Page length told us nothing about rank in our sample.
Content briefs still open with a word target: match the average length of the top ten, to the hundred. Five tables we already had carried a position beside a word count, so we ran the correlation on each. On our own corpus, 116 pages from the Google (us) top 10 across 20 keywords, the mean Spearman correlation between length and position is +0.024. NeuronWriter’s own competitor tables, 15 keywords, give −0.063. Clearscope’s, 3 keywords on a trial quota, −0.195. Scalenut’s, 8 keywords, −0.123. Surfer SEO’s, 20 keywords, +0.063. We report the five apart and never average them, because each counts words its own way. Not one of them landed far enough from zero to set a word target by.
What each source measured
Only the first row is scored against the frozen Google (us) top 10 we built before touching any tool. The other four are each tool grading its own SERP on its own day, so the crawl, the ranks and the word count are all its own. The exact fields sit under the table.
| Source | Positions come from | Mean ρ | Median ρ | Keywords | Pages | Rows dropped |
|---|---|---|---|---|---|---|
| Our corpus | Our frozen Google (us) top 10 | +0.024 | +0.086 | 20 | 116 | 30 of 146 |
| NeuronWriter | Its own SERP snapshot | −0.063 | −0.033 | 15 | 424 | 39 of 463 |
| Clearscope (reference only) | Its own SERP snapshot | −0.195 | −0.134 | 3 | 50 | 13 of 63 |
| Scalenut | Its own SERP snapshot | −0.123 | −0.107 | 8 | 126 | 21 of 147 |
| Surfer SEO | Its own SERP snapshot | +0.063 | +0.052 | 20 | 187 | 13 of 200 |
← swipe the table sideways for the rest of the columns
No total row. A mean over these five would state a number none of them measured.
Our word count is our own tokenizer run over the extracted body text, navigation and footers included. The other four report their own counts and none documents the method, so the same page can come out hundreds of words apart depending on who counted. The positions differ too, because each source crawled on its own day and saw a different set of pages. The rows sit beside each other and stop there.
Verbatim from the report, one block per source.
- Our corpus — our own corpus
- Positions: data/keywords/fixtures/serp-truth-v1.json (Ahrefs serp-overview, country=us, top-10, fetched 2026-08-15/16 & 2026-08-18)
- Word count: data/raw/serp-truth-v1/{keyword_id}/{position}.md (Tavily extract, extract_depth=basic)
- Counting rule: clean() strips fenced code, markdown images/links (anchor text kept), bare URLs and HTML tags; then non-[a-z0-9\s-] -> space and whitespace split. No stopword removal, no length filter, no boilerplate (nav/footer/related-posts) stripping.
- Excluded: non-web SERP entries (video); URLs whose Tavily extract came back empty (extraction failure != a zero-word page)
- NeuronWriter — NeuronWriter Select-competitors table (self-reported)
- Positions: NeuronWriter's own crawled rank (google.com/en, captured 2026-08-19)
- Word count: NeuronWriter's own words column
- Counting rule: NeuronWriter's, undocumented — NOT comparable to own_corpus token counts
- Excluded: rows with words 0/null or no rank
- Clearscope — Clearscope Competitors table (self-reported)
- Positions: Clearscope's own SERP snapshot, desktop column (captured 2026-08-20)
- Word count: Clearscope's own word_count column
- Counting rule: Clearscope's, undocumented — NOT comparable to own_corpus token counts
- Excluded: analysis_failed rows, word_count 0/null, rows without a desktop position
- Scalenut — Scalenut Competition table (self-reported)
- Positions: Scalenut's own crawled rank
- Word count: Scalenut's own word_count column
- Counting rule: Scalenut's, undocumented — NOT comparable to the own_corpus token counts in cross-findings-2026-08-21.json
- Excluded: rows with word_count 0/null or no rank; exact-duplicate rows deduped first
- Surfer SEO — Surfer Organic Competitors table (self-reported)
- Positions: Surfer's own crawled position
- Word count: Surfer's factors.body.wordCount.all
- Counting rule: Surfer's, undocumented — NOT comparable to the own_corpus token counts in cross-findings-2026-08-21.json
- Excluded: rows with word_count 0/null or no position; exact-duplicate rows deduped first
How this was measured
Inside each keyword we paired every page’s position with its word count and ran Spearman on the pair, the same statistic and the same sign convention the tool tests use. Read it this way: +1 means the longest page ranked first and the shortest ranked last, 0 means length carries no rank information, and a negative value means the shorter pages ranked higher. We publish every keyword rather than the mean alone, because a mean near zero hides two different results: steady noise, and violent disagreement.
Nothing new was fetched. The tables were already in the repo from the tool runs, so this cost no quota and no API call. The script that recomputes the first three is scripts/tests/content-opt/cross/mine_findings.py; a tool added later carries the same block in its own report, and the page reads it from there. The fixture and the ground truth are on the methodology page.
Definition as recorded in the report: Per keyword, Spearman rho between -position and word count; positive = longer pages rank higher. Three sources reported separately and never averaged together: each pairs its own position column with its own word-count column.
Our corpus, keyword by keyword
116 pages across 20 keywords, positions from the frozen SERP truth, word counts from the text we extracted. The pages run from 59 words to 12,715, so the range is wide enough for a length effect to show up. The mean is +0.024 and the median +0.086.
| Keyword | Pages | ρ | Median words | Shortest–longest |
|---|---|---|---|---|
| best ai for math | 4 | −1.000 | 1,400 | 143–1,958 |
| ai for coding | 4 | −0.600 | 165 | 99–355 |
| ai voice generator | 7 | −0.321 | 924 | 409–1,453 |
| ai for teachers | 7 | −0.321 | 1,225 | 374–3,693 |
| digital nomad visa countries | 7 | −0.286 | 4,743 | 1,765–10,192 |
| best help desk software | 4 | −0.200 | 3,221 | 1,499–6,469 |
| best ai writing tools | 5 | −0.100 | 3,980 | 2,946–9,528 |
| gpu comparison | 7 | −0.071 | 406 | 151–7,023 |
| best laptop for programming | 5 | 0.000 | 1,726 | 445–2,684 |
| best crm for small business | 6 | +0.086 | 1,648 | 583–2,010 |
| ai for customer service | 6 | +0.086 | 2,564 | 1,586–3,353 |
| ai for small business | 7 | +0.143 | 2,246 | 849–3,519 |
| ai for business automation | 6 | +0.143 | 1,792 | 452–4,859 |
| best ai chatbot | 5 | +0.200 | 6,611 | 1,202–8,878 |
| ai content detector | 8 | +0.286 | 2,052 | 612–3,163 |
| ai presentation maker | 6 | +0.314 | 1,454 | 59–2,488 |
| best ai for writing | 4 | +0.400 | 3,463 | 997–9,528 |
| cost of living comparison | 6 | +0.429 | 526 | 117–1,484 |
| ai for marketing | 6 | +0.638 | 2,951 | 1,520–6,279 |
| best ai seo tools | 6 | +0.657 | 3,690 | 1,081–12,715 |
← swipe the table sideways for the rest of the columns
The per-keyword values run from −1.000 on best ai for math to +0.657 on best ai seo tools, and every keyword rests on 4 to 8 pages. At that sample size one page moving two positions flips the sign. 11 keywords came out positive, 8 negative, 1 flat. Quote one row on its own and you can show whichever direction you came in with. The mean over 20 keywords is the smallest unit worth citing here.
30 of the 146 web URLs came back empty from the extractor — paywalls, bot blocking, pages that assemble their text in the browser. We dropped them instead of recording them as zero-word pages. That removal is not random. The pages we cannot read skew towards JavaScript-heavy tool sites, and those sites tend to carry short body copy, so the sample that survived is missing short pages more often than long ones. Had they been readable, they would more likely have pulled this correlation down than up. Treat +0.024 as the friendlier end of the range.
What the tools’ own tables say
Each of these tools builds a competitor table per keyword and prints a word count beside each rank. Those two columns are the tool’s own, so no table here checks our numbers. We put the same question to four more tables.
NeuronWriter
424 rows kept of the 463 its tables held, across 15 keywords. Mean ρ −0.063, median −0.033: 6 keywords positive, 9 negative. The per-keyword swing is the same story as ours: −0.607 on best crm for small business to +0.255 on best ai for writing.
| Keyword | Pages | ρ | Median words | Shortest–longest |
|---|---|---|---|---|
| best crm for small business | 27 | −0.607 | 3,435 | 411–8,007 |
| gpu comparison | 30 | −0.326 | 656 | 2–15,547 |
| best ai seo tools | 28 | −0.219 | 3,417 | 11–13,051 |
| cost of living comparison | 28 | −0.208 | 781 | 41–5,938 |
| best laptop for programming | 28 | −0.180 | 1,646 | 5–9,624 |
| ai voice generator | 30 | −0.179 | 1,266 | 25–6,144 |
| ai for business automation | 26 | −0.159 | 2,469 | 14–6,135 |
| ai for small business | 29 | −0.033 | 1,767 | 28–6,318 |
| ai for customer service | 30 | −0.028 | 2,196 | 290–8,740 |
| ai presentation maker | 30 | +0.008 | 1,317 | 28–16,608 |
| best ai chatbot | 25 | +0.063 | 1,993 | 60–10,636 |
| digital nomad visa countries | 29 | +0.181 | 3,848 | 138–19,184 |
| ai content detector | 29 | +0.235 | 1,871 | 65–5,698 |
| ai for coding | 27 | +0.253 | 1,650 | 7–6,889 |
| best ai for writing | 28 | +0.255 | 3,173 | 72–11,971 |
← swipe the table sideways for the rest of the columns
Clearscope: a reference row, not a result
50 rows kept of the 63 its tables held, across 3 keywords. Mean ρ −0.195, median −0.134: 0 keywords positive, 3 negative. 3 keywords cannot show a direction; change any one of them and the answer changes. The row is here as a third independent table, not as a finding.
| Keyword | Pages | ρ | Median words | Shortest–longest |
|---|---|---|---|---|
| ai voice generator | 19 | −0.356 | 1,518 | 228–2,936 |
| gpu comparison | 14 | −0.134 | 1,444 | 238–17,967 |
| cost of living comparison | 17 | −0.096 | 1,570 | 253–5,141 |
← swipe the table sideways for the rest of the columns
Scalenut
126 rows kept of the 147 its tables held, across 8 keywords. Mean ρ −0.123, median −0.107: 3 keywords positive, 5 negative. The per-keyword swing is the same story as ours: −0.389 on cost of living comparison to +0.093 on ai for small business.
| Keyword | Pages | ρ | Median words | Shortest–longest |
|---|---|---|---|---|
| cost of living comparison | 13 | −0.389 | 617 | 54–3,116 |
| gpu comparison | 11 | −0.345 | 1,151 | 24–4,840 |
| ai presentation maker | 16 | −0.218 | 876 | 233–20,666 |
| ai voice generator | 17 | −0.187 | 593 | 231–1,874 |
| ai content detector | 18 | −0.028 | 1,390 | 154–8,472 |
| best ai chatbot | 15 | +0.020 | 3,322 | 281–6,449 |
| ai for coding | 19 | +0.069 | 2,667 | 86–3,896 |
| ai for small business | 17 | +0.093 | 1,905 | 239–3,531 |
← swipe the table sideways for the rest of the columns
Surfer SEO
187 rows kept of the 200 its tables held, across 20 keywords. Mean ρ +0.063, median +0.052: 11 keywords positive, 9 negative. The per-keyword swing is the same story as ours: −0.612 on best crm for small business to +0.867 on best ai chatbot.
| Keyword | Pages | ρ | Median words | Shortest–longest |
|---|---|---|---|---|
| best crm for small business | 10 | −0.612 | 3,126 | 1,379–9,158 |
| ai voice generator | 8 | −0.524 | 1,403 | 648–2,436 |
| best help desk software | 8 | −0.381 | 4,567 | 2,442–8,500 |
| gpu comparison | 10 | −0.285 | 1,095 | 2–8,360 |
| best ai for math | 10 | −0.176 | 1,022 | 97–3,037 |
| ai for marketing | 10 | −0.091 | 2,666 | 499–6,761 |
| ai for customer service | 10 | −0.079 | 2,297 | 1,537–4,876 |
| ai for coding | 10 | −0.067 | 2,871 | 64–7,357 |
| best ai for writing | 9 | −0.067 | 5,324 | 2,999–10,411 |
| cost of living comparison | 9 | +0.050 | 916 | 123–4,672 |
| ai for business automation | 10 | +0.055 | 2,595 | 467–4,294 |
| ai for small business | 8 | +0.143 | 2,523 | 1,505–3,514 |
| best ai writing tools | 8 | +0.143 | 5,107 | 1,584–10,411 |
| best laptop for programming | 9 | +0.267 | 2,775 | 362–7,805 |
| digital nomad visa countries | 10 | +0.273 | 5,163 | 2,356–11,466 |
| best ai seo tools | 10 | +0.285 | 6,279 | 1,058–15,907 |
| ai content detector | 10 | +0.370 | 1,926 | 776–3,654 |
| ai for teachers | 9 | +0.483 | 2,509 | 863–6,720 |
| ai presentation maker | 10 | +0.612 | 1,242 | 693–5,622 |
| best ai chatbot | 9 | +0.867 | 2,599 | 403–7,662 |
← swipe the table sideways for the rest of the columns
The furthest any of the five sits from zero is Clearscope’s −0.195, on 3 keywords. A correlation that small, on that few keywords, is not a case for writing shorter either. What the five agree on is narrower and duller: inside a Google top 10, length does not order the results.
The number describes the pages that already rank; it does not rank a page. Matching it costs writing time, and none of these tables shows it buying a position.
What would overturn this
We would rather be corrected than quoted. Any of the following would change what this page says.
- A positive correlation that survives a bigger sample. Mean ρ above +0.3 across at least 10 keywords, with the sign holding keyword by keyword rather than averaged into place.
- The 30 unreadable pages, recovered. Fetch them with a renderer and recompute. They are the largest known bias in our own row, and if the sign flips once they are in, this page is wrong.
- A cleaner word count. Ours keeps navigation, footers and related-post lists, so a short article on a heavy template can outweigh a long one on a light template. Strip the boilerplate, recount, publish the result.
- Length tested against everything else. Spearman reads one column at a time, and our SERP truth shows top tens dominated by high-authority domains. Hold DR, page type and intent fixed, find length still paying off, and the reading here changes.
- A controlled test. This page describes pages that already rank; it cannot tell you what happens when you lengthen one. Publish matched pages where only length changed and the longer ones rose, and this becomes the weaker evidence.
- A mistake in our own work. The report file, the keyword fixture and the script are in the repo. Find an error in the pairing or the statistic and we will rerun it and correct the page.
Update log
| Date | Change |
|---|---|
| 2026-08-21 | Our corpus added: 20 keywords, 116 pages we read, mean ρ +0.024. |
| 2026-08-21 | NeuronWriter added: 15 keywords, 424 rows from its own table, mean ρ −0.063. |
| 2026-08-21 | Clearscope added: 3 keywords, 50 rows from its own table, mean ρ −0.195. Reference only; trial quota. |
| 2026-08-21 | Scalenut added: 8 keywords, 126 rows from its own table, mean ρ −0.123. |
| 2026-08-24 | Surfer SEO added: 20 keywords, 187 rows from its own table, mean ρ +0.063. |
Another table with a position column and a word count gets its own row, not a place in an average.
Pages carrying affiliate links say so at the top. A commission cannot move a test result: the keyword fixture is frozen, the runs are scripted, and we keep the raw responses. More on who we are and how we make money.
ToolVerdict. “Page length told us nothing about rank in our sample.” https://toolverdict.ai/findings/word-count-vs-rank. Last updated 2026-08-24.
Quote the source with the number. Our corpus reads +0.024 over 20 keywords of real Google (us) positions and 116 pages, measured 2026-08-21. The other 4 come from the tools’ own tables, and there is no combined figure to quote.