How to Use Claude for SEO: What It Does Well, and the Data It Needs
Claude is strong at the language half of SEO and blind to the data half. Here is which jobs to give it, how to hand it your own Search Console numbers, and the one mistake that makes its answers look right when they are not.
Use Claude for the parts of SEO that are language work: rewriting a title against the searches a page already gets, turning a page into a brief of what it is missing, grouping a long list of searches by what the searcher wants. Give it your own data for anything that involves a number, because Claude has no live view of Google and will fill the gap with plausible figures if you let it.
That split decides almost everything about getting useful answers. The rest of this guide is the jobs worth handing over, how to get your real numbers into the conversation, and how to catch the answers that sound right and are not.
TL;DR:
- Claude is good at the language half of SEO: titles, meta descriptions, briefs, grouping searches by intent, explaining a results page you paste in.
- It has no search volumes, no rankings and no view of your Search Console. Ask for those without giving it data and it will often answer anyway, with numbers that have no source.
- Paste or upload your own Search Console export, or connect a data source over MCP, and ask it to work only from what you gave it.
- A Search Console export stops at 1,000 rows, and its query rows never add up to the page totals. Tell Claude both, or its arithmetic will be wrong in a way that looks fine.
- If your site lives in a repo, the same work goes further in Claude Code, where the edit happens in the file and can be checked afterwards.
What is Claude actually good at for SEO?
The jobs where the input is text you already have and the output is better text. Claude reads a page, a list of searches or a results page faster than you do and writes a first draft that is usually close. Those jobs, in rough order of how much time they save:
- Rewriting a title and meta description. Give it the current title, the searches the page appears for and what the page actually offers. Ask for three options that use the searcher's words and promise only what the page delivers. You pick, and you check the length.
- Grouping searches by intent. Paste 200 searches and ask it to sort them into "wants to learn", "wants to compare" and "wants to buy", with the reason on each line. This is the slow part of keyword work done by hand, and Claude is reliable at it as long as you read the edge cases.
- Finding what a page is missing. Paste your page and the headings of the pages ranking above it, then ask what they answer that yours does not. The output is a gap list, not a verdict: you decide which gaps the searcher cares about.
- Writing a brief. Turn that gap list into an outline with the questions each section must answer. Briefs are where Claude is most useful and least risky, because a person still writes or edits the page.
- Structured data. Ask for JSON-LD for an article, a product or a FAQ, from the page's real content. Validate it with Google's Rich Results Test before shipping, because a valid-looking block can still name a type Google does not use for that page.
Notice what these share: none of them needs Claude to know anything about Google that you did not paste in.
What can't Claude know without your data?
Anything measured. Claude does not know how many people search for a phrase, where your page ranks, how many clicks it earned last month or whether Google has indexed it. Those numbers live in Google's systems and in your Search Console, not in a language model.
The trap is that it often answers anyway. Ask "what is the search volume for project management software for agencies" in a plain chat and you may get a confident figure, a difficulty score and a trend, with no source behind any of them. The answer reads exactly like a real one. Most guides on this topic name the same limit for the same reason: it is the mistake that costs the most, because a plan built on an invented number looks identical to a plan built on a measured one.
Two habits close most of the gap:
- Ask where each number came from. If the answer is "estimated" or "typical for this kind of search", treat it as a guess and drop it.
- Tell it to work only from what you provide. "Use only the data in the attached file. If a figure is not in it, say so." This turns an invented number into a visible gap, which is what you want.
Try it
Point an SEO agent at your own site, not just an essay about one.
QueryScope holds the plan for your site, reads your real Search Console, and hands the coding agent you already have one change at a time to rank for the searches you decided to win. From $49/month, everything included.
How do you give Claude your own Search Console data?
Two ways, and they suit different jobs.
Export and upload. In Search Console, open the Performance report, filter to the page or date range you care about, and use Export. You get a spreadsheet you can attach to the chat. This is the fastest route for a one-off question about one page or one month.
The export has two limits Claude will not know unless you tell it. It stops at 1,000 rows, which Google describes as "representative examples of your data", so a site with thousands of searches exports a slice. And the query rows never add up to the page or site totals, because Google leaves rare searches out of the query table for privacy while still counting their clicks in the totals. Say both in your prompt. Otherwise Claude will sum the rows, report a total lower than your real one, and the mistake will look like a finding. The export limits in detail and why page-level and query-level numbers disagree cover both.
Connect a data source. Claude can call outside tools over the Model Context Protocol (MCP). In the Claude apps you add one under Customize, then Connectors, as a custom connector, and Anthropic's help page lists it on every plan, with the free plan limited to one. Connect a Search Console server, a keyword data provider or an SEO tool you already pay for, and Claude can fetch live figures instead of guessing them. The upside is fresh data on every question. The cost is that you now have to judge which tool to trust when two of them return different answers, and read what each one asks permission to access.
A workflow that holds up
For one page, start to finish:
- Pick the page from your data, not from a hunch. In Search Console, find a page that appears for searches on the first page of results but earns few clicks. That is the cheapest page to improve, because Google already considers it relevant.
- Export its searches for the last three months and upload them, with the two limits above stated in the prompt.
- Ask Claude what the searches have in common and whether the title matches the words people actually use.
- Ask for three title and meta options that use those words and claim only what the page delivers. Pick one yourself.
- Make the change and write down the date. Google needs to recrawl the page before anything can move, and a title change usually takes a few weeks to show in the numbers.
- Check the same searches on a date you chose in advance. Compare the clicks for those searches, not the whole site, because the rest of the site moves for its own reasons.
Step 6 is the one almost everyone skips, and it is the only step that tells you whether Claude helped. How long SEO changes take to show has the waits per kind of change.
When does Claude Code do this better?
When your site is a codebase. In the chat app you copy a suggestion, open the file, paste it and deploy. In Claude Code the agent reads the page in your repo, makes the edit, runs your build and shows you the diff, so the language work and the change happen in one place. Connect your Search Console to it and it picks the page from your real numbers too.
That setup has its own catches, mostly about which instruction the agent follows when a skill and a data connection cover the same job. Claude Code SEO: skills, MCP and the loop that works walks through the wiring, where to put an instruction so it sticks, and what the agent still cannot see however you set it up.
Either way, the rule is the same one this guide started with. Let Claude do the writing and the sorting, give it your own numbers for everything else, and measure each change on a date you picked before you made it. SEO work stops at the click: whether a better title brought visitors who sign up is a question for your analytics, not for Search Console or Claude. For the one-line meaning of any Search Console metric, the glossary has it with the caveat that comes with it.
Sources
- Search Console Help, Export data directly from a Search Console report ("Data is truncated to 1,000 rows of representative examples of your data"; "Totals shown are actual totals, not totals with any row limits applied").
- Google Search Central Blog, A deep dive into Search Console performance data filtering and limits, 2022 (rare queries are left out of the query table while still counted in page and site totals).
- Claude Help Center, Getting started with custom connectors using remote MCP (custom connectors are available on Free, Pro, Max, Team and Enterprise plans, with Free users limited to one; added under Customize, then Connectors).
- Google Search Central, Rich Results Test (validates structured data and shows which rich results a page is eligible for).
Try it
Put an SEO agent in your editor.
Ask your coding agent what to work on. QueryScope reads your real Search Console, holds the searches you decided to win, and hands back one move at a time with the proof it worked. From $49/month, everything included. 1, 3 or 10 sites.