Cold-start log: testing Google’s “unique value” verdict with one long-tail term and zero backlinks

This is a pre-registered experiment, not a post-mortem

Most “how I did SEO” essays share a fatal flaw: they are written after the result is known. The author picks the actions that worked, drops the ones that did not, and narrates luck as methodology. Such essays cannot be falsified, so they cannot be learned from.

We write it differently. This essay publishes on the launch day of the 2026 China native multimodal LLM tier list and states expectations and falsification conditions, not results. All observable outcomes are appended weekly to the changelog, including the parts that prove us wrong.

The subject of the experiment is this site: 160 static pages, zero link-building budget, zero community promotion, zero paid distribution.

The judgement being tested

Google decides something is “worth indexing” by whether it has unique value, not by whether a human wrote it.

Both halves matter, but the second is more often ignored:

• AI-written content can be indexed, provided it offers something nobody else has.
• Human-written filler is thrown into “Crawled – currently not indexed” all the same, because it offers nothing nobody else has.

What the search engine faces is not a “human vs AI” problem, it is a replaceability problem. If an essay about Chinese multimodal models, with the byline and the site name blacked out, is indistinguishable from fifty others, then to the indexing system it is duplicate content — no matter who wrote it, how many words it has, or how many images it carries.

The real meaning of the Search Console status “Crawled – currently not indexed” is not “not good enough”; it is “we already have something much like this.”

So the core hypothesis of this experiment:

Under zero backlinks and zero paid distribution, page replaceability is the only variable determining indexation rate and ranking.

If that holds, all effort belongs on “making pages irreplaceable”, not on links, publishing frequency or social distribution.

Site structure: how the 160 pages are distributed

80 pages per language, 160 across Chinese and English:

Category	Pages per language	Notes
Home	1	Head-term carrier
Tier list	1	Full ranking of 32 models, sortable by 8 dimensions
Model index	1	Entry to all dossiers
Model dossiers	32	One page each with scores, benchmarks, scenarios, sources
Comparison index	1	Entry to all pairs
Comparison pages	10	Pairwise deep comparisons
Article index	1	—
Essays	23	This is one of them
Tool index	1	—
Interactive tools	4	Picker, cost calculator, context converter, compare lab
Info pages	5	Methodology, experiment, FAQ, changelog, about
Total	80	× 2 languages = 160

The key ratio in this structure is 42 : 28 — 42 data pages (dossiers plus comparisons) against 28 content pages (essays plus info pages). Data-heavy is deliberate: a data page is irreplaceable by construction, because it is computed from raw scores; a content page has to earn the same property through extra work.

The anti-replaceability method: one data artefact per page

Each of the 23 essays is bound to a data artefact that can only be recomputed here. The test is strict: if the artefact can be derived directly from public data available elsewhere, it does not count.

#	Essay	Bound data artefact	Why it cannot be derived elsewhere
1	What native multimodality means	N0–N4 native grades + shadow tier list	The grading standard is ours; the counterfactual re-ranking depends on our 26 scores
2	Tier list methodology	Rank variance over 2,000 weight perturbations	Needs full 8-dimension raw scores plus a fixed seed; others publish only composites
3	Benchmark traps	Rank-divergence index (internal gap + external cross-cycle + absence)	Internal gaps need all 8 dimensions; absence analysis needs both community heat and comparison shortlists
4	Chinese OCR field guide	Mini tier list over 8 untested document classes + the neural class	The 8 classes are our own; public benchmarks do not cover them
5	Long video understanding	Temporal tolerance curves / TTA	TTA is a constructed metric with no public counterpart
6	GUI agent benchmark	Failure recovery rate + coordinate error P50/P95	Public benchmarks report success rate only, never recovery
7	Cost per 1,000 images	2D and 9-dimension strict Pareto frontiers	9D dominance needs all 8 dimensions plus prices; no second copy exists
8	Enterprise selection playbook	Five veto gates + sensitivity analysis	The gate definitions and order are our construction
9	The frontier wave	Same-window alignment table + lineage delta table	The delta table needs per-dimension historical scores of same-family predecessors
10	Why cold models survive	The cold-model irreplaceability matrix	The “exclusive scenario” column comes from per-model human judgement on this site
11	This essay	A replaceability audit of all 160 pages + pre-registered falsification conditions	Self-referential, and updated over time
12	The security-video field guide	An 8-class security mini board + surveillance decay curves	The security classes are ours; surveillance data cannot enter public benchmarks
13	The industry exclusivity map	A mapping of 15 industry scenarios to exclusive models	The scenario list and mapping are our construction
14	The cross-review experiment	A judge-bias matrix (self-preference + family bias)	Cross-review data can only be reproduced here
15	Auditing the “Ge Fei” playbook	A reproducibility split of five hypotheses	The reproduction chain is ours

This table is also this essay’s answer: “unique content value” is not a slogan; it is an engineering constraint you can check page by page. Before writing any page, ask once — does this page carry something that cannot be derived elsewhere? If not, the page should not be written.

Three recomputation entry points keep the artefacts alive:

• Compare lab: change weights and dimensions, watch the ranking move in real time
• Cost calculator: change resolution and token counts, watch the bill move
• Model picker: tick hard constraints, watch the candidate set shrink

Keyword strategy: one head term, a field of long tails

There is exactly one head keyword and the whole site is organised around it: 2026 中国国产原生多模态大模型对比排名天梯榜.

It appears in the home page title, the tier list H1, and the body of all 23 essays — twenty-three out of twenty-three. That is not keyword stuffing — it is a genuine query intent that nobody currently occupies. A Chinese-speaking buyer looking for a ranking of domestic multimodal models finds either single-model reviews, translations of English leaderboards, or rankings that predate the October 2026 releases.

Long tails are laid out by “question type” rather than by word stem:

Long-tail intent	Carrier
X versus Y	10 comparison pages
How much do Chinese multimodal models cost	Cost calculator + cost per 1,000 images
Chinese multimodal models with private deployment	Enterprise selection playbook
Which model is accurate at Chinese OCR	DeepSeek-OCR 2 dossier + Chinese OCR field guide
Long-video understanding comparison	Long video understanding + context converter
GUI agent phone automation	GUI agent benchmark
Ascend domestic chips LLM	GLM-5.3 dossier + playbook gate G4
How is a ranking computed	Methodology + tier list methodology
Frontier release comparison	The frontier wave
Cold / open-source multimodal models	Why cold models survive + InternVL 4.0 dossier
Surveillance-video multimodal models	The security-video field guide + Guanlan dossier
Industry-scenario LLMs (finance / e-commerce / games)	The industry exclusivity map

Of these, “private deployment support”, “Ascend LLM” and “cost per 1,000 images” have no Chinese content with a checkable answer in our survey. Those are the three likeliest traffic entries for this site and the three least likely to be covered by big sites — the data is too fragmented and maintenance cost exceeds return; only a site dedicated to it would do this.

The two new long tails this cycle (“frontier release comparison”, “cold/open-source multimodal models”) are even closer to zero competition: the first is timely but unbanked, the second banked but low-volume. We are betting that the first’s indexation speed covers the second’s long-tail seepage.

Why two expansions: from 24 to 32 models

Expanding the sample is itself an anti-replaceability move and belongs in the experiment log.

The first expansion’s two additions — InternVL 4.0 and NeuroXess-N1 — improve nothing about the ranking; they improve the evidence spectrum: the site-wide evidence grades move from “13 A / 11 B / 0 C” to “14 A / 16 B / 2 C”. On an all-A/B board the evidence-grade column is decorative; with the first C, the column gains information for the first time.

At the same time, InternVL 4.0 introduces a new analysis object — absence divergence (the inversion of open-source community heat and comparison presence) — and NeuroXess-N1 brings one of the site’s two grade-C entries and its only non-standard modality. Expanding the sample turned three concepts that could only be talked about (evidence discount, leaderboard absence, modality coverage) into computable columns. Twenty-four models could not do that.

The cost is recorded honestly too: 32 models dropped the perturbation experiment’s never-moving models from 3 to 1, and widened the rank ranges of Keye-VL-2.0 and BAGEL-2. The board’s error bars grew — the real price of expanding the sample, and we accept it, because bigger error bars are more honest than a biased sample.

The second expansion (26 → 32) has a different rationale: the first eight cold models proved that coverage becomes a traffic entry, but all eight sat on the technology-form axis (open source, specialist, embodied, edge, academic). What was missing is the other half — the industry-scenario axis. Finance, e-commerce, games and surveillance each contribute one model, and with the two hardware-vendor entries the site gains a new analysis object for the first time: scenario exclusivity — 15 industry scenarios where only one or two models can compete, a mapping that exists nowhere else. The cost is recorded too: the second expansion kept the never-moving count at one (Qwen3.5-397B has not moved across all three expansions) but merged the perturbed ranges of ranks 27 through 32 — the order among the last six is essentially weight noise.

Two falsification conditions

A pre-registered experiment needs explicit failure criteria, or it is self-talk. We set two:

Falsification condition one: replaceability does not work. Ninety days after launch (by 2027-01-16), if fewer than 40% of the 160 pages are indexed by Google and the unindexed ones cluster in “Crawled – currently not indexed”, the core hypothesis is falsified: under zero backlinks, unique content value is not sufficient to drive indexation, and domain authority plus links remain necessary conditions. → The conclusion should then change to: cold start must solve links and domain trust first, and content differentiation second.

Falsification condition two: artefact density does not work. If indexation exceeds 40% but all 23 essays draw less organic search traffic than the average of the 42 data pages over the same period, then “one data artefact per page” is falsified — artefacts help ranking but not clicks, and the correct form for long-form is something else. → The conclusion should then change to: long-form should serve the internal links and semantic coverage of data pages, not chase rankings itself.

Neither is a “if we fail, the environment was bad” escape hatch. Each points to a specific, opposite conclusion.

Three weaknesses we volunteer

A pre-registered experiment without honesty about where we might be wrong has no sincerity.

One, sample size n=1. One site, one launch, one point in time. Even if indexation hits 85% after 90 days, that cannot separate “unique value works” from “the October 2026 algorithm happened to be friendly”. All we can do is publish the process data — page inventory, per-page artefacts, timeline — so others can reproduce it on their own sites.

Two, “unique value” is self-judged. The 12-row artefact table above is filled by us, and we believe it cannot be derived elsewhere. That judgement may be wrong — some site we have never seen may already have done the N0–N4 grading. The mitigation is a hard criterion (can it be derived directly from public data?) and logging any counterexample found in the changelog.

Three, AI-assisted writing may itself be misread. Content on this site is AI-assisted, which is public and stated on the about page. By this essay’s opening judgement, that is not an indexation barrier — replaceability is the barrier, not authorship. But if indexation is low after 90 days, critics will say “see, AI content doesn’t work”, and we will be unable to separate the AI question from the artefact-uniqueness question. That confound we cannot eliminate, only admit.

Timeline

Date	Event
2026-10-18	All 160 pages live, this essay published, pre-registration takes effect
2026-10-25	First Search Console coverage scrape, logged in the changelog
2026-11-18	Day 30: indexation rate, clicks and impressions published for the first time
2026-12-18	Day 60: click comparison, 42 data pages versus 28 content pages
2027-01-16	Day 90: both falsification conditions adjudicated

Node results — including results that falsify this essay — will be written into the changelog as they are.

If you want to reproduce this experiment

You need three things, none optional:

1. A set of recomputable raw data. Ours is 32 models × 8 dimensions = 256 score points, all public on the model dossiers. Without raw data there are no artefacts, and pages become replaceable by default.
2. A discipline of asking “can this be derived elsewhere?” Once per page, before writing. If the answer is yes, do not write it. This cuts page count hard — our 23 essays were selected from more than forty candidates.
3. A failure criterion written in advance. Written before launch, not after.

None of the three has anything to do with “writing skill”, “depth” or “expertise”. They are structural, therefore copyable. That is why this experiment is worth running: if it succeeds, the reason for success can be taken away by others.

Scoring criteria and evidence grades are on the methodology page; the audit of “can the methodology itself be reproduced” is in the “Ge Fei” playbook audit.