Surfer SEO Review (2026): All 20 Fixture Keywords — Its Content Score Did Not Track Google Rank

The app lives at app.surferseo.com. The unit of work is a Content Editor document: one keyword, one country, one analysis. Each document returns a competitor table — ten URLs with Surfer’s crawled position, a 0–100 content score, a 0–10 domain score and a word count — a panel of recommended terms with suggested usage ranges, and an editor that grades your draft as you type. The trial let us create 20 documents for the fixture and 5 more for the re-run, 25 of the plan’s nominal 30 a month.

We ran on the Pro trial, which ended 2026-08-31, and spent $0. The API sits on a higher tier than the trial, so the whole test went through the UI. Pricing was not recorded this round; see what we couldn’t test.

The content score is the number the product asks you to raise, 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 Surfer’s competitor table, and correlated the 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.

We report that correlation twice, for the same reason as on the other reviews. Our ground truth came from Ahrefs, and some of its URLs came back cut short. A strict path match misses those rows; a lenient rule, which counts a truth URL as matched when it is the unique prefix of exactly one URL Surfer captured, rescues 11 of them here, at the risk of pulling in a same-domain page that is not the one that ranked. On this run the two rules agree on the sign; the review still shows both, because neither is the corrected value.

The third framing is the tool grading its own homework: the same score against the position Surfer crawled itself, over all 187 rows that carried one. Bigger sample, easier test, same answer.

best crm for small business produced no correlation at all: none of its 7 ground-truth URLs matched a row in Surfer’s table at the same path. Under the lenient match it gains a row, which is why that framing counts 19 keywords to the strict 18.

10 of the 20 keywords rest on four matched pages or fewer, and the report marks them so nobody quotes them alone. Drop every flagged keyword and the finding survives: 10 keywords, 66 pages, mean ρ −0.194. The per-keyword values run from -0.800 on best ai for math to 0.800 on best ai chatbot: 7 keywords positive, 11 negative. A mean that close to zero, built from swings that wide, is noise around zero, not a small effect.

Unlike Scalenut’s run, the sign holds here: strict and lenient both land below zero, a move of 0.058 between them. That does not rescue the score. On the findings page we ask how far each figure sits from zero, and none of these three gets past ~0.1.

Surfer’s competitor table carries a second number none of the other tools showed us: a 0–10 domain score per URL. It is a domain-level metric, not a judgment of the page’s text, so it is not the content score under another name and we did not fold it into anything. We did run the same correlation on it, because it was in the table.

The domain score sat further from zero than the content score, under both answer keys, and on the same side both times. We are not calling that a finding. It is 20 keywords, we ran no significance test, per-keyword values swing from −0.667 to +0.949, and +0.185 is still well short of anything we would call predictive. What the pair of rows is fair evidence for is narrower: on this sample, a number about the domain came closer to the ordering of a top 10 than a number about the page’s text. That is consistent with what every run on this site keeps showing — the things a content score cannot see are the things doing the ordering.

URLs are matched on the exact path, with the protocol, www, trailing slash and query string stripped. Same domain at a different path does not count. Of the 146 ground-truth web URLs for these keywords, 101/146 appeared in Surfer’s tables at the same path. Not every matched row could be correlated: 13 of the 200 competitor rows carried no content score, because Surfer’s own fetch of that page failed — expatistan and reddit pages mostly, the sites that block crawlers — which leaves 96 rows for the content-score correlation and 101 for the domain score. High domain overlap, 84.6% of its top-10 domains matching Google’s, hides a lower URL overlap. The two are not the same claim.

The tables needed no cleaning: 0 duplicate rows, 0 rows without a position. Surfer’s SERP snapshot and our frozen truth were also not taken on the same day, so a URL that fails to match may be a crawl difference or a SERP that genuinely turned over; this report does not separate the two.

Surfer publishes a real term panel, so this is the same measurement we ran on NeuronWriter, Clearscope and Scalenut, with one difference stated first. Those tools attach a numeric importance to each term, and their top 30 is the top 30 by that number. Surfer attaches none. Its only ranking is the order the panel displays, so our top 30 is its first 30 as shown. Both are “the tool’s own first 30”; the ordering source differs, and a comparison across tools inherits that difference.

We matched those 30 terms, on unigrams and bigrams, against the top 30 terms our ground truth extracted from the body text of the pages that rank. About one term in six is a term the ranking pages share, and the list covers a third of those shared terms. Across 20 keywords it found 200 shared terms in total.

The spread across keywords is wide. Best case, best help desk software, recalled 0.500 of the shared terms. Worst case, best ai chatbot, recalled 0.067. Same tool, same settings, same day.

And the day matters more here than on any other review, because the consistency section below measured it: re-run a keyword a day later and the panel this test scored is not quite the panel you get. Read the P@30 and R@30 above as the hit rate of the 2026-08-24 list, dated like everything else on this page.

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.

Half of it is. For 5 keywords we created a brand-new Content Editor document 25.61–25.81 hours after the first — new document, not a reopened one, because reopening only tests whether our extraction repeats, and each new document costs a trial credit. The five keywords are the same five the Frase and NeuronWriter consistency runs used, so the three measurements sit on the same words.

The scores held. Across 44 URLs present in both runs with a content score, the mean absolute change was 1.43 points, the largest 9, and 23 of them (52%) did not change at all. The domain score did not move once: 46 of 46 URLs, zero drift. Read a one-or-two-point move in either score as noise.

The term list did not hold. Overall Jaccard overlap between the two days’ recommended-term sets averaged 0.746 with a worst case of 0.524, and 3 of the 5 keywords fell below the 0.7 threshold our methodology treats as unstable. In plainer units: on cost of living comparison, 20 of the first day’s 80 recommended terms were gone the next day.

The clean cell is the one that settles where the churn comes from. On cost of living comparison, the two days’ competitor sets were identical — Jaccard 1.0, the reference frame did not move — and the term list still swapped 20 of its 80 entries, with the surviving terms shifting an average of 11 places. Nothing outside Surfer changed; the list did. The counter-case sits two rows down: ai content detector, competitor set also identical, term list Jaccard 1.0, every score to the digit. The same measurement can read “stable”, so when it reads “unstable” that is the tool, not our ruler.

Two of the five keywords (gpu comparison, ai voice generator) did see their competitor sets turn over (Jaccard 0.667), and for those the drift is confounded: a different SERP means a partly different analysis input, and we do not apportion the blame. The conclusion rests on the clean cells.

What this changes in practice: cite Surfer’s term list with a date on it. It is “the terms Surfer recommended on 2026-08-24”, not “the terms Surfer recommends”. The scores you can treat as repeatable — repeatably unrelated to rank, per the section above.

Every competitor row carries a word count beside its position, so we asked the question we ask of every table that does: do longer pages rank higher? Inside each keyword, Spearman between Surfer’s own position and its own word count, over 187 rows across 20 keywords: mean ρ +0.063, median +0.052, 11 keywords positive and 9 negative.

Both columns are Surfer’s, and its word count is undocumented, so this row sits beside our corpus and the other tools’ tables on the length-versus-rank finding as a fifth source and is never pooled with them. Five sources, five numbers, no average.

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. The full procedure is on our methodology page.

Surfer’s API sits above the trial tier, so we drove the app through its UI and read the term panel and the competitor table out of the rendered editor’s component state rather than off the page’s markup, whose class names change with every build. One document per keyword, created with the country set to United States and left otherwise at defaults. The keyword and country on each capture were read back programmatically and checked against the fixture before anything was computed — a necessary step, because Surfer’s keyword field silently autocompletes: typing one keyword and confirming carelessly can create a document for a different one. Two extractor versions ran, 1 capture on v1, 19 captures on v1.1; the device setting exists only in the creation wizard, not in the captured state, so this page treats “desktop” as an operator record rather than captured data.

Before publishing, an independent script recomputed every headline number from the raw captures without importing the run’s code, and wrote the strict and lenient score-to-rank figures to a sidecar file that this page reads. The scripts are scripts/tests/content-opt/surfer/{extract_report.js,run.py}. First tested 2026-08-24, consistency re-run 2026-08-25, on the Pro trial, against app.surferseo.com.