A GEO QA Checklist for Content Factories

A composite example: a charging-accessory brand runs five storefronts with six writers and two translators, publishing 30 pages a week. On Wednesday the German store ships a wall-charger page that still reads 120 V and carries a US certification badge. By Friday support answers warranty questions. Final review: 90 seconds, headline, typos, hero image.

At content-factory scale, quality is not a writing problem: it is a checklist of yes/no items, a few gates that genuinely stop publication, and a sampling loop that repairs the checklist.

Final review checks fluency; the misses are facts

Rework comes from three sources, none prose.

Numbers. Model, capacity, wattage, price, warranty — one value off the master data or the policy page.

Structure. No sentence can be lifted alone, or one paragraph argues three things, so extraction drops the block whole.

Market. Units, voltage, plug, return window, local certification. Readable after translation, unusable in the market.

Self-review fails because the writer checks clarity, not whether the number matches master data. Quality surfaces after publication — support tickets, returns, AI answers quoting last quarter’s price — the costliest detection.

The counterexample is the scoring rubric: accuracy, readability, voice, 1 to 5. Three months in, every draft scores a 4. A rubric measures the reviewer’s mood; a checklist measures facts.

Twenty items, four groups, ten hard stops

Longer is not safer. Four groups of five cover nearly every rework ticket.

Group	Checks	Who	Hard stop
A Facts	Specs match master data; price, stock, warranty match the system of record; shipping terms match the policy page; certification sourced; comparison names a model on sale; no unsourced “first”	Category or support	Yes, scriptable
B Structure	An 80–120 word answer unit near the top; one claim per paragraph; tables carry units; no paragraph opens with “It”; the claim survives a covered headline	Editor	Yes, human but samplable
C Market	Units, voltage, plug; warranty and return window per market; local scene words; names match the entity sheet; local regulatory basis	Local ops	Yes, with a local signer
D Assets	Internal links live and on-topic; alt text; canonical and hreflang; schema matches the body; no two pages chase one question	Engineering	Yes, in CI

A and D belong in the build, where failure fails the pipeline, not a reviewer’s memory — see GEO checks in CI. No validator knows whether a warranty window clears in Germany, so C needs a local signer; B stays human: an editor covers the headline and reads the paragraph cold.

Only ten become hard stops: all five of A, B’s answer unit, C’s warranty and plug, D’s links and structured data. Every extra gate adds a reason to route around it — block all twenty and within a week everyone files exceptions. “Price is accurate” cannot be decided; “price equals the policy page” can.

Three gates: no signature, no publish

Gate strength lives in the stop, not the signature: a signature assigns blame, not quality. Run them in order.

Gate	Timing	Signer	Blocks	If late
1 Self-check	Handoff	Writer, script-assisted	What A and D decide	No entry to gate 2
2 Facts	Within 48 hours	One person: category, support or engineering	All of A, plus B’s answer unit	Returned; no silent default
3 Local	Before publish	Local market ops	All of C	Market waits; others ship

“Auto-approve” is how gates die: nobody has time, the system releases after three days, and within a week every page takes that path. Ship two days late instead.

Add a withdrawal right: anyone can mark a published page under review or roll it back without justification. No blame, so people use the gate.

Example: 24 pages, six stopped

Example — page counts, stop reasons, and rework rounds are invented; no real brand or batch is described.

Stop	Pages	Cause	Handling
Gate 1, script	1	Link to a merged URL	Relinked, released same day
Gate 2, facts	3	Two prices off the policy page; one unsourced “military grade”	Returned for rewording
Gate 2, structure	1	Opening was brand vision, no answer unit	80 words added, second review
Gate 3, local	1	US plug in the German page; US return window copied	Market held, republished

Eighteen of 24 released, six stopped, five fixed the same day (Example). Clustering is the point, not counts: facts stops mean master data is wrong, and another item fixes nothing until the spec table is unified — see whether your brand entity exists. Structure stops mean the template needs work.

Sampling audits the checklist, not the writer

Each week, pull 10% of published pages, at least five, stratified by page type — product, category, comparison, policy — not by writer. One question: run it through the gates again — would it stop?

Bucket every miss: the item is missing, nobody ran it, or the master data was wrong. The fixes differ — add an item, change the process, repair the data — and most teams pick the first. Three rounds later, the miss rate has not moved.

Cap rework: a page stopped twice goes back to the template, not to the writer. After three rewrites a writer deletes sentences until nothing can be faulted — and nothing can be quoted.

Tie rework rate to writer performance and writers learn to write safe sentences: the adjectives go, and the quotable sentences with them. Sample only cited pages and the uncited ones, carrying last quarter’s price, go unchecked.

Keep the sample aligned with the AI visibility baseline: same language, region, and prompt set, or you cannot tell better content from a volatile answer.

Where this breaks

First, a team of two or three. Twenty items turn into homework nobody files; start with eight: price, warranty, answer unit, units, links, schema, names, rollback.

Second, untrusted master data: a checklist only distributes errors more neatly.

Third, marketplace listings and feeds, which run the platform’s own field validation. This list governs the owned store.

Fourth, regulated categories: medical devices, toys, air-shipped batteries. A checklist does not replace legal review.

Fifth, loading it all at once. Add one item a quarter. Ten at once is the same as none.

This month, cut review requirements to twelve decidable items, run them on the next ten pages, and log how many stopped, who signed, and how many rework rounds. A month later, turn the error class you missed into item thirteen. The standard sits on the content and semantics hub; the checklist is what phase three of the 90-day GEO roadmap hands over.