How many words does a website have? Claude alone vs. Claude with Word Counter

Why ask an assistant at all?

“How many words does our website have?” is the first number in every website translation quote, and the first question in most migration and content audits. More and more, it’s typed into an AI assistant instead of a tool. So we tried it, with the assistant most likely to get it right: Claude Code, Anthropic’s agentic tool, which can browse, run shell commands and write its own programs.

The site was wxrks.com — ours. It’s a good test: a Webflow site with a long blog, a mega-menu on every page, and complete translations into 10 languages. Its sitemap lists 10,609 addresses.

How we ran it

• Claude Code 2.1.220, model claude-opus-5 (1M context), on 3–4 November 2026 — non-interactive, with the full log of every step kept.
• A new, empty folder for every run: no memory, no project instructions, no MCP servers except the one being tested.
• The same permissions every time — fetching web pages, web search, curl, wc and python3 — what you’d approve if Claude asked. Spending caps of $5 for the quick question and $8 for the careful one.
• Two questions. The quick one: “How many words does the website wxrks.com have?” The careful one: “I need an accurate word count of the entire website wxrks.com for a translation quote: every page, the total number of words, how many pages there are, and how much of the text is repeated across pages. Count it as accurately as you can.”
• Then both again with Word Counter’s MCP server connected — a copy running the same code as wordcount.wxrks.app, at the free plan’s speed — and nothing else changed.
• Times and costs are Claude Code’s own figures. Costs are at API prices; on a Claude subscription the same work comes out of your usage limits instead.

Round 1 — the quick question, Claude on its own

Claude didn’t guess. It downloaded the homepage, noticed it was a Webflow site, found the sitemap (18.7 MB, 10,609 addresses) and worked out that 9,457 of them were translations. No HTML parsing library was installed, so it wrote one in Python, tested it on three pages, and set two crawlers loose in the background: one for the 1,152 English pages, one for the rest.

11.2 minutes and $1.30 later, it reported:

English site (complete): 2,256,008 words across 1,149 pages — 1,660,196 of those are unique body content — the other ~596k are the nav megamenu + footer (519 words of chrome) repeated on every page.

Then it said it would give the total for every language once the translated pages were done — and ended its turn. Nothing wakes a non-interactive session back up, so that total never came. In an interactive session it would have carried on, for more minutes and more tokens.

It’s a resourceful answer. It also counts every word in the HTML of each page — the first words of its homepage count are the menu: “Platform The wxrks Platform Unleash translation super-powers High-level overview tools…” — and the 1,660,196 it offers instead comes from subtracting a guessed 519 words of “chrome” from every page.

Round 2 — “count it accurately”, Claude on its own

Asked for an accurate count for a quote, Claude went all in. It downloaded the sitemap again, wrote a crawler and fetched 10,515 pages — every language, not just English. Then it spent most of an hour on its own HTML parser: separating the menus and footers from the content, and chasing why some product copy looked hidden (Webflow’s conditional-visibility classes) before deciding the parser was right.

After 52 minutes, 79 tool calls and $5.96 it delivered a thorough report: 2,242,781 words on the 1,152 English pages, 700,928 of them menus and footers and 1,541,853 of page content; 1,114,098 new words after removing repeats, a 50.3% repetition rate; and the same for all ten translations.

That is serious work, and its numbers for the main site are in the same range as Word Counter’s. It also fetched nine times the pages an English quote needs, took the sitemap for the whole website — so the help center, the docs and the community never came up — and the next site starts from zero again.

Round 3 — the quick question, with Word Counter

Same question, Word Counter connected. Claude loaded the tools and made one call, count_words("wxrks.com"). Word Counter started crawling — opening each page in a real browser — and at 50,683 words, on page 26, paused the way it does for every large site: with an estimate of the whole from the sitemap, the repetition so far, the biggest sections and pages, and a question.

45 seconds and $0.19. The estimate — 2,241,748 words on about 1,150 pages, from the sitemap — got the main site’s pages about right and its words high: the first 26 pages were mostly long blog guides, and it extrapolated from them. And no sitemap could tell it about the subdomains the full crawl would find. That’s why it’s labelled an estimate, and why Claude asked before counting everything.

Round 4 — the full count, with Word Counter

Asked for an accurate count for a quote, Claude found the paused count, carried it on with continue_word_count, and checked in with get_word_count — 37 times over 35.1 minutes — while Word Counter worked through the site at the free plan’s speed: four pages at a time, each opened in a real browser. Following links rather than the sitemap, it went well past the main site, into 11 subdomains.

The result: 2,696,246 words on 2,494 pages — 1,718,712 of them on the main site’s 1,050 pages, 977,534 on the help center, the developer docs, the community forum and eight more subdomains. 1,064,373 words (39.5%) repeat a sentence found on another page, leaving 1,604,010 new ones; 799 duplicate pages were counted once.

Claude then did what an assistant is good at: it read the numbers. It suggested leaving the machine-generated API reference and the user forum out of the quote (2,176,673 words without them), pointed out that the forum’s ?sort= pages repeat each other, and named 1,604,010 new words as the figure for a vendor pricing by translation memory. One slip: it called the result “~20×” the estimate. It was 1.2×.

Side by side

On the 1,021 pages both crawled, Claude’s script counted 2,166,342 words and Word Counter 1,691,655: 28.1% more, a median of 478 words a page. That’s the menu and the footer. On a quote priced per word, it’s the menu, translated 1,149 times.

Claude’s own correction — subtract 519 words of “chrome” from every page — landed within 4% of Word Counter’s count for the main site. It deserves credit for that. But it was one guessed number for 1,149 different pages, it said nothing about repetition, and the sitemap it trusted left out 977,534 words on 11 subdomains.

Why the numbers differ

1. Menus and footers, on every page

A one-off script counts every word in a page’s HTML. On wxrks.com that includes a mega-menu and a full footer — 478 words, at the median — on all 1,149 pages. Word Counter opens each page in a browser and counts the main content: menus, footers, sidebars and forms are left out of each page — they get translated once, not once per page. Where a site builds its footer without the HTML that says so, as Webflow does here, the leftover links show up as the report’s most repeated text (1,047 times) — counted once as new, and as repetition everywhere else.

2. What a visitor actually sees

Raw HTML includes text a visitor never sees: in Round 1, Claude’s script noted that Webflow hides elements with display:none — and counted them anyway; in Round 2 it spent much of its hour deciding what to do about them. On many sites raw HTML also misses text that only arrives with JavaScript (wxrks.com renders on the server, so not here). Word Counter counts the page after it has loaded, as a visitor would read it.

3. Repetition

A translation quote prices repeated sentences differently from new ones. Word Counter splits every page into sentences and found 39.5% of the words repeating text from other pages — the CTAs, the disclaimers, the blocks reused across posts. Asked for it, Claude measured repetition too — 50.3% for the English pages, menus and footers included, since its script had counted them on every page. Not asked, it never mentioned it.

4. Where the site actually ends

A sitemap lists what its owner remembered to list. wxrks.com’s doesn’t include its help center, its developer docs, its community or eight other subdomains — 977,534 words between them. Word Counter follows the links a visitor could follow, across the whole domain, and the report shows each subdomain on its own line, so leaving one out of a quote is a decision, not an accident. (You can also ask it to count one host only.)

5. The language you translate from

wxrks.com is already translated into 10 languages: 9,457 of its 10,609 addresses are translations. A quote counts the source, so Word Counter detected English and set the translated versions aside without fetching them, naming each one in the report. Claude, left to itself, set about crawling all 10,609 — and in Round 2, after fetching them, it rightly said the English count is the basis for the quote.

6. Time and money

Building and running a crawler cost Claude 11.2 minutes and $1.30 for the quick question, and 52.1 minutes and $5.96 for the careful one — and the next site starts from zero again. With Word Counter the crawl runs on Word Counter’s servers and Claude spends its tokens on the question and the answer: $0.19 for the estimate. Following a 35-minute count cost $1.34, nearly all of it Claude re-reading the conversation at each of its 37 checks — asking once more at the end would cost cents.

What the report shows

Every count started through the MCP server is also a Word Counter report, with every page, a CSV and a share link. Here is the one behind Round 4.

Connect it to your assistant

It takes a minute. Get a personal connection URL, paste it into Claude, ChatGPT, Cursor or VS Code — or add it to Claude Code with one command — and ask about any website.

Notes on the method

• One site, one model, one day. Assistants don’t answer the same way twice; run it again and the steps will differ, though the menus will still be in the HTML.
• Word Counter’s count isn’t a law of nature either: it’s a definition — the main text of each page, as rendered — and the report shows exactly what it covers and what it left out.
• The session logs, with local file paths removed, are here: Round 1 · Round 2 · Round 3 · Round 4.