NeuronWriter Review (2026): We Scored 110 Ranking Pages — Its Content Score Didn’t Predict Rank

The app lives at app.neuronwriter.com and the unit of work is an analysis: one keyword, one country, one report. Each analysis returns a competitor table with ranks and content scores, a term list weighted by importance, and an editor that scores your draft as you write. The plan you buy is a monthly allowance of those analyses.

We ran on Bronze, $23 a month for 25 analyses and 15,000 AI credits. The API starts two tiers up, at $69 a month, so this whole test was run through the UI. Prices checked 2026-08-19.

The score is the number the product asks you to work toward, so we tested it against the thing you actually want. For each keyword we took the real Google (us) top 10 out of our ground truth, found those URLs in NeuronWriter’s competitor table, and correlated the content score it gave each page with that page’s real position. Positive means higher-scoring pages ranked higher. Near zero means the score carries no rank information.

There is an easier framing, and we report it too: the same score against the rank NeuronWriter crawled itself, over every competitor it listed. That sample is five times larger, but the tool is marking its own work and its ranks are not Google’s. We keep the two apart and lead with the harder one.

2 of the 20 keywords produced no correlation at all. For best crm for small business not one of its competitor URLs matched a page in the real top 10. For ai for coding exactly one URL matched, which is not a sample. Another 5 rest on three or four pages each, and the report marks them so nobody quotes them alone.

Drop all 7 flagged keywords and the finding survives: 13 keywords, 91 pages, mean ρ -0.079. Whichever subset you take, the answer sits on zero.

The second row is the generous reading and it still comes to nothing. 0.105 over 564 rows, with 25–30 competitors per keyword, means the score barely tracks even the ordering the tool produced itself.

The competitor tables carry two kinds of junk. Both matter if you plan to do arithmetic on that data, and neither is visible unless you export it.

10 rows were exact duplicates, the same page listed twice with identical fields. We deduplicated on the whole row, because a duplicated point is counted twice by a rank correlation and quietly shifts the result. Another 48 rows had no rank at all: video carousels, short answers and other SERP extras that the table lists next to the organic results. The own-rank framing drops those rows, since they have no rank to correlate. The Google framing keeps them, because it uses our recorded position rather than NeuronWriter’s.

NeuronWriter publishes a real term list, so this is a like-for-like measurement rather than the substitute we had to use on Frase. We took its top 30 terms by its own importance score and matched them, on unigrams and bigrams, against the top 30 terms our ground truth extracted from the body text of the pages that actually rank.

About a quarter of what it offers is a term the ranking pages share, and its list covers about a third of those shared terms. Useful as a checklist, not as a specification.

The spread across keywords is wide. Best case, ai presentation maker, recalled 0.500 of the shared terms. Worst case, best ai chatbot, recalled 0.100. Same tool, same settings, same day. Whether the list is worth following depends on the keyword, and the tool gives you no signal about which kind you are looking at.

This run now covers the same 20 fixture keywords as the Frase run, so the two precision figures rest on the same keyword set. They are still not a head-to-head: Frase publishes no term list, so its number is a stand-in built from keyword research. Same definition, same keywords, different kinds of output.

You can rebuild the ground truth yourself. Our Top 10 Term Finder runs the same extraction on any keyword you type: fetch the live top 10, pull the body text, keep the terms those pages share. Compare that against whatever your tool hands you.

We have no timing number and will not invent one. To save hours, the first round created its analyses concurrently and collected them later, which measures the batch rather than the job; the top-up round recorded no timing either. Two analyses happened to yield usable ceilings, 162 s and 186 s. A ceiling is not a median. Getting a real distribution takes a serial run, and that costs analyses we have not spent yet.

Every tool runs against the same frozen fixture and the same ground truth (serp-truth-v1 + terms-truth-v1): the live Google (us) top 10 for each keyword, the body text of those pages, and the terms those pages share. Country, language and device stay constant across tools so the numbers remain comparable. The full procedure is on our methodology page.

Bronze has no API, so we drove the app through its UI and extracted the competitor tables and term lists from the rendered pages, in two rounds: 15 keywords on 2026-08-19 and 5 keywords on 2026-08-21. The report records which round each keyword came from. The extractor changed once, during the first round: 4 captures on v1, 16 captures on v2. The later version reads more fields; both feed the same statistics, and every raw capture is archived on our side, so these numbers do not depend on NeuronWriter keeping our account data.

The scripts that produced this run are scripts/tests/content-opt/neuronwriter/{ui_capture.py,extract_analysis.js,report.py} and scripts/tests/content-opt/neuronwriter/{extract_analysis.js,topup.py}. We re-run the score-versus-rank and term tests quarterly and re-check pricing monthly; the date at the top of this page changes when we do.

First captured 2026-08-19, completed 2026-08-21, on Bronze, against app.neuronwriter.com.