canonicals pointing to
a different URL, internal links to URLs returning 3xx or 4xx, and redirect
chains longer than one hop. Run it, save reports/crawl_audit.csv, and
report counts by issue with the command you ran.

The first instruction matters most. The sample crawl of about 1,800 URLs has dozens of columns per row, and reading it whole fills the context and the bill. Asking for the header first makes Claude Code learn the columns and write a script that treats every row the same way, and that you can rerun next month.

On the sample data, the script finds 14 practice-area pages whose canonical points to a trailing-slash URL that redirects back to the page itself. Claude Code can fix the template that builds the canonical, and flatten redirect chains in a file such as next.config.ts or .htaccess so every old URL points straight at its final destination. Both changes should arrive as a diff on a branch that a person reads before anything deploys. Our technical SEO audit checklist covers what else a full audit includes, and the free SEO checkers help spot-check single pages once a fix is live.

How to analyze a Search Console export: striking distance and cannibalization

Two analyses repay the effort: queries where a page sits just off page one, and queries where two of the site's own pages split the impressions. Both depend on the right export, which is where most attempts go wrong.

The export button in Search Console stops at 1,000 rows, Google says in its guide to performance data limits, and each tab arrives as its own summary. Short of exporting one page at a time, that file cannot tell you which page earned which query, and cannibalization is exactly that question. The Search Analytics API returns query and page together, up to 25,000 rows a request and 50,000 a day per site and search type, according to Google's API documentation. The bulk export to BigQuery lifts that cap but still omits anonymized queries, those not issued by more than a few dozen users over two to three months. The interface counts them in chart totals but not in rows, so no export's rows add up to the headline figure.

With a query-and-page file in exports/gsc/, the logic for both analyses fits in a few lines, which is the point: you can read it and check it.

import pandas as pd

# columns: query, page, clicks, impressions, position
df = pd.read_csv("exports/gsc/query_page_2026-09.csv")

# Striking distance: average position 8 to 20, with enough impressions to matter
striking = df[df.position.between(8, 20) & (df.impressions >= 100)]

# Cannibalization: queries where two or more pages each take 20% or more of impressions
share = df.impressions / df.groupby("query").impressions.transform("sum")
split = df[share >= 0.2].groupby("query").filter(lambda g: len(g) >= 2)

The thresholds are choices, so tell Claude Code to state them in every report. In the sample data, "wrongful dismissal lawyer toronto" shows the pattern: the employment law hub and an older blog post on severance split the impressions at an average position near 11, and neither earns many clicks. Claude Code handles the next step well, reading both pages and proposing which should own the query and what the other becomes: merged into the hub with a redirect, or rewritten for a different query. Our guide to building topic clusters around pillar pages covers choosing the owner. The redirect decision stays with a person, because it is the one that can lose traffic.

Averages need care. Position is an average across impressions, so a page at 11 may sit at 4 for some searches and far lower for others. Compute click-through rate from summed clicks and summed impressions, never by averaging the CTR column, which is exactly what a model asked for "the average CTR" will often do.

Can Claude Code write content briefs and refresh old pages?

Yes, with one rule: build the brief from evidence in the folder, not from the model's sense of what people search for. Prompt lists written for chat assistants routinely asked for "search volume estimates" and "keyword difficulty". A language model has neither, so asking produces plausible numbers with nothing behind them.

A brief Claude Code can stand behind uses the queries the page already gets impressions for, the site's other pages on the topic (so the new page does not compete with them), and the internal links it should receive. For a refresh, give it the page's source file and its queries, and ask for proposed changes with a reason for each before any edit. On the sample site, the wills guide gets impressions for questions about cost that it never answers, so the brief adds that section and a link to the fees page.

Drafting is where Google's own guidance applies. Its documentation on generative AI content, last updated on 1 October 2026, says AI can help with research and structure, that it is "critical to manually factcheck and review all AI-generated content for accuracy", and that generating many pages without adding value for users may violate its spam policy on scaled content abuse. Claude Code can draft a refresh in minutes. The fact-check takes as long as it always did, and it is the part that protects the client.

How to generate and validate schema markup with Claude Code

Structured data is where access to the codebase matters most, because the fix is almost always in a template. Ask Claude Code to build JSON-LD from fields the template already renders, such as the H1, the CMS dates and the byline, so the markup cannot drift from what visitors see.

Google's rules make that drift the main risk. Its structured data guidelines say "Don't mark up content that is not visible to readers of the page" and that markup "must be a true representation of the page content". They also say valid markup does not guarantee a rich result. One change dates a lot of older advice: Google stopped showing FAQ rich results in Search on 7 May 2026. FAQPage remains a valid schema.org type, but a refresh that adds it to win expandable questions in the results is chasing a feature that is gone.

Validation works best in two layers. First, a check Claude Code writes and runs against the built pages:

Write scripts/check_schema.py. For every HTML file in the build output,
parse each application/ld+json block. Fail if a block is invalid JSON,
if Article.headline differs from the page's H1, or if dateModified is
earlier than datePublished. Print the file and the reason for each failure.

Second, put a sample of templates through Google's Rich Results Test and the Schema Markup Validator, which catch problems the script was never told to look for.

How to turn an SEO workflow into a reusable Claude Code skill

Once a job works, package it so it runs the same way next month and for the next client. A skill is a folder with a SKILL.md file, kept in .claude/skills/ inside a project or ~/.claude/skills/ for all of them. Anthropic's skills documentation says Claude Code uses a skill when a request matches its description or when you type its name after a slash, and that skills "can bundle and run scripts in any language". That suits SEO work: the instructions say when and how, and the script does the counting.

.claude/skills/striking-distance/
  SKILL.md
  scripts/striking_distance.py

---
name: striking-distance
description: Find Search Console queries at average position 8 to 20 and the page that should own each. Use for striking distance, page-two or quick-win requests.
---
1. Run the bundled scripts/striking_distance.py on the newest file in exports/gsc/.
2. State the thresholds used and the row counts before and after filtering.
3. For the top 20 queries by impressions, read the ranking page and propose one change.
4. Never estimate search volume. Write results to reports/striking_distance.md.

Type /striking-distance in a session. To run it unattended, for example from a weekly scheduled job, the non-interactive mode accepts a skill name in the prompt, and --allowedTools limits what it may do without asking:

claude -p "/striking-distance" \
  --allowedTools "Read,Write,Bash(python .claude/skills/striking-distance/scripts/striking_distance.py *)"

Our guide to Claude skills for SEO automation goes further, with skills for keyword clustering, content audits and schema, and how to chain them into one workflow.

What does a Claude Code SEO job cost?

On Pro, Max or Ultra there is no per-job bill. Work draws down an allowance with a five-hour limit and a weekly one (verified Ultra accounts have only the weekly one), published as multiples of Pro rather than message counts, so you learn your ceiling by reaching it. On an API key every session is billed per token.

Claude Code resends the conversation at every step, so most of a session's input is cache reads, which Anthropic's pricing page puts at 5% of the normal input price on Opus 5.5 and Sonnet 5.5. Here is list-price arithmetic for one audit-sized session; the token counts are illustrative assumptions, not measurements.

Line itemTokens (assumed)Opus 5.5Sonnet 5.5
Uncached input60,000$0.24$0.12
Five-minute cache writes60,000$0.30$0.15
Cache reads1,200,000$0.24$0.12
Output, including thinking40,000$0.80$0.40
Total$1.58$0.79
In an audit-sized session, output is the largest cost on either model, though it is the smallest token count.
Uncached input (60K)$0.12$0.24
Cache writes (60K)$0.15$0.30
Cache reads (1.2M)$0.12$0.24
Output and thinking (40K)$0.40$0.80
Sonnet 5.5Opus 5.5
Cost of one assumed audit-sized Claude Code session by line item, Sonnet 5.5 against Opus 5.5, US dollars
RowSonnet 5.5Opus 5.5
Uncached input (60K)$0.12$0.24
Cache writes (60K)$0.15$0.30
Cache reads (1.2M)$0.12$0.24
Output and thinking (40K)$0.40$0.80
Worked example with assumed token counts, not a measured session. List prices per million tokens from Anthropic's pricing page, as of November 2026: Opus 5.5 $4 input, $5 five-minute cache write, $0.20 cache read, $20 output; Sonnet 5.5 $2, $2.50, $0.10 and $10.

Output is the biggest line despite the smallest count, because it costs five times the input price and thinking counts as output. That makes effort the setting that moves a bill most: leave routine jobs at medium or low. The other common way to inflate a session is reading a raw export into the conversation instead of scripting it, since every later step resends it.

High-volume labelling is cheaper still. Tagging 20,000 search terms by intent on Haiku 5.5 through the API, with 400,000 input tokens of terms and instructions and 200,000 output tokens of labels, costs about 14 cents at list price before any thinking tokens. The same job costs $2.80 on Sonnet 5.5 and $5.60 on Opus 5.5. At Haiku's price the expensive part is checking a few hundred labels by hand, which you should do before trusting the rest.

Labelling 20,000 search terms costs 14 cents on Haiku 5.5 and $5.60 on Opus 5.5, before thinking tokens.
Haiku 5.5$0.14
Sonnet 5.5$2.80
Opus 5.5$5.60
List-price cost of labelling 20,000 search terms, US dollars, by Claude model, before thinking tokens
Haiku 5.5$0.14
Sonnet 5.5$2.80
Opus 5.5$5.60
Worked example: 400,000 input tokens and 200,000 output tokens, sent in requests under 100,000 tokens each, at list prices from Anthropic's pricing page, November 2026 (Haiku 5.5 $0.10 and $0.50, Sonnet 5.5 $2 and $10, Opus 5.5 $4 and $20 per million tokens). Thinking cannot be switched off on any of the three, per Claude Code's cost documentation, so real totals are higher.

What Claude Code should not be trusted with

Arithmetic it did not run comes first. Any total, rate or before-and-after figure should come from a script whose code you can read. If a number appears in a report with no command behind it, send it back.

Data it cannot see comes next. Claude Code has no search volumes or rankings unless a file or a connection provides them, and asked for those numbers it will supply confident estimates. Live connections exist: MCP servers, added with claude mcp add --transport http <name> <url>, or a script with read-only Search Console API credentials. Anthropic's MCP documentation warns that servers which fetch external content can expose you to prompt injection. For client work, exports in a folder are slower but far easier to audit, because you know exactly what the model saw.

Changes to anything live belong behind a person. Template edits, redirects, robots rules and ad account changes should go through a branch and a diff that someone reviews. A wrong canonical or an overbroad negative can do damage that takes weeks to show, which is why Claude Code's permission prompts are worth keeping on for client work.

The last is a claim that the work worked. Claude Code will compare clicks before and after a change and call the difference a result, but seasonality and Google's updates move those numbers too. The honest comparison measures changed pages against pages nobody touched over the same weeks. That is how AIO Copilot reports every SEO change, and the free Did it work? check runs the same test on two Search Console exports. Our piece on proving an SEO change worked with a control group explains the method.

Used this way, Claude Code changes who does the tedious part of SEO far more than it changes the work. The export still has to be the right one, and the redirect still needs someone willing to put their name to it. What goes away is the afternoon of lookups and pivot tables, which was never the part of the job that needed a person.

Questions

Is Claude Code free?

No. Claude Code needs a Pro, Max, Ultra, Team, Enterprise or Claude Console (API) account, and the free Claude plan does not include it. Pro costs $20 a month, or $17 a month billed annually. Max costs $100 a month for five times Pro's usage, and Ultra, which replaced Max 20x, costs $200 for ten times Pro's weekly usage with no five-hour limit once the account is verified. On an API key you pay per token instead.

Which Claude model should I use for SEO work in Claude Code?

Claude Code starts on Claude Opus 5.5 on every paid plan and the API. Keep it for multi-step diagnosis and site-wide code changes. Switch with /model sonnet for routine analysis on Sonnet 5.5, which costs half as much per token, and use Haiku 5.5 for high-volume labelling. All three default to medium effort, which you change with /effort.

Can Claude Code connect to Google Search Console?

Only through a connection you set up, such as an MCP server or a script that calls the Search Console API with your credentials. Otherwise it works from exports you put in its folder. The export button in Search Console stops at 1,000 rows and does not pair queries with pages; the API returns query and page together, up to 50,000 rows a day per site and search type.

Can Claude Code do keyword research?

It can cluster and label keywords you give it, and find queries in your Search Console data that no page on the site targets. It has no search volume, difficulty or ranking data of its own. If you ask it for those numbers it will produce plausible estimates with no source, so feed it exports instead.

Is it safe to put client data through Claude Code?

That depends on the account. Since Anthropic's August 2025 terms update, sessions on Free, Pro and Max accounts, including Claude Code, can be used for model training if the user allows it, with five-year retention in that case and 30 days otherwise. Team, Enterprise and API use are under commercial terms that the change does not cover. Agencies should check the privacy setting or run client work on a Team seat or an API key.

Do I need to know how to code to use Claude Code for SEO?

Not to start. You describe the job in plain English and approve the commands it wants to run, and a desktop app spares you the terminal. You do need to read a short script well enough to see what it counts, and to read a diff before anything reaches a live site.

New studies, by email

Our own studies, and what each new release changes for SEO, AEO and Google Ads.

Or get your price for the app