Skip to content

ChatGPT vs Claude for SEO Work: Five Real Tasks, Side by Side

Same five SEO jobs, same files, same prompt: an audit, keyword intent, a brief, internal links and a fact-check. Claude Code found all 20 planted audit issues and Codex 16. On the brief, with no answer key, each grader preferred its own maker's work.

ChatGPT vs Claude for SEO work, 5 real jobs: a terminal scorecard with audit 20 vs 16, intent 43 vs 53, links 16 vs 6 (graded by Opus) and fact-check 8 vs 8
Table of Contents13 sections

Get my next blog post first

Real tests on Claude and SEO tools, with the numbers and what failed. About once a week.

Unsubscribe anytime. Privacy

I use both every day. Claude Code builds and audits my sites, and Codex reviews the plans and drafts. So instead of comparing them on vibes, I gave them the same five SEO jobs.

Same files, same prompt, one run each, and a score sheet written before either tool ran.

The short version: Claude Code did the work better, and I still want Codex checking it.

Key Takeaways

  • The audit split them most: Claude Code found all 20 planted issues, Codex 16, and Codex called a "Page not found" page a landing page.
  • Both caught all 8 numbers I'd broken in a fact-check. Codex also flagged 4 correct rows because it counted a different way.
  • On keyword intent Codex had more exact matches with the key, 53 of 55 against 43.
  • On the content brief, with no answer key, each grader preferred its own maker's work. With a key they mostly agreed.
  • One run per task, on my own work. Treat it as a field test.

Quick answer: which one I use for what

Claude Code gets the jobs where a miss costs you money: audits and internal links. In this test it found every planted audit issue and wrote the link list both graders liked more.

Codex gets the second read. It matched my keyword intent key more closely, and in October it caught an audit miss Claude had made.

So whichever tool does the work, the other one checks it.

What I compared, and what I didn't

I tested the coding tools, the ones that run in a terminal. Both read files, run commands and write their output to a file.

That's how I actually do SEO work with them (my Claude Code settings are a post of their own).

Claude CodeCodex CLI
MakerAnthropicOpenAI
Version2.1.3110.160.1
ModelClaude Opus 5.5gpt-6-luna, reasoning effort high
Date5 November 20265 November 2026
Could useRead and write files, run curl and PythonRead and write files in its folder, network on
Couldn't useWeb searchWeb search

Each run started in a fresh, empty folder with only that task's files copied in. Same prompt, word for word. Neither tool could see the other's output or the answer keys.

What this doesn't cover: the ChatGPT and Claude chat apps, writing quality as a ranking factor, and anything that needs more than one run to call.

The five jobs

#JobInputHow it was scored
1Technical SEO auditA test site with 20 planted issues, URL onlyAnswer key written before the site was built
2Keyword intent55 real keywords from my researchIntent key written before the runs
3Content briefSix pages ranking for "how to get cited by chatgpt"Every attributed claim checked against the pages
4Internal linksNine of my posts, one not yet publishedSentence found word for word, target real, not already linked, then judged
5Fact-checkA copy of my Perplexity post with 8 numbers changed, plus its dataThe 8 planted errors

Each output was graded twice, blind: once by Claude Opus and once by Codex, with the names swapped for A and B. Where the two graders disagreed, I say so.

The scorecard

#JobClaude CodeCodex
1Audit, 20 planted issues20 of 20 (both graders)16 of 20 (Opus), 16.5 (Codex)
2Keyword intent, 55 keywords43 exact, 54 with the key's second choice53 exact, 53 with the second choice
3Content briefOpus: 113 of 115 claims right. Codex: 99 of 111Opus: 78 of 82. Codex: 55 of 55
4Internal links, 12 each12 of 12 valid. Judged 16 of 24 (Opus), 21 of 24 (Codex)12 of 12 valid. Judged 6 of 24 (Opus), 17 of 24 (Codex)
5Fact-check, 8 planted errors8 of 8, plus 3 fair extra flags8 of 8, plus 4 wrong flags
Time, all five13 minutes15 minutes

Job 1: the audit

This is where they split the most. Same test site, same prompt: "do a full SEO audit", a URL and curl, nothing else.

Claude Code found all 20 planted issues, including the hard ones: a noindex sent only in an HTTP header, and a page that says "Page not found" while returning a 200.

Codex missed three and half-found a fourth:

  1. The header-only noindex on /subscriptions/.
  2. The soft 404 on /gifts/.
  3. A 302 on a permanent move. It saw the redirect but not that it should be a 301 (Opus gave no credit, Codex half).
  4. A near-duplicate 1kg product page.

Codex called /gifts/ "a reachable, self-canonical landing page" and suggested you add it to the sitemap.

It also got one redirect backwards.

Opus counted four wrong findings in Codex's audit, the Codex grader one. Claude Code had one that both graders agreed on: that the home page links to one product twice. It links once.

(The Codex grader found two more in Claude's audit, both arguable.)

Job 2: keyword intent

Codex matched the key on 53 of 55 keywords, Claude Code on 43.

Most of Claude's misses were one call: "digital marketing for lawyers" or "seo for dentists" as commercial, where the key says informational.

The key allows a second answer on those. Count it, and Claude reaches 54, and the gap nearly closes. One run can't tell you whose judgement is better.

Claude helped me draft the intent key, before either run. Codex still matched it better. Read this job as the softest of the five.

What about clustering? Claude made 13 topic groups. Codex made 24, a few with one keyword each.

Job 3: the content brief

Both briefs were mostly accurate on numbers, and the statistics the graders checked traced back to the right page.

The graders disagreed on the rest, and each one preferred the brief its own maker wrote: Opus rated Claude's 5 for usability and Codex's 4, and Codex rated its own 5 and Claude's 3.

So I checked the hardest claim myself. Codex's brief named one page's publisher "Searchbloom", four times. That name isn't anywhere in the six pages.

Both lists passed every mechanical check, all 12 suggestions each:

  • the sentence really is in the post, word for word
  • the target post and its #section exist
  • that post doesn't already link there

Those checks don't catch everything, though. Opus found one Codex suggestion whose anchor words already link to a different post.

So the graders judged whether a reader of that sentence needs that page. Both preferred Claude's list, Opus by a lot (16 against 6 out of 24) and Codex by a little (21 against 17).

Claude also noticed one of my nine posts wasn't published yet, and left it out on its own. I hadn't planned that test. It was my mistake in the input.

Job 5: the fact-check

I took my Perplexity post and changed 8 numbers: 210 sources instead of 200, 96 sites instead of 116. Then I handed over the real data.

Both caught all 8. They differed on the extras:

Claude CodeCodex
Planted errors found8 of 88 of 8
Fair extra flags3: a figure from another test, a sentence the planted error made untrue, page types that exist for 4 more questions2: the same figure, and a claim the data can't check
Wrong flags04 correct table rows, counted by unique page instead of by row

Codex's four wrong flags came from counting each page once, where my table counts every row. You'd only spot that if you know how the data was built.

The failure that changed how I audit

The /gifts/ page has tripped up audits before. In October I tested five Claude SEO skills on this same site.

One audit said every page it checked had a meta description and a canonical.

/gifts/ had neither. Codex, grading that run, caught it.

So today Claude caught what Codex missed, and in October it was the other way round. Would I trust either one alone with a client's site? No.

I run an audit with one, have the other grade it against the HTML, and open the pages they disagree on myself.

How to test them on your own work

  1. Pick a job you already know the answer to. A site you audited, a post you fact-checked.
  2. Write the answer key first, before either tool runs.
  3. Give both the same folder and the same prompt, word for word.
  4. Score against the key, and count the wrong claims as well as the hits.
  5. Run it again next month. Both tools change often.

Common Questions

Is Claude or ChatGPT better for SEO work?

In my five-job test on 5 November 2026, neither won everything. Claude Code did better on the technical audit (20 of 20 planted issues against 16) and on internal links. OpenAI's Codex did better on keyword intent (53 of 55 against 43). Both caught all 8 planted errors in a fact-check.

Can Claude Code run a technical SEO audit?

Yes. Given only a URL and curl, Claude Code found all 20 issues planted on a test site, including a noindex sent in an HTTP header and a soft 404. It also made one wrong claim, so check its findings against the page HTML.

Is Codex the same as ChatGPT?

No. Codex CLI is OpenAI's coding tool: it runs in a terminal, reads and writes files and runs commands. It uses OpenAI's GPT models, like ChatGPT, but this test didn't cover the ChatGPT chat app.

Can one AI fairly grade another AI's SEO work?

It went better with an answer key. On the audit and the fact-check, both scored against keys, Claude Opus and Codex mostly agreed. On the content brief, with no key, each grader preferred the brief its own maker wrote.

Want my next blog post before anyone else?

I run real tests on Claude and SEO tools, like 10 voice models on one Mac or 25 audits of a site with 20 planted problems. You get the numbers, including what failed. One email per post, about once a week.

Unsubscribe anytime. Privacy

Hiring for AI search? Let’s talk.

Email me the job description, or book thirty minutes now. You’ll hear back within one working day.

Put your email and confirm it from your inbox. The file comes straight after.

You'll also get new posts, about once a week. Unsubscribe anytime. Privacy