Templates and Personalization: Avoiding Repetition After 132 Articles

A November review at a composite cycling-accessories brand in Munich: 60 drafts built from one template — scene, four proofs, comparison table, limits paragraph. The content lead opens three winter-glove drafts: the openings differ by four words. In a sampled assistant answer, all three collapse into one sentence, cited to the earliest URL; the other two earn no mention.

A template guarantees the floor, but avoiding repetition is a separate job: keep the modules, change the coordinates of who, under what constraint, and giving up what, then gate synonym rewrites before publish.

One template, one sentence

A template fixes structure; writers, left alone, fix the content too. What decides whether two drafts are the same piece is proposition and evidence, not syntax.

Extraction works on sentences and paragraphs. Two pages with the same proposition and parameters read as two copies of one fact; retrieval keeps one source, and the copy is never compared, only dropped. Vector search is harsher: a synonym rewrite lands near the original and does not read as new evidence.

“German winters are cold, so gloves need insulation” becomes “German winter temperatures are low; gloves require thermal performance.” The words change, the proposition does not, and the evidence is still zero. The rewrite clears every literal plagiarism check and fails the identity judgment. The template also removes the first question: why does this page exist? Ten drafts answer “what to do when your hands are cold” with ten new openings.

A module library fixes order, not content

Fix two things: field order and paragraph boundaries. Everything else is a slot, holding evidence a third party can check, never an adjective.

Module	Fixed	Changed every time	Passes when
Conclusion	position, proof, contrast, next step	scene, unit basis, exclusion	title hidden, four sentences stand
Evidence	spec, test condition, source	values and source links	a stranger can recheck each one
Comparison	dimensions and units	rival, price band, conditions	answers “when should I not buy”
Limits	position and length	the specific condition that fails	at least one verifiable limit
Action	exactly one move	the move and the retest	checkable within a week

Granularity follows liftability, not layout: a paragraph arguing two topics is cut in half or dropped whole. Hence the boundary with the PDP component library: components decide how a page looks, modules how it talks. Same field order, different acceptance tests.

The angle matrix: coordinates, not adjectives

How many distinct pages a product supports depends on how many coordinate sets you can carve, not synonyms.

Axis	Values	Decides
Buyer stage	problem-aware, comparing, pre-purchase check	how the conclusion advances
Objection	price, weight, ease of use, after-sales	which proof leads
Scenario constraint	temperature, duration, terrain, company	test conditions and unit basis
Market constraint	units, certification, return window, payment	facts localization must replace

One value per axis is a coordinate. Two topics sharing a coordinate should merge, not publish side by side. A topic that moves on one axis while evidence stays identical is a pseudo-angle, the same piece again. Buyers shift axes mid-session: “which brand” becomes “does it hold at 28°F for an hour,” then “what does a return cost in Germany.” Sorting topics by axis pre-builds the question chain as pages.

Five steps and a gate

Order matters: extract modules and set coordinates first, or writers draft what feels good and then hunt for a coordinate.

Step	Action	Owner	Passes when
1 Extract	Pull five modules from the ten most-cited pages of the last six months	Content lead	modules cover most live pages
2 Set coordinates	Fill four axes per topic; merge duplicates	Content, category	no two topics share a coordinate
3 Swap evidence	At least one verifiable proof per coordinate	Category, supply chain	it points to a spec, test, or third party
4 Gate	Literal overlap plus side-by-side answer units	Editor	5-gram overlap under 0.25 and one sentence cannot summarize both
5 Review	Judge angle and evidence only	Editor	grammar is the template’s job

Step 4 has two layers. The literal layer is a 5-gram Jaccard score, a hard limit; under 0.25 only rules out copying. The semantic layer is human: read both conclusion sentences; if one covers both pages, they duplicate whatever the vocabulary. The gate blocks publishing, not drafting. It fits phase two of the 90-day roadmap: once modules and coordinates exist, capacity repeats.

Example: one glove, two coordinates

Example — brand, product, and figures are invented.

	Coordinate A	Coordinate B
Buyer	commutes 20 minutes each way, always cold	rides 50 miles or more, palms go numb
Season and market	winter rain, Germany	summer, United States
Conclusion	for the rider who commutes through a wet German winter	for the rider whose hands fail after hour three
Proof	back-of-hand temperature over a 30-minute ride at 23°F, touchscreen response, waterproof rating against rain intensity	palm pad thickness, breathability, pressure distribution after four hours
Limit	does not seal in sustained downpour	thick padding dulls bar feel; not for crit sprints

Same skeleton, no overlap in proposition, scene, or proof. Retrieval will not merge them. Each page needs a next step: “see the other piece” fails the moment the paragraph is extracted alone.

Where this breaks

First, facts stay identical sitewide. Specs, warranty, price, and return window cannot fork for variety; a model that retrieves two conflicting versions discounts both. Personalization covers interpretation, scene, and exclusions.

Second, a narrow category cannot yield enough coordinates. Pseudo-angles dilute topical focus; merge and deepen instead of adding articles.

Third, do not run a module library without an editor. It degrades into a fill-in form: every cell has words, none is citable — the commonest failure of content at scale.

Fourth, finer modules raise the cost of every sitewide change. Version them and roll major edits out in batches.

Fifth, the duplicate gate does not travel. A threshold calibrated on English proves nothing in German or Japanese; rerun the answer-unit comparison in the target language.

This month, pick a category with 20 or more published pages and copy every answer unit into a table. Read in pairs; where one sentence covers two pages, merge them, keep the primary URL, and spend the freed capacity on a new coordinate. Two weeks later, check whether citations concentrated on the survivor or vanished with the page — the four layers of AI visibility tell you which. The sequence sits on the content and semantics hub.