Fix Product Master Data Before You Do GEO

Tuesday morning, a Berlin support inbox: a customer wants to return an outdoor power station because an AI assistant told her it would run a kettle and an induction plate at once. Plug both in and the unit trips its protection. Operations edits the product page; the spec table was already right. The EU marketplace listing quotes surge watts as continuous. The distributor spreadsheet says “high output”. Fixing any one of them fixes nothing.

When product master data disagrees with itself, GEO does not optimize anything — it puts better copy on top of a wrong number, and AI quotes that number with your brand attached.

AI compares fields, not sentences

Ask “will this run a kettle” and the model has one job: line up wattage, capacity, outlets, and compatibility across candidate SKUs. Copy can differ by market. Fields cannot.

Breakage happens in three places. Definition collision: rated and surge watts share one field called “power”, so every downstream comparison is false. Unit collision: capacity is Wh on the US site, mAh on the EU site, and the conversion needs a voltage nobody published. Change propagation: a spec is corrected on the site while the feed, listing, and distributor sheet sit three months behind.

The third is expensive because it is invisible. Retrieval reads every version it finds and takes the self-consistent one, not the one you meant — and across three channels, often none is. In the product graph, a product is an entity assembled from attributes; contradict them and the graph is broken however good the page looks.

Draw the boundary before you clean anything

Most teams treat narrative as master data and attributes as copy. Reverse it: narrative decides whether AI cites you, attributes decide whether the citation is right.

Data domain	What it holds	Which GEO step depends on it	Owner
Identity	Brand, model, GTIN, variant relations	Whether AI recognizes five pages as one product	Merchandising
Attributes	Dimensions, materials, wattage, capacity, compatibility	“Will it work for me” questions	Category and product
Compliance and fulfillment	Certifications, warranty, returns, ship-from country	Whether AI recommends and agents order	Legal and operations
Commercial	Price, stock, promotions	Whether the quoted price is credible	Pricing and operations
Narrative	Claims, use cases, images, video	Whether you get extracted and believed	Content and brand

The table allocates authority, not taxonomy. Nothing in the first four rows should be editable by one market’s operations team alone.

Three decision rights you cannot merge

Master-data projects rarely fail on software. They fail because nobody owns what a field means.

Split three ways. Market and category own semantics — which attributes buyers ask about, a demand trade-off rather than a technical question. Operations owns values and units and answers for accuracy. IT owns the pipeline: one authoritative system, sync, and validation, so changes are entered once.

Both extremes fail. IT hands back fields like erp_pwr_max_kw that nobody fills and no buyer would phrase. Content rewrites the copy while the feed stays put, so AI keeps reading the old value.

The shared artifact is one field card: definition, unit, allowed values, missing-value rule, source system, owner, and who to notify on change. Not documentation engineering: the contract that makes the field citable.

A minimum viable governance: 12 fields, five steps

Skip the full attribute library. Pick twelve fields buyers ask about and AI compares, then run one end-to-end publication; that argues for budget better than a PIM rollout does.

Step	Action	Owner	Passes when
1 Select	Pull the twelve most-asked attributes from tickets, site search, and AI question logs	Operations and category	Each field maps to three real questions
2 Define	Write the field card: definition, unit, allowed values, missing rule	Category and IT	One SKU reads the same value in all three channels
3 Source	Name one authoritative system; everything else syncs read-only	IT	A change is entered in one place
4 Publish	Generate PDP, feed, and structured data from that source	IT and operations	Spot-check 30 SKUs, three channels agree
5 Audit	Route changes through the same workflow; monthly sample	Operations	A change log and a sample result exist

Step 2 is the watershed. Skip it and the rest become a faster way to copy wrong data into more places. Turning attributes into buyer language is separate work — an AI shopping answer covers it; this piece only guarantees the raw material.

Example: one capacity, three channels

Example — a hypothetical outdoor power brand; no real company, no figures. The US page lists capacity in Wh, the EU listing in mAh, the distributor sheet “about one kilowatt-hour”. Each may be accurate. None is comparable.

Asked whether the unit can fly, the model cites capacity and still cannot answer: air rules read Wh, and mAh needs voltage. The buyer gets a hedge, the brand a lost answer slot.

The fix touches no copy. Capacity moves to Wh, with voltage as a required companion field. Rated and surge wattage split into two fields. Compatibility becomes a closed list of allowed values, not free text per site.

Acceptance has two numbers: 30 sampled SKUs agree across three channels, then the same question set returns your published values in AI answers — the AI visibility baseline, not a content review.

Where this breaks

A few hundred SKUs. One field card and a shared sheet are enough. A PIM is the waste here, not the discipline.

Sales concentrated inside a marketplace. The platform’s listing fields are your master-data interface. Complete the required attributes before building your own.

A product line changing within the year. Freeze the field cards, delay the library. Migrating a premature structure costs more than waiting.

No field owner exists. Run a read-only inventory first and list every conflict across the three channels; the conflicts argue for the project better than a business case does.

Expecting it to fix “not being recommended”. Correct fields stop AI from reading you wrong, not from skipping you — recommendation is off-site authority work, and this lane sits inside the strategy and market hub.

This month, export 90 days of support tickets containing “will it”, “how big”, or “how long”, then pick the twelve attributes buyers asked about most. Put the values for your 30 best-selling SKUs from the product page, the feed, and one marketplace listing into one sheet; mark every disagreeing cell. Fix the three fields with the most conflicts, a card each: definition, unit, allowed values, missing rule, source system, owner. Rerun the same questions next month and compare the numbers in the answers.