Opus 5.5 for SEO: what it does, and what it still needs.

1. What Opus 5.5 changes for SEO work.

Opus 5.5 is the model Claude's own docs now point to first: they tell developers to start with Claude Opus 5.5 for most workloads, and the launch post says it performs at the level of Fable 5.1 on most work while costing 40% less to run than Opus 5. If you are still choosing between models, the comparison of Claude models for writing covers the whole lineup. For SEO, three of the changes matter more than any benchmark.

It follows your writing rules.

The launch post says Opus 5.5 puts the most important information up front, is less likely to use jargon or idiosyncratic phrases, and follows the writing rules you give it. That is the part of SEO writing that used to need the most hand editing: a style guide that holds for 2,000 words, banned phrases that stay banned, and an answer in the first paragraph instead of a warm-up. Give it your rules once and check that they survive to the last section. The Claude vs ChatGPT writing comparison covers how that stacks up against the other assistant.

It holds a whole site in one session.

Opus 5.5 has a 1M-token context window and can write up to 128K tokens in one response. In SEO terms, that is room for your full sitemap URL list, a Search Console export of the queries each page ranks for, your brief, and the pages ranking for the keyword today, all in one conversation. That turns a cannibalization check or internal link mapping from your real URLs into a single-session job instead of a spreadsheet project.

It costs less to run at volume.

On the API, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. In the Claude apps it is available on the Pro, Max, Team and Enterprise plans. Our model guide works out what one article costs on each Claude model, and at these prices the writing step is the cheap part of an SEO article.

What did not change: its knowledge stops at June 2026, it has no keyword database, and it cannot see your rankings. The launch post itself says benchmark margins at this level have become a less reliable guide to real-world differences. That is a fair rule for SEO too. Judge Opus 5.5 on your own drafts, against your own checks.

2. What Opus 5.5 still cannot do on its own.

Ask whether AI can do SEO now and the honest answer splits in two. The writing part, yes: give Opus 5.5 an exact brief and it will write you a strong article. But a brief is the output of four other jobs, and the model does none of them unprompted. Those four jobs are what an automated SEO pipeline exists for, and they are the four parts in the video.

1. It does not know what people search for. There is no measured monthly volume and no difficulty score inside the model. Ask for one and you get a number shaped like a measurement.
2. It does not know what is already on your site. Unless your pages are in the conversation, it will happily plan an article you already have.
3. It does not check its drafts against sources by default. A clean draft and a correct draft look the same. Verification has to be its own step.
4. It does not publish, and it does not do it every day. A chat ends with text in a window. Rankings come from months of steady publishing.

So an AI SEO agent is the model plus the four parts around it. The version-free guide to using Claude for SEO covers the do-it-yourself setups (a Project, Claude Code skills, MCP connectors) that hold for any Claude model. The rest of this page takes the four parts in order: what to do with Opus 5.5 by hand, and what the job looks like when it runs on its own.

3. Part 1: keywords with real volume and difficulty.

Every keyword has two numbers you want before you write anything: how many people search for it each month, and how hard it is to reach page one. Take the two examples from the video, both US figures. The keyword “seo” gets about 90,000 searches a month, but its difficulty is 88 out of 100, so a small site has almost no chance at it. “Local seo for plumbers” gets about 590 searches a month with a difficulty close to zero, so a small site can rank for it and collect those visitors every month.

Opus 5.5 cannot produce either number. There is no keyword database inside the model, and its knowledge ends in June 2026, so any volume it states is an estimate written in the same confident voice as a fact. The pattern that works: export real numbers from a keyword tool or your Search Console data, paste the list in, and give the model the jobs it is good at. It groups the list into clusters, labels search intent, and picks the keywords a site your size can win. Tell it in plain words never to estimate a volume or a difficulty.

This is the first part the agent takes over. The SEO Agent pulls the real monthly volume and difficulty for your market and builds a 30-day plan from the keywords a site like yours can win. Every keyword card shows both numbers, so you can see why it was picked. Turn on approval and the plan waits for your yes. Leave it off and it works through the plan one keyword a day. The AI SEO software overview walks through the rest of the product.

4. Part 2: what is already on your site.

If two of your pages go after the same keyword, they compete with each other on Google, and often neither ranks as well as one page would. That is keyword cannibalization. A model that plans articles without seeing your site walks straight into it, because every topic looks fresh from inside an empty conversation.

Opus 5.5 handles this check well once it has the material. Paste your sitemap URL list and a Search Console export of the queries each page ranks for, then the planned keyword list, and ask it to flag every planned keyword an existing page already targets, with the URL. For most sites that inventory fits in its context window. The catch is that the inventory goes stale the day you publish, so the check has to run again before every plan. Connecting Search Console over MCP keeps the export live instead of pasted.

The agent runs this before it plans anything. It reads every page in your sitemap and the keywords those pages rank for, and a keyword you already have a page for never makes it into the plan. Every article it publishes joins that index, so it does not write the same article twice. On a WordPress site running on autopilot, that is the difference between a growing archive and a pile of near-duplicates.

5. Part 3: quality checks and scaled content abuse.

This is where AI content gets sites in trouble, and it is not because a model wrote it. Google has said on its Search Central blog that its focus is on the quality of content rather than how it is produced. What its spam policies do not allow is generating many pages without adding value for users, which Google calls scaled content abuse. The risk is volume without checks, not the model.

A better writer makes this harder to spot, not easier. Opus 5.5 produces fluent, well-ordered prose, and a fluent draft with an invented statistic looks exactly like a fluent draft without one. The fix is a separate pass with its own instructions: list every factual claim, attach the source that supports it, and cut or rewrite what has none. Run it as a different prompt from the one that wrote the draft, so the check is not grading its own homework. That is the job an AI fact-checking pass does on every draft.

In the agent, the check is not optional. Before it writes, it reads the pages that already rank for the keyword, and it links to its sources inside the article. Then every draft has to pass a quality check before it can go live. If a draft fails, it is fixed and checked again, without anyone touching it.

6. Part 4: publishing every day.

The last part is the one most people give up on. One article does little on its own, because SEO builds up over time, and the sites that do well tend to keep publishing for months. A session with Opus 5.5 ends with text in a window. Somebody still has to format it, make the images, paste it into the CMS, fix the links, and do it all again tomorrow. That daily grind is what autoblogging with a quality gate is meant to remove.

The agent connects to whatever your site runs on: WordPress, Webflow, Shopify, Wix, or any stack through a webhook. Each article goes up with its images already made, and listicles about tools use real screenshots of those tools. If an article is updated later, the live post is updated instead of duplicated. Connect Google Search Console and the last 28 days of clicks and impressions show on the dashboard. Then turn on auto-publish, and each day's article goes live on its own.

One thing no model changes: a brand-new site still takes a few months to start ranking. The useful move is to keep publishing checked articles every day while those months pass, which is cheap to keep up when nobody has to sit in a chat window to do it. The full pricing for the daily plan is one flat monthly price.

A worked example: one keyword, from pool to published.

Here is the whole loop for one article with Opus 5.5 as the writer, done by hand. Say the site belongs to a small agency that does marketing for trades businesses. The numbers are the US figures from the video.

1. 01 · POOL

   seo: about 90,000 searches a month, difficulty 88. Skip, a small site cannot win it. local seo for plumbers: about 590 a month, difficulty near zero. Pick.

2. 02 · SITE CHECK

   Sitemap URL list and a Search Console export pasted in. No existing page ranks for local seo for plumbers or a close variant, so it goes into the plan. Had the agency services page already ranked for it, the keyword would be dropped and that page improved instead.

3. 03 · BRIEF AND DRAFT

   One prompt to Opus 5.5 with the rules, the sources and the allowed URLs (below).

4. 04 · CHECK

   A second, separate prompt lists every factual claim and quotes the source line behind it. Anything marked UNSUPPORTED is cut or rewritten.

5. 05 · PUBLISH

   Title, meta description, images and schema added, pasted into the CMS, and the keyword plus the new URL logged in the inventory so the next plan skips it.

The draft prompt for step 3. Everything in square brackets is material you paste in, which is where the facts come from:

Write an article of about 1,800 words targeting "local seo for plumbers".
Readers: owners of plumbing businesses with one to five vans.

Rules:
- Answer the main question in the first paragraph.
- No statistic, price or date unless it appears in SOURCES.
- Headings are plain questions a plumber would ask.
- Link only to URLs from ALLOWED URLS, at most one per section.
- Never use a phrase from BANNED PHRASES.

SOURCES: [the three pages ranking now, plus your own notes]
ALLOWED URLS: [your real pages]
BANNED PHRASES: [your list]

And the check prompt for step 4, run in a fresh message:

List every factual claim in the draft below.
For each claim, quote the line from SOURCES that supports it.
If no line supports it, mark it UNSUPPORTED. Do not rewrite anything yet.

Opus 5.5 handles steps 3 and 4 well and quickly, and its habit of following the rules you give it is what keeps the draft inside the sources. Steps 1, 2 and 5 are data and plumbing: a keyword export, a sitemap and Search Console export, and a CMS session. Done by hand, that is the part that eats the afternoon, and it has to happen again tomorrow.

Common mistakes when using Opus 5.5 for SEO.

1. Asking Opus 5.5 for search volume. It will answer. The number is an estimate written like a measurement, from a model whose knowledge stops in June 2026. Bring measured numbers in and tell it never to estimate.
2. Reading fluency as accuracy. A better writer produces more convincing errors. Keep the claim check as its own step, on every draft.
3. Planning without the sitemap. Every topic looks new from an empty conversation. Paste the inventory first, or you will write the page you already have.
4. Scaling before the check exists. Thirty unchecked drafts a month is the pattern Google's scaled content abuse policy describes. Volume is only worth having once every piece clears the same bar.
5. Switching models to fix a process problem. When drafts come out wrong, the gap is usually missing data or a missing check, not the model. If the model really is the issue, see which Claude model writes best.
6. Expecting a new site to rank in weeks. No model changes how long a new domain takes to earn trust. Plan for months of steady publishing.