Skip to content
Doctor SEO
Free SEO audit

AI for SEO: Using the Models as Tools

What Claude Can and Cannot Do for SEO

A thirteen-task scope table for Claude in SEO: what it does with nothing connected, what has to be plugged in before it works on real data, and what stays human. With the documented failures, named and dated.

Updated 29 September 2026 19 min read

By the end of this lesson you will have a scope table for Claude: task by task, whether it does the job with nothing connected, what you must plug in first, and which part stays yours. The question that brings people here is whether Claude can do their SEO. The answer is neither yes nor no: it turns on whether the data the task needs sits inside the conversation or outside it.

The demand data says something odd here. In the Google Autocomplete harvest run for this course on 20 September 2026, three of the tasks most written about return nothing: the seeds claude schema markup, claude internal linking and claude api seo each echo the query back with no variants. The seed claude seo audit returns ten suggestions, six asking for a skill, a repository or a tool.

Every figure below carries a date, because connector inventories, tool counts and context windows have a shelf life of months here. Anything older than a quarter is a lead to re-check, not a fact to quote.

What you’ll learn

  • Tell apart the three products people call “Claude”, and what each is for in SEO.
  • Use a thirteen-task scope table to see which jobs need real data connected first.
  • Apply a data-location test to any task the table does not list.
  • Say what was measured when Claude and ChatGPT were compared on SEO tasks.
  • Size a crawl export against the one-million-token window by measuring, not guessing.
  • Name the documented failures, with author, date and the limits of each case.

What Claude sees when nothing is connected

With nothing connected, Claude sees only what is in the conversation: your text, your attachments, and whatever the tools you have granted it return. It does not crawl your website, it does not hold search volume data, and it cannot open your Search Console. Lawrence Hitches put the boundary in one sentence on 2 April 2026, in Claude for Technical SEO: Audits, Crawl Fixes & Schema: Claude analyses crawl data, it does not crawl sites.

That line matters because the failure is silent. Ask for an analysis of your site with no export attached, no connector enabled and no web access, and you will get an analysis: well organised, using your brand name, assembled from what a model has read about sites like yours. Nothing in the reply announces that no file was read. So the first habit is mechanical. Before accepting any output that contains a fact about your site, name the artefact it came from.

Three surfaces share one name, and they do not share settings

Three products get called Claude, and the difference decides which SEO jobs are possible at all. claude.ai is the chat interface, with Projects, Artifacts and MCP connectors. Claude Code is the terminal agent: it works on files, runs scripts and preprocesses large exports. The Claude API is the programmatic route, for an operation that repeats over thousands of rows.

They do not share configuration. Anthropic’s platform documentation, checked on 20 September 2026, states that custom Skills do not sync between the API, claude.ai and Claude Code, and that Skills on claude.ai are per user rather than organisation-wide. Package your audit as a Skill in the chat and you do not have it in the terminal. One number sets expectations: the official anthropics/skills repository held about 168,900 stars on 20 September 2026 and contains no SEO skill. What exists is community work such as AgriciDaniel/claude-seo at about 16,700, none of it independently evaluated in the research behind this course.

The scope table: thirteen SEO tasks and where the limit falls

This table is the centre of the lesson. Each row answers three questions: does Claude do the task unaided, what has to be connected before it works on real data, and what stays human whatever you connect.

Task Unaided? What it needs connected What stays human
Keyword research with search volumes No. It answers anyway A real source: DataForSEO, Ahrefs or Semrush Deciding whether the term is worth the work
Clustering a keyword list you supply Yes Nothing; you paste the file Deciding whether the grouping matches real intent. Measuring that is its own lesson
Crawling your site No, with no nuance available A crawler: Screaming Frog shipped a native MCP server in v24.0, 19 May 2026, or an export Deciding the crawl is trustworthy before you read anything into it
Analysing crawl data Yes Nothing, while the file fits The order of the fix list: prioritisation is a business call
Reading your Search Console No A community MCP server; no official connector exists (Passionfruit, 28 March 2026) The interpretation. A drop in clicks does not say why
Server log analysis Yes, on the file you hand over Nothing, or preprocessing for millions of lines Knowing which questions a log cannot answer
Generating JSON-LD schema Yes, and among the better jobs Nothing Validation in the rich results test
Proposing a redirect map Partly, through embeddings Both URL lists A human review of the mapping before it ships; validation is a lesson of its own
Planning internal linking Yes, the plan Your URL inventory; executing it needs the CMS Approving the plan dry before anything is written
Writing to your CMS Yes, technically A WordPress MCP server; Rank Math published an official one, 27 July 2026 All of it. The worst documented outcome in this level is here
Drafting briefs and copy Yes Nothing Acceptance sampling per batch, serious defect defined in advance
Outreach and link building It produces the text Nothing All of it. One of the tasks practitioners most often decline to automate
Being cited by AI search engines Not this lesson’s job Nothing here Worked separately, in GEO and AIO

Read it by columns. The second says no wherever the data lives outside the conversation: volumes, crawling, Search Console. The fourth says all of it wherever a mistake cannot be undone in a minute, meaning the CMS and anything reaching a person. Everything between is delegable with a check on top.

Two rows deserve their numbers. Keyword.com’s State of AI and Automation in SEO, 1 January 2026, n=97, found 38% of respondents using AI for technical SEO audits; among the 79% who could automate a task and chose not to, 40% named technical SEO and 51% named link building. Small, self-selected sample: an ordering of tasks, not a measurement of the profession.

The data-location test, for any task the table does not list

For an SEO task not in the scope table, run one test before delegating it: name the artefact carrying the data the task needs, and name who puts that artefact into the conversation. It catches most bad delegations, because the failure it targets is not the model being wrong but the model being fluent about something it was never given. There are three outcomes and no fourth.

  1. Already in the conversation. The data is in a file you can attach or text you can paste, and the task is a transformation of it: clustering, summarising, restructuring, drafting. Delegate it, and design the check before the first batch.
  2. Connectable. The data exists somewhere real and a connector, an export or a tool call can bring it in: your crawler, your analytics, a keyword API, your CMS. Delegable once that link exists and not before, because until then it is the first category in disguise.
  3. Not available at all. Nobody can supply the artefact, because it does not exist in readable form: what a competitor plans next quarter, what a search engine will do after its next update. The task is out.

One verdict the test never returns: hand it over and walk away. Keyword.com’s survey found exactly 1 respondent in 97 describing any workflow as fully automated (1 January 2026, n=97, with 56% of respondents in teams of one to five people). A test like this decides how much of a task leaves your hands, not whether you leave the room.

Claude or ChatGPT for SEO: what was measured and what was not

One published comparison runs Claude and ChatGPT against concrete SEO tasks, and knowing how little it proves is as useful as knowing the result. SE Ranking’s Claude AI vs. ChatGPT: Which AI Tool is Better for SEO?, published on 7 July 2025 by Yulia Deda and updated 24 June 2026, ran four tasks and reported three outcomes. On JSON-LD schema, Claude’s markup returned 1 non-critical issue in the rich results test against 5 for ChatGPT’s. On content briefs the comparison also favoured Claude. On blog titles ChatGPT came out ahead.

The methodology is the part nobody quotes and the part that matters. That comparison is qualitative: no accuracy metric, no sample size, each of the four tasks run once. Presenting those outcomes as a verdict on two products over-reads a single pass through four prompts.

The difference that decides most cases is architectural. ChatGPT’s own thing is the custom GPT with Actions: Paul Shapiro documented one on 28 November 2023, updated 27 December 2023, that queries real search volumes through DataForSEO behind an OpenAPI schema, and it cannot be rewritten for another engine. One reader told him the schema section bewildered them, a fair description of the entry cost. Claude’s own thing is the MCP connector ecosystem pointing at SEO tools: the Ahrefs connector was Anthropic-verified with 61 tools on 20 September 2026, Semrush’s MCP endpoint was last updated 5 August 2026, and Rank Math shipped its own on 27 July 2026. The question is not which writes better prose, but which reaches your data.

Artifacts and the API: an output format and a scale route

An Artifact is a document or small application Claude builds inside the conversation, which you view, manipulate and download instead of reading as chat text. It is a view of data you supplied and adds no information, so a dashboard built on invented volumes is invented volumes with colours on. The published evidence on Artifacts in SEO is thin: the only comparison located is SE Ranking’s, 7 July 2025, where Claude’s year-on-year report carried line graphs, pie charts and interactive elements against ChatGPT’s mostly text, qualitative and with no sample size.

The Claude API is the other end of the same question, and it earns its complexity when one operation repeats across thousands of rows with nobody reading the results one at a time: rewriting metadata for a catalogue, classifying intent, extracting entities. Three pieces matter there, all verified in Anthropic’s platform documentation on 20 September 2026. The Batch API processes asynchronous batches at 50% of the standard cost, prompt caching cuts the cost of resending the same long context, and the Models API returns max_input_tokens per model, the only honest way to know how much fits. Two warnings attach. The agent workflows published for this kind of work ship a set of write-safety controls around every live change (Lawrence Hitches, StudioHawk, 12 May 2026), and those controls are a subject of their own later in this level. And Skills used through the API have no network access, so a Skill meant to look something up live fails there.

The one-million-token window: measure it, do not estimate it

The context window is how much text Claude can hold in front of it at once, and it invites people to paste an entire crawl and start asking questions. Anthropic’s model documentation, checked on 20 September 2026, gives these limits.

Model Input context Maximum output
Opus 5 1M tokens 128K tokens
Sonnet 5 1M tokens 128K tokens
Fable 5.1 1M tokens 128K tokens
Haiku 4.5 200K tokens 64K tokens

Pasting works to a point and then stops abruptly, and the limit is not a number of URLs: it depends on how many columns you exported, so a 50,000-row export with four columns and one with forty are different problems. The measurement takes two minutes. Send one request containing a sample of your real columns, read the input token count the API returns, and multiply. My own working rule, not Anthropic’s, is to stay under a third of the window, because the conversation keeps growing while you work. For millions of URLs, Hitches recommends the opposite of the paste reflex (2 April 2026): preprocess in Claude Code and hand over the summary.

One cost goes almost unmentioned: MCP tool definitions consume context before you ask anything. The Ahrefs connector exposes 61 tools, and a Search Console server and a Semrush server add roughly twenty each. No published measurement of that overhead in SEO was found for this course, so it stays unquantified: connect what a session needs, disconnect the rest.

The failures that are actually documented

Published cases of what goes wrong number fewer than five, against thousands about what works, and share a shape: the failure appears where the model acts unverified.

Template duplication under CMS write access. Will Scott reported this one in Search Engine Land on 28 August 2026. Two landing pages were asked of an assistant holding write access to a content management system, and what arrived was the site’s homepage duplicated twice over, title tags swapped, nothing else touched. Neither page took an impression or a click, the homepage itself did not climb above position nine, and a second site produced the duplication independently. The article declares no sample: two sites is a pattern to watch for, not a rate. It could not be verified in this review, so it is cited by author and date, unlinked.

Third-party estimates treated as fact. ContextBolt’s Claude SEO Experiment, published 23 June 2026, put a week of live SEO work through Claude. Three behaviours came out of it: estimates from third parties returned as though they were measurements; a keyword whose own difficulty score marked it unwinnable was recommended anyway; and invented markup was offered as the format that wins AI citations. Loose instructions also produced keyword-stuffed over-optimisation. No sample size is published: one site, one week, one team, author given only as “David”.

Agent-driven crawls. Rich Voller’s agency guide to Screaming Frog’s v24 MCP server, 26 May 2026, updated 12 June 2026, documents five distinct failure modes, not one. The first is enough here: a crawl set to 1 URL per second that got through 500 URLs in 20 seconds, because the MCP path does not inherit the interface speed setting. The other four are mechanically unrelated to it and to each other, and they get a lesson of their own later in this level. No sample size is published: this is one agency’s testing. A crawl far faster than agreed is a server problem before a data one, which puts it in technical SEO.

None of the three is the model writing a bad paragraph. Each is the model acting, or being believed, with nothing checking the result against a source.

Where this lesson stops and Level 4 begins

What this lesson does not cover is how an AI search engine decides what to cite. Different job, different evidence, and it lives in Level 4 on GEO and AIO. Splitting them is a structural choice about this course, not a claim about the field.

Whether generative engine optimisation is a discipline of its own or ordinary SEO under a new name is genuinely disputed. One camp holds that GEO is largely a vendor invention and competent SEO already covers it; the other sells it as a new speciality with its own techniques. The available evidence splits by surface rather than by camp: an AI answer assembled on top of ordinary search results rewards much of what organic search already rewarded, while an assistant retrieving from an index of its own does not respond to the same playbook. That is a qualitative reading, and the citation-overlap figures behind it belong to Level 4 with their sample sizes attached. Neither camp has data that settles the question, and this course does not settle it. This lesson only declines to blur the two jobs, and the rest of this level works through the connected-data side of the table above.

Common mistakes

  • Asking Claude for search volumes and believing the answer. A volume produced in a chat is text shaped like data. The fix: connect a real source, or use the model to group and prioritise, never to measure (ContextBolt, 23 June 2026, one site, no sample size published).
  • Granting CMS write access before watching it work dry. Granting write access is the worst documented consequence in this lesson: two pages at zero impressions and zero clicks (Search Engine Land, Will Scott, 28 August 2026, two sites, no sample declared). The fix: treat write access as a decision of its own rather than a setting, and keep the model proposing until you have watched it work. The CMS lesson in this level carries the checklist.
  • Treating “it generated it” as “it is validated”. This happens most with schema, one of the better unaided tasks and so the easiest to trust too far. The fix: every piece of JSON-LD goes through the rich results test before it ships.
  • Counting the time saved without counting the setup. The circulating figures are real and partial: six to eight hours down to minutes for a metadata rewrite, three to five days to under an hour for internal linking (StudioHawk, 12 May 2026). None nets off MCP setup, the subscription or the human review those authors insist on. The fix: time your third run, then add the review.

The short version

  • Claude analyses crawl data, it does not crawl sites (Lawrence Hitches, 2 April 2026). Where the data lives outside the conversation, meaning volumes, crawling and Search Console, it answers anyway and sounds the same.
  • The scope table’s second column is the decision: if you will not connect the source, do not delegate the task.
  • The one published task-level comparison (SE Ranking, 7 July 2025) gave three outcomes: schema to Claude at 1 non-critical issue against 5, briefs to Claude, titles to ChatGPT. Qualitative, no sample size.
  • Artifacts are an output format and add no information. No official Search Console connector exists; the ones that do are community-built (Passionfruit, 28 March 2026).
  • Opus 5, Sonnet 5 and Fable 5.1 accept 1M input tokens with 128K output; Haiku 4.5 is 200K and 64K (Anthropic, 20 September 2026). For millions of URLs, preprocess.
  • 1 respondent in 97 called any workflow fully automated (Keyword.com, 1 January 2026, n=97), and the documented failures cluster where the model acts unchecked: two cloned pages at zero impressions and zero clicks (Search Engine Land, Will Scott, 28 August 2026, two sites), five agent-crawl failure modes from one agency (Rich Voller, 26 May 2026).

Frequently asked questions

Can Claude crawl my website?

No. Claude analyses crawl data; it does not crawl sites (Lawrence Hitches, 2 April 2026). It needs a crawler to hand over the data: Screaming Frog has shipped a native MCP server since version 24.0 of 19 May 2026, and an attached export always works. One thing gets confused with this. Anthropic’s support page on its web crawlers documents three distinct agents that fetch pages from the web, and renders no absolute date, only a relative one. Those agents fetching a page is not Claude crawling your site.

Will my whole crawl fit in a million tokens?

That depends on how many columns you exported; no URL count is reliable away from your own file. Opus 5, Sonnet 5 and Fable 5.1 accept 1M input tokens with 128K output, Haiku 4.5 200K with 64K (Anthropic, 20 September 2026). Measure rather than estimate: send one request with a sample of your real columns and read back the input token count. For millions of URLs, preprocess with Claude Code (Hitches, 2 April 2026).

Can using Claude to write get my site penalised?

Not for the drafting as such, on anything published. The scaled content abuse policy is indifferent to how a page was produced, turns on purpose and on value to the reader, and names no page count anywhere in its text (Google Search Central, last updated 28 August 2026). Heavily AI-written pages also rank: Ahrefs put the share at 9% at 80% AI or more, across 331,000 pages holding top-ten positions, published 27 July 2026, its authors stating plainly that their detector is not Google’s. What none of that settles is where scaling stops and spamming starts, and this course does not decide it. Site-level consequences are worked in advanced SEO strategy, measurement in SEO analytics.

Sources

Carry on with the course

This is the first lesson of the level, so there is no previous one. Next is Claude Skills for SEO: Install, Use, and Write Your Own. The level index is AI for SEO: Using the Models as Tools, and all nine levels sit on the free SEO course.