Frase Review (2026): We Scored 38 Ranking Pages — Its Score Didn’t Predict Rank

The current app lives at next.frase.io. Work goes one way: a research run gathers keywords, sub-topics and SERP results, you derive a brief from that research run, you generate a draft from the brief, and an audit scores a URL or a pasted draft from 0 to 100. You cannot feed a keyword straight into a brief, and you have to add a site to the workspace before research will run at all.

The 7-day trial asks for no card and includes 5 articles, 30 research runs, 50 audited pages, 250 API requests. Paid plans are Starter at $39 per month billed annually ($49 month to month) and Pro at $103 per month billed annually. Prices checked 2026-08-18.

The score is the thing Frase asks you to act on, so we tested it directly. For each keyword we took the real Google (us) top 10, sent every URL to Frase’s page audit with that keyword attached, and recorded the SEO score it returned. Within each keyword we computed the Spearman correlation between the score and −position. A positive value means higher-scoring pages rank higher. A value near zero means the score carries no rank information.

Knowing a page’s Frase score told us almost nothing about where it ranked. A page scoring 80 was no more likely to outrank a page scoring 40 than the other way round.

For ai presentation maker the correlation ran strongly negative, at -0.747. The pages ranking first and second on Google, Canva’s and Adobe’s own product pages, both scored 16 for SEO. Frase extracted 24 and 35 words of body text from them. Pages ranking fifth and eighth scored 80 and 83. The score is reading text volume and on-page structure. A brand landing page has little of either. A long affiliate listicle has plenty.

The direction is the finding here, not the exact coefficient. With 6–8 pages per keyword, any single keyword’s ρ is noisy. What survives the noise is a mean near zero and an unstable sign, so we have no evidence the score predicts rank.

Half of it does. We sent 5 pages back through the audit 25.0–25.1 hours after the first run: same URLs, same keywords, same endpoint. The SEO score came back identical on 5 of 5. The GEO score moved on 3 of them, and elevenlabs.io lost 9 points.

Read that next to the rank test. The SEO score repeats itself exactly and still doesn’t track rank, at a mean ρ of -0.115. It is stably wrong, and stable is worth something: two runs on the same page agree, so when the number moves you know you moved it. The GEO score can’t claim that much. Optimize a page against it and the same page can read 9 points lower the next day.

One URL per keyword: for each keyword the rank test covered, we re-audited the top-ranked page that had returned a score. That is 5 pages, not the 20-keyword fixture the rest of this page runs on.

So use the SEO score if you use either, and read a small GEO move as noise rather than progress.

Our standard test compares a tool’s recommended terms against the terms the current top-10 pages actually use. Frase publishes no such list, so we substituted the closest thing it does return: the top 30 entries of its research keyword output, matched against our ground-truth top 30 terms on unigrams and bigrams. That compares keyword research against term coverage, which is not the same job. A low number here mostly means Frase does not do this job, not that it does it badly.

The spread is wide. Best case, digital nomad visa countries: recall 0.400, precision 0.188, with country names and visa phrasing coming through. Worst case, ai for customer service: precision and recall both 0.000. Not one of its 40 suggested keywords matched a high-frequency term on the pages that rank.

The two columns above show how it fails. On gpu comparison, 29 of 30 suggestions contain a term the ranking pages share. But nearly all of them contain the head word, so between them they cover only 13 of the 30 terms those pages actually use. The suggestions are mostly restating the query back at you.

The SERP overlap of 60.8% has a mechanical cause. Frase pulls up to 20 results and reports ranks past 10, so forum and retail URLs such as reddit.com and newegg.com enter its top-10 view of a SERP where Google’s actual top 10 has none. Check that against a real SERP before you trust its competitor set. Two keywords sat at the extremes: digital nomad visa countries matched 8 of 8, best crm for small business matched 1 of 7.

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.

We generated 5 drafts through the full pipeline at the API’s minimum target of 500 words. They came back at 690–905 words in 64.8–106.5 seconds. From those drafts we pulled 20 checkable factual claims and tried to trace each one back to a source.

The error was a number. The draft stated that Adobe Express turns a text prompt into a finished presentation in under two minutes. Adobe’s own page says “in minutes” and carries no such figure, so the two came from nowhere. Two of the unverifiable claims have the same shape: a specific-sounding range with nothing behind it.

All 5 drafts share one skeleton: three H2 sections, the last always titled “The Bottom Line”, each section opening with a bolded sentence. The target keyword goes into running prose verbatim, which reads badly. Every outbound link points at a page the research step had already fetched.

Frase scored its own drafts 75–81 for SEO and 69–74 for GEO. The pages that actually rank first for ai presentation maker scored 16.

The serial queue is the real constraint. One research run takes about a minute, which is fine. Twenty submitted at once do not run in parallel, so a batch of keywords is a coffee break.

Next up: NeuronWriter, run against the same 20 keywords and the same ground truth.

Every tool we test runs against the same frozen fixture of 20 keywords (keywords-v1) 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 are held constant across tools so the numbers stay comparable. The full procedure is on our methodology page.

We drove this run with a script rather than by hand, so the same commands reproduce it. We keep the raw API response for every audit, research run and draft on our side, so these numbers do not depend on Frase keeping our trial data. 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.

Last tested 2026-08-18, on the 7-day trial, against next.frase.io.