The SERP Teardown Method

Four research threads ran in parallel: the commercial tool market, the open source and DIY layer, the evidence on what actually moves a page into the top 10, and the evidence on search intent and SERP diagnosis. The findings converge on something uncomfortable and useful: most of what this category sells is a thin wrapper around data you already pay for, and several of the things it sells hardest are contradicted by the only controlled tests anyone has run.

You have four SEO skills installed. Here is what each one actually does:

The one sentence, quoted in full from content-writer:

That is the whole thing. Top 5, not top 20. No data pull, no scoring, no gap quantification, no link-authority read, no page-type classification, no kill criteria. It is a reminder to look, not a method.

This matters more than a normal tooling gap, because looking at a SERP and explaining how to win it is the product. It is the free Snapshot, and it is the first half of the $4,500 Sprint. The capability sitting closest to the revenue is the one with the least machinery behind it.

I ran the method against SEO consultant Boulder, geo-located to Boulder, desktop, on 26 July 2026. The result is a good demonstration of why "study the top pages and write a better one" is the wrong instinct.

Page one is not ten organic slots

It is a three-result local pack, then nine organic results, then related searches. Here is what occupies those nine:

What that actually tells you

Five of nine slots are directories and social profiles. One is somebody else's listicle. Only three are service pages. That is a strong prior against a service page, not a hard ceiling of three: a service page can displace a directory if Google's read of relevance changes. But you should plan against the prior. And Chris Raulf already holds four of the nine through his own site plus his LinkedIn, Yelp and Facebook profiles. Site diversity matters here: Google has said you usually will not see more than two results from the same site, though it is a tendency rather than a hard cap. Profiles on other people's domains do not count against it at all, so he has effectively built a moat out of third-party presence.

The incumbent page, pulled apart in one API call

One call against Raulf's /boulder-seo/ page returned the full heading tree, 5,359 words, 43 internal links, 13 external, readability indices, title-to-content consistency of 0.875, an on-page score of 95.24, and a last-modified date of 28 November 2025. His H3s are the interesting part, because they are the objection list he built the page around: We tried SEO before and got burned. Our Boulder competitors dominate Google and we can't figure out why. We hear about AI search changing everything. We don't have time to become SEO experts.

Two things to be precise about. That URL is his local-pack landing page, not one of the nine organic results above; his organic entry is the homepage. And at 5,359 words it is a long page, which sits in tension with the extraction research in section 6. Both can be true: length appears to work against being quoted by an AI answer, and it clearly is not stopping him ranking. Those are different systems, and section 6 keeps them apart.

Either way, that heading tree is the content brief, and it took one request costing fifteen thousandths of a cent.

Cost per brief, verified against live pricing pages

The DIY figure is real arithmetic, not a guess. Costed end to end at three depths:

At Standard depth, a hundred keywords costs about $13. Against a $4,500 Sprint the data cost is a rounding error, which is the real point: the entire economic argument sits in the analysis quality, not the API bill. One line item dominates and should stay behind a flag, which is pulling ranked keywords for rival domains at roughly $0.40 a keyword.

What is genuinely differentiated, and what is a wrapper

Three things in this market cannot be rebuilt on commodity data:

• Ahrefs' backlink index. Ten-plus years of accumulated crawl history. If you sell link-gap analysis, this is the one asset you cannot replicate. Note the index-size comparisons are misleading though: DataForSEO's 2.8T counts live links only, while Ahrefs' 35T counts cumulative including lost links.
• Semrush and Similarweb clickstream traffic panels. Genuinely different signal from rank times volume times CTR modelling. But useless below roughly 5,000 visits a month, which is every Front Range local business you will ever pitch.
• Screaming Frog's extraction engine. $279 a year for something nobody has bothered to reimplement.

Everything else in the content-optimizer category is the same loop: fetch the top 10, count words, extract headings, compute term frequency, render a score out of 100. Semrush's On Page SEO Checker is exactly that computation behind a $139 to $549 a month interface.

This is the part worth owning, because the evidence says most tools measure the wrong things. Fields are grouped by whether the research supports them.

Capture per SERP

• Task metadata: query, timestamp, locale, user location, device. Not optional. AI Overview prevalence swung from 6.49% to 24.61% to 15.69% inside eleven months, so an undated capture is worthless. Video carousels trigger roughly 400% more on mobile. And geo-locate the test rather than using your own Boulder IP.
• Page-type mix and its entropy. The single most informative derived field and the one nobody publishes at scale. High entropy means a mixed SERP, which means a differently-shaped page can take a slot. Low entropy means the SERP has decided what it wants.
• Site-class mix: first-party business site, directory, editorial publisher, forum, video, social profile. This is what surfaced the real story on your Boulder term.
• Median referring domains across the organic 10. The realistic entry bar, and far more useful than Keyword Difficulty. Use the median, because one Wikipedia result wrecks the mean.
• SERP feature inventory with pixel position, not just presence. An AI Overview above the fold and one at position eight are different SERPs.
• Slot availability count: how many of the ten are actually winnable given site diversity and who already owns multiple positions.

Capture per result

• Position, URL, domain, page archetype (listicle, guide, service page, product, category, forum, video, docs, tool, news)
• Title, whether Google rewrote it, H1 through H3 tree
• Word count as a distribution, never a mean to match
• Referring domains to the URL and to the domain
• Publish date, last modified, and whether the date signals conflict (a 2019 slug plus a 2026 byline is a conflict you created for free)
• Schema types present, author presence and credentials
• Whether the page is cited in the AI Overview, and at what position
• Named entities and their salience, not only heading text. Headings are a weak proxy: they can be templated, navigational, or omit material the body covers in depth. Run the body through an entity extractor and compare which entities the ranking set treats as central against which yours mentions in passing. An entity gap is usually more actionable than a heading gap.
• Link quality, not just referring-domain count. Topical relevance, local relevance, and whether the linking pages themselves rank. The leaked documentation stratifies link value by index tier, so a link from a barely-indexed page is worth close to nothing. Ten editorial links from Colorado business media are not ten directory profiles.

Do not capture, or capture only as curiosity

Nine steps, and the first four are gates: if any fails you stop, change the target, or switch to a different playbook entirely. Saying so is the most valuable thing in the deliverable.

1. Gate zero: what kind of SERP is this, and what does the query want?

   Before anything else, classify two things. Query intent (informational, commercial investigation, transactional, navigational, or local) determines which page types can rank at all and whether an AI Overview will eat the clicks. And SERP shape: is this organic-dominant, local-pack-dominant, AI-Overview-dominant, shopping, or big-publisher territory?

   This routes everything downstream. A local-pack-dominant SERP does not run this playbook at all, it runs the one in section 7. A navigational SERP for somebody else's brand is usually not winnable at any effort, and you should say so rather than sell a plan. Do not skip this because the keyword "looks" commercial. On my own Boulder run, DataForSEO classified "Boulder SEO" as navigational, and the SERP confirmed it: five of the first seven results were one competitor's brand assets.

   One caution from Google's own rater guidelines, section 12.2: "If you research the query on Google, please do not rely on the top results on the SERP. A query may have other meanings not represented on Google's search results pages." Google tells its own raters not to read intent off the SERP. It is still the best proxy available. It is a proxy.

2. Gate one: can your page type rank here at all?

   Classify all twenty results by archetype. If eight or more of the top ten are a format you cannot credibly produce, stop. If the mix is genuinely varied, a differently-typed page can hold a slot. Be careful with the strong version of this rule though, because it is folklore: BrightLocal's data shows business websites still holding 53% of the top ten for "best chiropractor" and 63% for "best vet clinic". Mismatch is a heavy prior against you, not a hard block.

3. Gate two: is the link bar reachable in twelve months?

   Take the median referring domains across the organic ten, counted per ranking URL rather than per domain, since that is the unit that tells you whether a specific page is displaceable. Then read link quality too, not just count: topical relevance, local relevance, and whether the linking pages themselves rank. Ten editorial links from Colorado business media are not ten directory profiles.

   For a site under DR 30 these are the working bands. They are calibration heuristics, not measured win rates, and the honest move is to replace them with your own outcomes as engagements complete: 0 to 9 winnable on content and internal linking alone, 10 to 19 needs a handful of real links, 20 to 29 genuinely uncertain, 30 to 49 mostly fails as a single-page attempt, 50-plus should not be chased directly. Chase the long tail underneath instead and let cluster strength accrue.

4. Gate three: can the page even be indexed and served?

   Cheap, boring, and skipped constantly. A page can pass both gates above, target a genuinely winnable SERP, and still fail on crawlability, a stray noindex, a canonical pointing somewhere else, JavaScript-only content, or simply being orphaned with no internal links pointing at it. Check indexation and serving before writing a ranking plan, not after three months of wondering. Google's AI features have the same prerequisite: the page must be indexed and snippet-eligible or it cannot be cited at all.

5. Count the slots that are actually available

   Subtract the directories, social profiles and third-party listicles you cannot displace with a page. What remains is your real target count. On SEO consultant Boulder that number is three, out of nine. Then ask the separate question the content tools never ask: can you occupy the non-page slots instead, by getting listed, profiled or included?

   And then answer the follow-up question, which is the one that actually gets it done. "Get listed on Clutch" is not a plan. Each of those platforms has its own internal ranking system, and you need the specifics: what review count and velocity moves position, how complete the profile has to be, whether there is a paid inclusion tier and whether it is worth it, and what the category and service-line taxonomy rewards. Treat every directory that holds a slot as its own small SERP with its own algorithm. For third-party listicles the mechanism is different again and simpler: the author is a person, they are reachable, and being genuinely worth including is most of the work.

6. Build the coverage gap, not the word count gap

   Take the union of every H2 and H3 across the top 20 and subtract what your page covers. That set, ranked by how many of the twenty results address each item, is the legitimate core of what content optimizers sell.

   Then add the layer nobody ships. Google confirms AI Overviews issue a fan of related sub-queries, so expand the keyword into 12 to 15 of them, pull the real SERP for five, and build a coverage matrix: twenty pages down the side, fifteen sub-queries across the top, each cell scored yes, partial or no. Every serious practitioner describes this artifact and none of them sells it. iPullRank's Qforia publishes its fan-out prompt in full, so you can lift it straight into Claude. One useful refinement from DEJAN's research: feeding the actual SERP results into the fan-out prompt substantially beats prompting from the bare keyword.

7. Find the angle nobody has

   The heading union tells you the table stakes. The differentiator is what none of the twenty have: original data, first-hand testing, named expertise, a local specific. On the Boulder SERP, not one of the nine results shows a real client outcome with numbers. You have Bone Voyage. That is the gap.

8. Write for sentence-level extraction, and aim short on purpose

   Google's grounding extraction operates on individual sentences with a heavy bias toward opening paragraphs. Front-load the answer, write self-contained sentences that survive being lifted out of context, and keep bloated tables of contents out of the main content area where they compete for extraction slots.

   Then the finding that should change your word-count advice permanently. DEJAN analyzed 883,262 grounding snippets across 7,060 queries and found the average extracted chunk is 15.5 words, that there is a roughly fixed 2,000-word grounding budget per query regardless of how many sources are used, and that the budget is split by rank: the top source gets a median 531 words, the fifth gets 266. You are competing for share of a fixed pie, not expanding it.

   And coverage collapses as pages get longer:

   Page length	Words actually grounded	Coverage
   Under 1,000 words	370	61%
   1,000 to 2,000	492	35%
   2,000 to 3,000	532	22%
   3,000 plus	544	13%

   In that dataset, grounded words rose only slightly past about 540 while the share of the page being used collapsed. For extraction, density beats length.

   Do not turn this into a word-count rule, which is the mistake I nearly made here

   This is evidence about what gets quoted into an AI answer. It is not evidence about what ranks. A lower share of a long page being cited does not show that shorter pages rank better, convert better, or serve the reader better, and the incumbent in section 3 ranks first with 5,359 words. Run the two as parallel tracks: set organic length by what the task and the winning archetypes actually require, and win extraction by building self-contained, front-loaded answer passages inside whatever length that turns out to be. A blanket 1,500-word ceiling would walk a client into a thin service page on a query that wants proof density.

   Same caution on the diagnostic: a competitor over 2,000 words is a candidate for redundancy, not proven diluted. Check its passages before claiming an opening.

9. Audit for demotions before adding boosts

   Anchor mismatch, exact-match-domain demotion, nav demotion, conflicting date signals. Checking whether you are being held down is usually higher-yield than adding another optimisation.

Label every rule with the evidence behind it

The failure mode this method is most exposed to is presenting useful hypotheses as validated rules. The fix is cheap: carry a grade next to each one, so a client can see which parts are load-bearing and which are working assumptions you intend to test. This is the same T1 to T4 scale used in the research appendix.

The two T4 and T3-heuristic rows are the honest weak points, and naming them is what separates this from a sales document. Every band in the link table should be replaced with your own measured outcomes as engagements complete, which turns the weakest row into the most defensible one within about a year.

Kill criteria, stated plainly in the deliverable

A plan that never says "do not chase this" is a sales document. Fire these:

• Eight or more of the top ten are a page type you cannot produce
• Median referring domains exceeds what you can earn in twelve months
• The SERP is majority-owned by Amazon, Wikipedia, Reddit, YouTube, or government properties
• An AI Overview sits on an informational query with no transactional follow-through, so you would be competing for a sub-1% click-through slot
• The page has been live and crawled for six months without entering the top 50

Everything above is an organic playbook. For a Front Range service business it is frequently the wrong one, and the demonstration query proves it: SEO consultant Boulder puts a three-result local pack above every organic result, and the method as written walks straight past it.

What actually determines pack position

Three inputs, and only one of them is on your website.

• Proximity to the searcher, which is the single strongest factor and the one you cannot change. This is why a single rank check from the city center is nearly meaningless. A business can sit at position 2 outside its own front door and vanish four miles away.
• Prominence, which is review count, velocity, recency and content, plus how much the wider web talks about the business. This is where the work is.
• Relevance, which is the profile itself: primary and secondary categories, services, attributes, description, photos, posts, and the local landing page it points at.

The audit, in order

1. Eligibility and verification first

   Is the profile verified, unsuspended, and eligible for the category at all? Service-area businesses have different rules from storefronts. This is the gate, and it fails more often than people expect.

2. Primary category, then secondary

   The primary category does most of the work and is the highest-leverage single field in the whole profile. Pull the primary category of every business in the pack and compare. If all three competitors share a category the client does not have, that is usually the finding of the engagement.

3. Measure on a grid, never at a point

   Because proximity dominates, a single coordinate tells you almost nothing. Sample rank across a grid of coordinates spanning the actual service area and report the map, not the number. The honest deliverable is "you rank 2 to 4 within two miles and drop out entirely past five," which is a completely different conversation from "you rank 3."

4. Reviews as a rate, not a total

   Count, average rating, velocity over the last 90 days, recency of the most recent, whether the owner responds, and whether review text mentions the services you want to rank for. Compare against the pack incumbents rather than an abstract target. A competitor with 28 reviews is a very different problem from one with 400.

5. NAP consistency and the citation set

   Name, address and phone consistency across the directories that matter in the market. This overlaps directly with the directory work in the organic plan, which is convenient: the same Clutch, Thumbtack and Yelp placements that occupy organic slots also feed prominence here. One effort, two systems.

6. The landing page the profile points at

   This is the only place the organic and local playbooks genuinely converge. The page the profile links to should be locally relevant and match the primary category. Note that this is exactly what the Boulder incumbent does: the page I pulled apart in section 3 is his local-pack landing page, not his organic entry.

7. Check for competitor spam, and decide whether to report it

   Keyword-stuffed business names, fake addresses, and lead-gen listings are common in local service categories and they are reportable. Sometimes the highest-return action in a local engagement is a redressal filing rather than anything you build.

The method above ends at arrival, which is a real gap: a page that reaches position 8 and falls to 15 within six weeks is a failed engagement, not a successful one. Two numbers set the stakes.

The maintenance protocol

• Re-pull the full teardown quarterly, and after every confirmed core update. Not just the rank. The thing that changes and gets missed is composition: a directory replacing a service page, an AI Overview appearing, a local pack being introduced. Those change the plan, and rank tracking alone will never show them.
• Watch page-type mix as the leading indicator. If the archetype mix shifts against you while your rank holds, that is your early warning. Rank is the lagging signal.
• Track the fan-out sub-queries separately, because losing those is invisible in head-term rank tracking and it is where the citation surface lives.
• Set a decay trigger. Define in advance what prompts a refresh: a drop of more than three positions sustained over two weeks, a competitor entering with fresher content, or a claim in the piece going stale.
• Keep the date signals honest on every refresh. Updating the byline without updating the content is a signal conflict, and the leaked documentation carries three separate date fields that are expected to agree.

Two deliverables, not one

This section is worth more than the rest combined, because every competitor in Boulder is selling at least two of these.

And one thing worth selling that almost nobody is

The finding hiding in plain sight, and the best sales argument you have

The Princeton GEO paper is widely cited for its headline, that adding statistics, quotations and cited sources lifts visibility by up to 40%. Almost nobody quotes its third table, which breaks the same result out by where the page already ranked:

The depth parameter does not do what you expect

Asking DataForSEO for depth 20 returns page one only, because max_crawl_pages defaults to 1. Depth also multiplies cost per ten results, and each search operator such as site: multiplies the task cost by five. Budget accordingly.

Any trend line crossing September 2025 is contaminated

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. Use October 2025 as the new baseline, and if a client shows you a chart with a cliff last September, that is what it is.

1. This week: OpenSEO against your existing key

   Free, MIT licensed, self-hostable on Cloudflare's free plan, bring-your-own DataForSEO key, ships its own MCP server plus prebuilt Claude Code skills. Roughly 8,100 stars and actively maintained. An afternoon of setup for keyword research, rank tracking, competitor insight, backlinks, site audit and AI visibility. Highest leverage item in this entire report.

2. This week: Screaming Frog, $279 a year

   The cheapest deep page extraction that exists, and version 24 ships a built-in MCP server. Two caveats: it cannot fetch a SERP at all (its "SERP mode" is a snippet pixel-width preview, not results retrieval), so you feed it URLs from DataForSEO. And list mode is not yet supported over MCP, which kills the obvious "hand Claude twenty competitor URLs" workflow for now.

3. Next sixty days: rent Thruuu at $33 a month

   Reads up to 100 results, has both a SERP API and a Brief API, costs $0.47 a brief. Use it to find out whether your brief format is right before spending fifteen hours building one. If it turns out you need things Thruuu does not carry, which it will, you will then build the right thing instead of the obvious thing.

4. Then: build the SEO Believer teardown skill

   Around the schema in section 5 and the seven steps in section 6. The reason to build is not cost. It is that you own the prompt, so the deliverable carries your methodology and your voice rather than a vendor's Content Score, and you can fold in link authority, keyword intersection and AI Overview citation data that no per-brief tool exposes. It is also the only thing here that makes SEO Believer its own proof of concept while every competitor resells the same eight dashboard logos.

5. Only when a client needs it: Ahrefs Lite at $129

   For the backlink index, which you genuinely cannot rebuild, and because clients recognize the DR number. Its MCP is included on every paid tier. Do not stack Content Kit and Brand Radar on top, which takes it to $427 a month for things you can build.

6. Rebuild keyword-targets.md on measured data

   Every number in it is currently an estimate, by its own admission. Run the real volumes, difficulties and SERP compositions, and record which Tier 1 terms fail the two gates in section 6. Some of them will.

The build stack, if and when you build

• SERP and page data: the DataForSEO client and MCP already wired in. One vendor, real positional rank, city-level geo-targeting, AI Overview citations from the same call.
• Extraction: trafilatura as the default, crawl4ai when you need JavaScript rendering. Deliberately not Firecrawl if any of this ever becomes a hosted client-facing tool, because it is AGPL-3.0 and the network clause bites.
• Semantic layer: fastembed with 256-dimension truncation. No PyTorch dependency, which is a materially better Windows install story.
• Steal one design decision: record the API cost of every report and cache aggressively so you never pay twice for the same SERP. That single habit is what makes a per-keyword pipeline safe to run at scale.
• Read but do not install SE Ranking's open-source seo-content-brief skill. It is the most thoroughly specified brief structure published anywhere, and it is a better starting spec than anything you would write cold.

If you would rather buy something agent-drivable

Only two tools in this entire market can be driven from Claude Code today. Frase ships an official MCP server that is pushed to daily, plus 50-plus documented REST endpoints, from $39 a month. NeuronWriter ships an official MCP server and a public six-method API at $57 a month yearly, and its evaluate-content method scores a draft without consuming credits, which is the cheapest correct scoring primitive available anywhere. NeuronWriter also happens to be the one tool that scored best in Ahrefs' independent cross-tool test. Surfer's own API is locked to the $299 tier and its MCP is announced with no date.