SEO audits with Claude: separate observed HTML from measured performance
SEO audits with Claude: why this skill requires a running script, and what it can never measure on its own.
By Educasium
Prepare an audit with Market SEO →
You ask Claude to "run an SEO audit" of your site expecting a full diagnosis — the kind an agency would bill several hours for — and you get back a text that comments on the homepage title and stops there, never touching real Google rankings or performance data measured from real visitors.
This needs to be said clearly, upfront: the Market-SEO skill relies on a bundled script that extracts a page’s HTML, titles, links and tags — but that script must run in an environment capable of executing code, such as Claude Code. Unlike other marketing packs in the same catalog, its instructions provide no manual web-search fallback if that script is unavailable. In a plain Claude.ai, ChatGPT or Gemini conversation without that capability, what you get is limited to analysis of text you paste yourself — a real result, but noticeably less complete than an audit run with the script, and one that never replaces a performance or ranking measurement produced by a dedicated tool.
This article details what this skill actually observes, where that observation stops, and how to get the performance and ranking data that a simple page review always lacks.
Summary
- Why an SEO audit is not just reading a page’s HTML
- Why this skill assumes a running script — and what remains possible without it
- What the pack actually produces
- Preparing an audit with Market SEO step by step
- What Market SEO does not guarantee
- Training to read an SEO audit without trusting an invented figure
- Frequently asked questions
Why an SEO audit is not just reading a page’s HTML
A complete SEO audit crosses three types of information that do not resemble each other: what is written in the page’s code (titles, tags, structure), what a search engine actually shows in its results (rankings, rich snippets), and what visitors experience once on the page (load speed, visual stability). A skill that only observes HTML answers the first category; the other two require tools and data that HTML alone never contains.
What Claude can observe directly in a page
From a page’s source code — pasted in or fetched by the pack’s script — Claude can pick out titles, meta descriptions, heading structure, internal and external links, and missing image attributes. That reading is reliable for what it covers: it describes what is actually written in the code, not what a visitor sees after JavaScript rendering nor what Google actually indexes.
What it can never measure on its own
Traffic, search rankings and Core Web Vitals cannot be inferred from reading code: they are measurements produced by specific tools — Google Search Console for rankings and traffic, PageSpeed Insights or an equivalent tool for Core Web Vitals. Without an export or a real measurement you supply, those figures stay absent from the audit, and should appear as such rather than as an invented estimate.
Why this skill assumes a running script — and what remains possible without it
It needs to be said plainly, as with other skills in the catalog that depend on an execution environment: Market-SEO combines automated page-data extraction with editorial analysis, and its instructions provide no manual web-search fallback if the script is unavailable, unlike other skills in the same marketing pack.
| Your situation | What is actually possible | What to do |
|---|---|---|
| Claude Code, with Python and the bundled script runnable | Extracting HTML from several pages, structured collection of titles, links and tags | Also supply Search Console and PageSpeed exports to complete the audit |
| Conversation in Claude, ChatGPT or Gemini, without running the script | Analysis limited to source code you paste yourself, page by page | Paste each page’s source code, and separately supply any available traffic or ranking data |
| A profile with no access to Search Console data | HTML-only analysis, with no measured performance data at all | Request access to the site’s Search Console, or have that part done by whoever holds it |
| Claude via the API, without a code execution tool | Same limit as a plain conversation: analysis of supplied text only | Supply page content as text and clearly document what was not measured |
What the script adds when it can run
In Claude Code, the pack’s script automates fetching HTML from several pages in sequence, avoiding repeated manual copy-pasting and covering more pages in one pass. It remains limited, however, to the HTML received by a plain request: a page that only renders correctly after JavaScript execution can appear empty or incomplete in that collection — a case to flag explicitly rather than ignore.
What you get without it, presented honestly
Without a runnable script, pasting a page’s source code into the conversation produces a real analysis of that specific page — not a useless degraded version, but page-by-page work rather than an automated pass over the whole site. Presenting that result as equivalent to an automated multi-page audit would be inaccurate; presenting it as a targeted analysis deliberately limited to what you supplied stays honest.
What the pack actually produces
The pack drafts an SEO-AUDIT.md file that pairs findings drawn from the pages’ code with recommendations, explicitly distinguishing what was observed, what was measured from an external source you supplied, and what remains unknown for lack of data.
Linking every recommendation to a finding or a supplied measurement
A useful recommendation rests on a problem visible in the code or on data actually supplied — two pages sharing the same title, an image with no alt attribute, a ranking reported by a Search Console export. A recommendation resting on neither remains a generic hypothesis, to be stated as such rather than presented as an established diagnosis.
Distinguishing received HTML from what is actually rendered
For a page built with a JavaScript framework, the HTML fetched by a plain request can differ sharply from what a visitor actually sees after full load. The audit must state which of the two versions it examined; otherwise a page that renders correctly in a browser but appears empty in the raw HTML would be flagged as a problem that does not exist for a real visitor.
Preparing an audit with Market SEO step by step
Step 1: Define the pages to review and their intent. Choose a limited number of representative pages — the homepage, a service page, a blog post — rather than an exhaustive list impossible to process by hand without the script, and state what each page is meant to help visitors find or decide.
Step 2: Gather already-available performance and ranking exports. Collect a recent Google Search Console export and, if possible, a PageSpeed Insights report; without this data, the audit must state so explicitly rather than letting a reader assume it was accounted for.
Step 3: Supply the pages’ source code, via the script or by pasting it in. In Claude Code, let the script fetch the HTML of the listed pages; in a plain conversation, paste that code yourself page by page, noting whether a page depends heavily on JavaScript rendering.
Step 4: Check that every recommendation cites its source. Reread the produced file and flag any recommendation that points to neither an observed element in the code nor a supplied piece of data; rephrase it as a hypothesis to verify rather than leaving it as a definitive finding.
Step 5: Compare the pack’s thresholds against Google’s current benchmarks. The thresholds built into the skill date from its last update; check them against current Google Search Central and web.dev documentation before applying a fix based solely on a potentially outdated benchmark.
What Market SEO does not guarantee
The skill never measures traffic, rankings or Core Web Vitals itself: those figures come solely from what you supply, and their absence should stay visible in the report rather than hidden behind an estimate. An audit displaying a complete score with no real performance data would give a false impression of rigor.
It also does not guarantee that a suggested change will improve a specific ranking: the link between a technical fix and a ranking change depends on many factors that neither this skill nor any tool can isolate with certainty. Presenting a recommendation as a guaranteed ranking gain would be a promise nothing here can back up.
In the audits we run for architect or self-employed clients, the most frequent mistake is not technical: it is confusing a site that "looks clean" in HTML with a site that actually gets traffic — two realities that do not always line up.
Training to read an SEO audit without trusting an invented figure
Telling a real technical finding apart from an unverified estimate requires a minimal understanding of how search engines work, which is built through guided practice. Educasium’s Master Claude training covers preparing and critically reading AI-produced audits, for profiles who have never taken SEO training before. For a business owner or a self-employed professional under the relevant scheme, the FIFPL covers part of the cost of this Qualiopi-certified training under the 2026 criteria set at €300/day and €900/year, with e-learning capped at 50% of the daily rate.
Frequently asked questions
Can Market-SEO give me my current Google rankings?
No, under no circumstances: no part of this skill accesses Google Search Console or any ranking-tracking tool. Rankings reported in an audit can only come from an export you supply yourself; without it, that section should stay marked as not measured rather than filled with a plausible-looking estimate. If a Search Console export exists but was not supplied in this session, ask for it explicitly before treating the report as complete on that point.
What happens if I use this skill in a conversation without running the script?
The analysis is then limited to the source code of the pages you paste yourself into the conversation, page by page, without the automated multi-page pass the script enables in Claude Code. The result stays real for the pages supplied, but it must be presented as a targeted analysis, not an exhaustive audit of the entire site. If the need covers several dozen pages, plan to switch to Claude Code to benefit from the script rather than repeating this manual copy-paste page after page.
Are the Core Web Vitals mentioned in the report reliable?
Only if they come from a real measurement you supplied, such as a PageSpeed Insights report or a Search Console figure; the skill itself cannot measure a page’s load speed or visual stability. Always check these figures against web.dev or the tool that produced them before treating them as final. A report that shows a Core Web Vitals score without stating its source should be treated as suspect, whatever the apparent quality of the rest of the audit.
How do I know whether a recommendation in the report is solid or speculative?
Check whether it points to a specific element observed in a page’s code or to data you supplied; if neither can be cited, treat the recommendation as a lead to verify rather than an established finding. A properly written report flags this distinction itself rather than leaving it to be discovered. If doubt remains after this check, explicitly ask Claude to cite the exact line of code or data point that justifies the recommendation.
An SEO audit produced with Claude stays useful for organizing technical and editorial work, but it never replaces real Search Console and performance-measurement data — those, not the report, say whether a site is actually seen and visited. That is the role of the Market SEO audit-preparation pack, to be used in Claude Code to benefit from the extraction script, alongside Google Search Central’s SEO Starter Guide and web.dev’s Web Vitals documentation to check every cited threshold.
Once the technical audit is in place, comparing your offer against competitors calls for a different method: our guide to comparing competitors with verified sources details how to avoid introducing the same approximations there.
Training 100% fundable via OPCO/FIFPL. Qualiopi-certified programme. To learn how to prepare and read an SEO audit with AI as part of our Master Claude training, contact Educasium and specify your status (employee, self-employed, business owner) and your goal.