SERP analysis tooling: the full research

Four threads ran in parallel: the commercial SaaS market, the open-source and DIY layer, the evidence base for what actually moves a page into the top 10, and the evidence on search intent and SERP diagnosis. Every price below was fetched from the vendor's own live pricing page on 26 July 2026, not from a review site. GitHub star counts and last-commit dates came from the GitHub API, package versions from the npm and PyPI registries.

Claims are graded, because most of what circulates as SEO fact is not:

This is the category that sells "analyze the top pages and tell me what to write." Two structural things happened in the last year: Surfer was acquired by Positive Group (announced October 2025, now branded Positive Surfer), and MarketMuse was absorbed by Siteimprove (completed 31 October 2024). The category has also split, with Frase and NeuronWriter going agent-native while Surfer, Clearscope and Scalenut pivoted their messaging to AI-visibility tracking. Nobody leads with "TF-IDF content editor" anymore.

What they actually measure

Surfer analyses the top 20 by default and now splits its Content Score into an SEO Score (keywords, NLP terms, "true density," heading structure, length versus SERP, images) and an AI Search Score (facts coverage, whether the page answers the primary question early). Notably its own documentation warns against the product's own headline metric: aim for 70 to 85, because "over-optimizing can actually hurt your final content."

Clearscope is refreshingly plain about its mechanism: it "calculates the importance of each term by how much the keyword appears in the competitors' articles." That is term frequency, stated by the vendor. It has the best practitioner reputation in the category for restraint, and a genuinely novel feature in AI Term Presence, which flags which suggested terms actually appear in Gemini and GPT answers for the query.

Page Optimizer Pro is the most differentiated and the least fashionable. It is the only tool here that pulls competitors' schema and compares it against yours, and it runs competitor content through the Google Cloud NLP API to show entities and categories. Its epistemology is single-variable live testing rather than correlation, which is a more serious claim than the others make and also the least independently verifiable, since the tests are not published.

Frase is the only one scoring author and E-E-A-T signals as a first-class metric, and its documentation is unusually honest: "Scores are feedback, not goals. Don't chase perfect 100s. Focus on beating competitor averages."

Is the content score defensible?

Short answer: it is a real but weak correlation, computed close to circularly, with no causal evidence in either direction.

• Surfer's own study (10,000 queries, top 100 results each, one million SERP entries, published July 2025): Content Score correlates with rankings at Spearman 0.28. By intent it ranges from 0.19 to 0.30. That is roughly 7 to 8 percent of rank-order variance.
• The circularity problem. The score is computed from the pages currently ranking in the top 20. Correlating it against current rank is close to correlating a variable with its own input. Under that construction 0.28 is arguably low.
• The only independent head-to-head is Ahrefs across five tools. Verdict: "weak correlations everywhere," with NeuronWriter scoring best. Their most damaging line, from inside the category: "in some tools, you can literally copy-paste the entire list, draft nothing else, and get an almost perfect score." Caveats worth carrying: only 20 keywords, Ahrefs sells a competitor, and they excluded Reddit, Quora and YouTube because the tools scored them zero, which removes exactly the results that most embarrass the category.
• No controlled split test of a content score exists publicly. Not from Kevin Indig, Aleyda Solis, Cyrus Shepard, Mike King, Lily Ray, Authority Hacker or anyone else. Neither does any test-retest reliability study, which matters because the benchmark is a moving SERP.

Semrush was acquired by Adobe, completed 28 April 2026, all cash at $1.9 billion. No price rises yet. The warning worth carrying is Cyrus Shepard's: Adobe's tooling leans enterprise, which historically means drift away from small operators.

Accuracy limits that matter for Boulder work

Independent testing puts Semrush traffic estimates as reasonable above 50,000 organic visits a month and "not accurate at all" below 5,000, with a 30 to 60 percent under-report on small sites. Ahrefs' estimates swing between minus 80 and plus 80 percent on newer domains. For local service businesses, treat all third-party traffic estimates as directional only and lean on Search Console.

One index-size correction, since the marketing comparison is misleading: DataForSEO's 2.8 trillion links counts live links only, while Ahrefs' 35 trillion and Semrush's 43 trillion count cumulative including lost links. The gap is real but nowhere near twelve-fold.

Screaming Frog, and an important limitation

Version 24.3, £199 per year (shown as $279). It cannot fetch a Google SERP. Its "SERP mode" is a snippet pixel-width preview where you upload titles and meta descriptions, not organic-results retrieval. You must source the top-20 URL list elsewhere.

What it does superbly is list mode plus custom extraction: paste twenty competitor URLs and pull headings, word counts, schema blocks, author bylines and dates via XPath or regex. Version 24 shipped a built-in MCP server, but with a catch confirmed by their own staff: list mode is not yet supported over MCP, which kills the obvious "hand Claude twenty competitor URLs" workflow for now.

Thruuu, the buy-versus-build spoiler

Analyses up to 100 results, extracts headings, word counts and on-page data, covers AI Overviews and AI-engine citation sources, and ships both a SERP API and a Brief API. Starter $13/mo yearly, Pro $33, Agency $66, with published per-unit prices of $0.47 per brief at the Pro tier. A 100-result brief at 47 cents with an API is close enough to DIY cost that it is the honest comparator for a build decision, not Surfer.

Sitebulb and Rank Math

Both are the wrong category for this job. Sitebulb has no SERP retrieval, no competitor analysis and no brief; it is a site crawler with the best plain-English audit output on the market. Rank Math's Content AI surfaces suggested keywords but exposes no top-N breakdown, and its credits are metered at one credit per word generated with keyword research costing 500 credits per operation.

DataForSEO, verified per-call costs

Blunt finding: not one of them does top-20 ranking-page analysis or produces an outrank plan. The category runs on a different primitive entirely, which is firing prompts at LLMs repeatedly, parsing the answers for your brand and the cited URLs, and charting it.

Why the whole measurement category has a statistics problem

Rand Fishkin and Patrick O'Donnell ran 12 identical brand-recommendation prompts 2,961 times through 600 volunteers across ChatGPT, Claude and Google AI. The odds of the same brand list appearing twice were under 1 in 100. The same list in the same order: under 1 in 1,000. What survived as a stable measure was membership in the consideration set, expressed as a visibility percentage across many runs, never a rank.

The variance stacks in ways nobody discloses. Querying ChatGPT via the official API returns an average of 7 sources while scraping the actual web interface returns 16, with only 8% source overlap on Perplexity between the two collection methods. Google's own AI Overviews and AI Mode cite the same URLs only about 13.7% of the time. Roughly 11% of domains are cited by both ChatGPT and Perplexity. Every dashboard in this category renders a single draw from that distribution to two decimal places.

My read: the measurement layer is a bubble consolidating into incumbents and DXP vendors, while the execution layer is not. Profound pivoting into Agents and Evertune into a ChatGPT Ad Agent are both admissions that dashboards will not sustain those valuations. Structurally, Ahrefs bundling prompt tracking at $129 and Semrush at $199 caps what a standalone tracker can ever charge.

The projects actually worth knowing

Content extraction: the 2026 answer

trafilatura (v2.1.0, June 2026) is the default choice, precision-oriented and admitting only about 6.6% boilerplate. rs-trafilatura scores highest on the current benchmark and is fastest at 44ms per page, but carries real bus-factor risk. crawl4ai (75,000 stars, Apache-2.0) is the right pick when you need JavaScript rendering or schema-driven extraction.

Two licensing flags for a consultancy

• Firecrawl is AGPL-3.0. Excellent scraper, but the network copyleft clause bites if you ever expose a hosted client-facing tool built on it. Fine for internal use, dangerous for a productised service.
• Several useful repos carry no license at all, including searchsolved and seo-audits-toolkit. Read them for method, do not vendor the code.

The free diagnostic nobody knows about

DEJAN's Grounding Snippet Extraction Tool uses Gemini's live search grounding and returns the exact URLs and exact sentences Google actually extracted for a query. Every other tool in this survey guesses at retrievability using a public embedding model. This one shows you the real output, and it is free. It is the closest thing to ground truth available to anyone outside Google.

The job has two halves: get a positional SERP, and get the ranking pages in analysable form. Most tools do one.

Among the paid content tools, only Frase and NeuronWriter ship official MCP servers today. NeuronWriter's evaluate-content method is worth singling out: it scores a draft without consuming an analysis credit or creating a revision, which makes it the cheapest correct scoring primitive available anywhere if you want a tight iteration loop.

Clicks are a ranking input, and this is now settled under oath

T1 From DOJ trial exhibits, Google's own engineers describe topicality as built from ABC signals: Anchors, Body, and Clicks, where clicks means how long a user stayed on a linked page before returning to the SERP. That is dwell time in all but name. Ex-Googler Eric Lehman on NavBoost: "it's just a big table. It says for this search query, this document got two clicks." Trained on 13 months of rolling data.

Pandu Nayak, asked directly by the judge where NavBoost sits in importance: "navboost is important, right. So I don't want to minimize it in any way. But I will also say that there are plenty of other signals that are also important." He ranked the document itself first, then topicality, page quality, reliability, localization, then NavBoost.

Site-level authority exists, and Google's denials were semantic

T1 The leaked Content Warehouse documentation contains siteAuthority. Google spokespeople had denied domain authority for years. The DOJ deposition is blunter still, on the page-quality signal: "Q is largely static and largely related to the site rather than the query."

This is the mechanical reason a low-authority site underperforms on merit. A static, site-attached quality score gates you before page content is even assessed. Combined with homepagePagerankNs being copied onto new pages as a proxy, and NavBoost having no table entry for a page with no click history, a new page on a weak domain is not merely unranked, it is unmeasured.

Internal linking is the best-evidenced controllable lever

T2 SearchPilot's controlled split tests, which use control and variant groups and so rule out seasonality and algorithm updates:

• Adding nearby-location links across ~8,000 regional pages: +7% organic traffic to pages receiving new links.
• Expanding the homepage footer with more internal links: +5% to destination pages.
• Moving city and clinic names to the front of title tags on a healthcare site: +8.5% at 95% confidence.
• Removing a Google Map component from location pages: minus 7%. Practitioners polled were evenly split and mostly expected an uplift, which is a good reminder that intuition is wrong about half the time here.
• Adding a smaller secondary intent to a title tag on a page with a dominant primary intent: minus 10%. This is the controlled evidence against "cover both intents."

Zyppy's 23-million-link study found traffic rising with incoming internal links up to roughly 45 to 50, then reversing, explained by sitewide navigation links pointing at low-traffic URLs.

Links, freshness, page experience

• Links are no longer top three, per Gary Illyes in 2023, and the leak shows link value stratified by index tier, so a link from a barely-indexed page is worth close to nothing. Judge link quality by whether the linking page itself ranks and gets traffic, not by DR.
• Link velocity is not a positive signal. Ahrefs tested it across 10,000 SERPs and states they "failed to distill any observable relationship." It appears in the leak only as a spam signal.
• Freshness is query-class-conditional, not universal, and the age data cuts the other way for most queries: 72.9% of top-10 pages are more than three years old and the average number-one page is five years old. What does matter is that your date signals agree, since the leak shows three separate date fields and a 2019 slug plus a 2026 byline is a conflict you created for free.
• Core Web Vitals, from Google's own FAQ: "trying to get a perfect score just for SEO reasons may not be the best use of your time." Hygiene, not strategy.

The base rates every client should see

Page-type mix, and the "best" modifier

BrightLocal classified the first ten organic results for local-intent terms. Overall: business websites 47%, directories 31%, business mentions 16%, forums 7%. The per-keyword breakdown is the useful part, because one word flips it:

Note this data also refutes the strong form of the content-type rule. Business websites still hold 53% for "best chiropractor" and 63% for "best vet clinic." Mismatch is a heavy prior against you, not a hard block. Google's own rater guidelines instruct raters to reward format diversity, and to treat a more-specific brand page as Highly Meets for a broad category query if the brand is popular and prominent. The operative variables are prominence and specificity distance, not format.

The overlap question, and why the answers disagree threefold

These are not measuring the same thing. Ahrefs counts SERP blocks including features, BrightEdge counts organic-only overlap, and the cross-engine figure covers non-Google assistants. Ahrefs states plainly that its own parser improved between its two studies so they are not directly comparable, and that AI Overviews moved to Gemini 3 in January 2026. Anyone quoting "AI citations collapsed from 76% to 38%" as a trend is over-reading the vendor's own caveat.

The grounding-budget study, the most commercially useful finding

DEJAN analyzed 883,262 snippets across 7,060 queries and 2,275 tokenised pages:

• Average grounding chunk: 15.5 words. Extraction is extractive, not abstractive, with non-contiguous sentences joined by ellipses.
• A roughly fixed 2,000-word grounding budget per query, remarkably consistent regardless of source count or page length.
• The budget is split by relevance rank: source one gets a median 531 words, source five gets 266. You compete for share of a fixed pie.
• Coverage collapses with length: under 1,000 words gets 61% grounded, 3,000-plus gets 13%. Grounding plateaus around 540 words no matter how long the page is.
• Heavy positional bias: opening paragraphs are extracted almost wholesale. Structural noise like tables of contents gets ingested as if it were prose, competing for slots.

Petrovic's conclusion, and it contradicts what the industry still sells: density beats length.

The GEO paper, including the table nobody quotes

The Princeton GEO paper (KDD 2024) is cited everywhere for "up to 40% visibility improvement." Its measured lifts on Position-Adjusted Word Count: quotation addition +41%, statistics +31%, fluency +28%, citing sources +27%, and keyword stuffing at minus 8%, worse than doing nothing.

Table 3 is the one that matters, and almost nobody quotes it:

These tactics roughly double visibility for the weakest source and actively hurt the strongest. It is a leveling effect. Caveats: the engine was a researcher-built pipeline over 2023-era results, not any production system; "visibility" means share of words in the answer, not clicks; and no independent replication exists on 2026 engines.

Schema: correlation says yes, the controlled test says no

T3 Across 6 million URLs, AI-cited pages were almost 3x more likely to have JSON-LD. That is the stat on the conference slides.

T2 Then Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched control pages: Google AI Overviews minus 4.6% (small but statistically significant), AI Mode +2.4% and ChatGPT ~0, neither significant. Four separate statistical tests pointed the same way.

Ahrefs publishes its own limitation: every page already had 100-plus citations, so this tests whether schema lifts an already-cited page, not whether it helps an invisible page get discovered. Separately, searchVIU tested five assistants during direct real-time fetch and found none of them used JSON-LD, hidden Microdata or hidden RDFa; all extracted visible HTML only. Keep schema for rich results, entity clarity, and Bing, which has been explicitly more schema-friendly. Do not sell it as an AI citation lever.

Brand mentions beat links, by a lot

Ahrefs' correlations across 75,000 brands: YouTube mentions 0.71 to 0.74, branded web mentions 0.66 to 0.71, branded anchors 0.527, Domain Rating 0.326, referring domains 0.295, backlinks 0.218, and number of site pages 0.17, which is effectively no relationship between content volume and AI visibility. Mention frequency beat mention reach, so twenty small creators may register more than one viral video.

The methodology flag Ahrefs discloses: brands were defined as domains above DR 40 whose top keyword has 800-plus monthly volume. That filter selects for large established brands and almost certainly inflates the correlation between "is a big brand" and every other variable. Directional, not a coefficient to plan against.

Is AI citation even worth chasing?

Be honest with clients about the economics. Pew Research, using metered browsing behavior from 900 US adults rather than vendor SERP scraping, found users who encountered an AI summary clicked a traditional result in 8% of visits versus 15% without one. Seer Interactive found being cited lifts click-through from roughly 0.6% to 1.08%, which is a real doubling of a very small number. And Google still sends roughly 345 times more traffic than ChatGPT, Gemini and Perplexity combined.

AI citation is a brand and defensive play. It is not yet a traffic channel.

Claims checked against primary sources and found wanting. Several of these are actively sold in Boulder right now.

A data break that contaminates trend lines

Google disabled the num=100 URL parameter around 10 to 12 September 2025. Rank-tracking bots had been generating impressions across positions 1 to 100 in a single request, so removing it removed the bot impressions. Across 319 analyzed properties, 87.7% of sites lost Search Console impressions and 77.6% lost unique ranking terms. Clicks were unaffected, average position improved arithmetically, and click-through rate rose mechanically.

Any longitudinal analysis crossing that date is contaminated. Use October 2025 as the new baseline. This is also a free credibility play, since most agencies quietly let clients believe that September cliff was somebody's fault.

Assumes one keyword, top 20 pages, DataForSEO pay-as-you-go, cheap models for the language work.

Three configurations: lean at $0.02 (enough for the free redacted Snapshot), standard at $0.13 (the right depth for a paid Sprint), full at $0.58 (only for commercially load-bearing terms). A hundred keywords at standard depth costs about $13.