Auditing the “Ge Fei” playbook for reproducibility: what earns AI-written content an index entry

Why audit a methodology that “works”

The “Ge Fei” playbook circulates widely among Chinese indie developers. Its core moves: pick keywords from search demand, derive the site from the keywords, ship fast instead of perfecting, scale content across pages, cold-start through community and links, then iterate on data.

It works. Plenty of sites got traffic from it.

But “works” and “is reproducible” are two different things. A method can work without being reproducible (it depends on the executor’s tacit judgement, the time window, luck), or be reproducible only for certain categories. In success stories the two look identical; only by taking them apart can you tell.

This essay splits the playbook into five falsifiable hypotheses and adjudicates each: which parts are a reproducible chain, which parts are an unreproducible result. The subject is ourselves — the 2026 China native multimodal LLM tier list was built on this method, so we are both auditor and audited.

The conclusion first: of five hypotheses, three hold, one holds partly, one fails. The one that fails is the most widely repeated and the most taken for granted.

Five hypotheses and verdicts

#	Hypothesis	Verdict	Basis
H1	Keyword-first (fix the search term, then decide what to build) beats content-first	Holds	Reproducible
H2	Shipping speed beats completeness	Holds	Reproducible
H3	Page scaling raises total indexation	Holds (bounded)	Reproducible, but marginal returns are positive only under a specific structure
H4	Unique content value can be proxied by “is it recomputable”	Partly holds	The proxy is effective but incomplete
H5	Links and community promotion are necessary for cold start	Fails	Not reproducible, and directionally wrong under 2026 judgement logic

H1 keyword-first: holds, fully reproducible

The hardest of the five. Its reproducibility comes from one fact: the search term is externally given, not author-imagined.

Our execution path replays step by step:

1. Fix the category: Chinese domestic multimodal model comparison
2. Check existing Chinese coverage: many single-model reviews, few comparisons; zero comparisons including the October 2032 models
3. Derive the query-intent gap: a long query with a year, a geographic qualifier and ranking intent
4. Set the head keyword: 2026 中国国产原生多模态大模型对比排名天梯榜
5. Derive the site structure backwards: a tier list page (carries the head term), 32 model dossiers (carry “how is X”), 10 comparison pages (carry “X versus Y”), 15 essays (carry question-shaped long tails)

Step two is the key, and the most skipped. Most failed content sites did not pick the wrong term; they never validated the gap — they found a high-volume term, discovered the top twenty are all authority sites, wrote a better article anyway, and nothing happened.

The reproducible check: put the head keyword into search and see whether any of the top twenty results structurally satisfies the query. If yes, change the term; if they merely brush against it, that is the gap.

H2 ship-speed-first: holds, because it turns hypothesis into observables

This is often misread as “rough is fine”. Its real value is not speed: only shipped things produce falsifying data.

This site went from raw data to all 144 pages live in one shot, with no staged rollout. Not because of a deadline, but because staging contaminates the experiment: if dossiers go live first and essays later, the essays’ internal link equity, crawl order and click distribution are all altered by the dossiers already existing, and you never learn what an essay does on its own.

The reproducible part: treat the whole site as one experiment delivery, not each page as a delivery. The cost is that the structure must be frozen before launch; the benefit is clean data.

H3 page scaling: holds, but the boundary is hard

“More pages, more indexation” held in 2020; in 2026 it holds only under a specific structure.

Within our 144 pages, the 42 data pages to 20 content pages ratio is not arbitrary:

• Data pages are irreplaceable by construction. Each of the 26 dossier pages carries that model’s 8-dimension scores, benchmark figures, scenario scores and source links. Those numbers are published here for the first time in structured aggregate. Facing such a page, the search engine has no “we already have something much like this” option.
• Content pages are replaceable by default. An essay on “how to choose a Chinese multimodal model”, without extra work, is indistinguishable from the two-thousandth of its kind.

So H3’s accurate form: what scales must be data pages; content pages need a separate case for existing. Every essay here is bound to a data artefact (see the 12-row artefact table in the cold-start experiment), which is essentially converting content pages into half-data pages.

Read H3 as “write more articles” and it fails. Read it as “publish more irreplaceable structured data” and it holds.

H4 proxying unique value with recomputability: partly holds

This is our own addition and the site’s core operating discipline: a page with a recomputable artefact that cannot be derived elsewhere has unique value.

Where it holds: the proxy is executable. “Unique value” is vague and writers self-justify; “can this number be looked up directly elsewhere” is binary and hard to self-justify against. It cut 30-plus candidate topics down to 12 here.

Where it is incomplete, three points:

One, recomputable ≠ wanted. A question nobody searches has no traffic even if its answer is unique. Both why cold models survive and the 8-class cold Chinese document mini board carry this risk — they are indeed the only ones on the web, but “which multimodal model is strongest at industrial QC” may have single-digit monthly searches. This is deliberately accepted: cold but unique beats popular but duplicated.

Two, the recomputability bar is falling. We can run 9-dimension strict dominance in the Pareto essay because we hold all eight dimensions. Anyone with our public data can run it too. The artefact’s moat is not the algorithm but the collection cost of the raw data — the human judgement behind 32 models × 8 dimensions = 256 score points, plus the weekly updates.

Three, recomputability cannot cover timely content. A breaking model release is valuable because it is fast, not because it is recomputable. The frontier wave is an attempt at a hybrid: using a recomputable artefact (the lineage delta table) to carry a timely topic — the release news exists elsewhere, the per-dimension deltas against same-family predecessors do not. Whether the hybrid works waits for the indexation data.

So H4 is judged partly holding: an effective filter, not a sufficient reason to exist.

H5 links and community promotion are necessary for cold start: fails

The most widely repeated one, and the one we rebut head-on.

Its origin is understandable: in the traditional SEO model, domain authority (PageRank and successors) is the core ranking input, and domain authority is mostly links. A new site without links has no authority, no ranking, so links must come first. The chain is internally consistent and was empirically true for a decade-plus.

But under 2026 judgement logic it points the wrong way. The reason returns to the distinction at the top of this essay: a search engine’s first filter for a new page is not “how authoritative is this site” but “is this page worth indexing”. The two are sequential:

• Fail the first → status “Crawled – currently not indexed” → more links change nothing, because the page is not in the candidate pool
• Pass the first → ranking competition begins → domain authority starts to matter

Links act on the second gate. Most new sites die at the first. Buying links for a “crawled – not indexed” page spends money on a stage that has not been activated yet.

That is the real reason this site runs zero links and zero promotion — not principle, but clear the first gate before worrying about the second. The mechanism is in the cold-start experiment: 144 pages, one artefact each, two pre-registered falsification conditions.

But H5’s refutation is conditional and we must say so: if this site’s indexation rate is below 40% after 90 days, then H5 is right instead — the first gate cannot be cleared on unique content value alone, and domain trust is a precondition. That adjudication is written in the cold-start essay, not rhetoric.

Until then, H5’s accurate form: links are a ranking amplifier, not an indexing knocker. Reversing the order is the most expensive mistake of the cold-start stage.

The reproducible chain vs the unreproducible result

Combining the five verdicts reveals a clean boundary:

The reproducible part (all process):

• The method for validating a keyword gap (H1)
• The experiment design of one-shot full launch (H2)
• The data-page to content-page structural ratio (H3)
• The per-page screening question “can it be derived elsewhere?” (H4)
• Clearing the indexing gate before the authority gate (H5)

The unreproducible part (all judgement):

• Why multimodal rather than another category
• The specific 8-dimension scores for the 32 models
• Where the five procurement-gate thresholds sit
• Which 18 of 30-plus candidate topics got cut
• Whether listing a grade-C BCI-company model is brave or a gimmick
• When to admit a judgement was wrong

The boundary matters because it decides whether this methodology is useful to others. The process part can be copied; the judgement part cannot. Any methodology claiming the latter is copyable is not selling method, it is selling certainty hallucinations.

Our way of externalising the judgement part is to turn it into checkable data: the rubric lives on the methodology page, every model’s source links live on its dossier, every adjustment lives in the changelog, and evidence grades A/B/C are published independently of the composite. You cannot copy our judgements, but you can check them, overturn them, and recompute them with your own — the compare lab exists for exactly that.

This is also our final answer to “what earns AI-written content an index entry”: not whether AI wrote it, but whether it turned judgement into checkable structure. An AI-generated ranking with 256 traceable score points and 12 recomputable artefacts is more worth indexing than a human-written “in-depth review” made of adjectives. The reverse holds too — a human-written filler stream and an AI-written filler stream are thrown into “Crawled – currently not indexed” in exactly the same way.

Authorship is not the variable. Replaceability is.

One line to summarise the five

Validate the gap first, ship all at once, scale data not articles, every page must answer “can this be derived elsewhere”, and — do not buy links before you are indexed.

Four of the five are process and can be copied directly. The rest is judgement: you can only make it yourself, you must publish it when done, and only then can it be overturned.

All raw data is on the model dossiers, all computation on the tools pages, all rubrics on the methodology. If you build another site on this process, we want to know the result — especially the falsifying parts.