One Product, Five Names: Aliases, Model Numbers, and Local Terms

A US DTC brand sells portable power stations. One week’s support inbox holds four names for the same unit: “solar generator,” “camping battery,” “backup power for my CPAP,” “the ES-1000.” The page says “ES-1000 Pro Portable Power Station.” Nothing on it is wrong. Ask an assistant what runs a CPAP overnight and four competitors come back; this brand does not. A rival that folds three wattage variants into one page, meanwhile, gets the 1,500 W runtime quoted for a 1,000 W question. One brand is invisible; the other is confidently wrong.

Buyers call one product five names, and a model resolves the name before reasoning about the product: alias coverage decides retrieval, alias boundaries decide confusion.

Recall happens before understanding

An AI answer runs in three moves: retrieve candidates, disambiguate the entities, compare specs. Retrieval is plainer than teams expect: string overlap, embedding similarity, how third parties phrase you. “Solar generator” against a page saying only “portable power station” is a jump some models miss.

Exact matching is why the model number survives translation and why it breaks. ES-1000 Pro, ES1000 Pro, and ES 1000 Pro are three strings; page, feed, and manual each pick one, and the model splits a SKU in three or drops it. An alias belongs on the product entity, not in a paragraph — see the Product Graph.

The failures cost differently. A recall miss is silence; you never learn you were skipped. Confusion arrives after the recommendation, and the return plus the review pull it back. So the first rule is one row per entity.

Five names, four to keep, one to wall off

Name type	Example (fictional)	Same entity?	Where it goes	Cost of misplacing it
Canonical name	Portable Power Station 1000W	Yes, one primary	Title, H1, feed title	Primary name drifts
Model number	ES-1000 Pro	Yes, one per SKU	Title suffix, spec table, beside GTIN	Variants read as separate SKUs
Marketplace category	Outdoor power station	Yes	Listing, feed, category page	Found on the platform only
Local vernacular	Solar generator, camping battery	Yes	One “also called” sentence, FAQ	Invisible to how buyers ask
Misspellings, legacy names	ES1000Pro, ES-900	Misspelling yes; legacy no	Misspellings out; legacy gets a page	Legacy specs land on the current model

Do not chase misspellings: ten spellings make a page a worksheet and retrieval barely moves. Maintain three: legal model variants, the vernacular buyers type, the marketplace category name.

A legacy model is a different entity with its own page and succession up front: “The ES-900 is discontinued; the ES-1000 Pro replaces it.” “Also known as ES-900” on the current page invites exactly that confusion.

Build the table: one row per entity, one per market

Step	Action	Owner	Passes when
1 Collect	Copy buyer wording verbatim from tickets, zero-result search, marketplace and ad queries	Content, support	25 of 30 questions hit the table
2 Define	One row per model: name, model string, GTIN, market; legacy separate	Product ops	No GTIN twice, no model in two rows
3 Classify	Mark each alias collect, search-term-only, or reject; each keeper gets a spot	Content	Nothing parked “for later”
4 Land	PIM row with owner and review date; page and feed read one source	Engineering	Ten SKUs match across page, feed, listing
5 Re-test	Ask in the buyer’s words; record cited page and model	Growth	Confusion measurably lower

Step 3 is where teams go soft: allow “park it for later” and the table becomes a list nobody dares delete while pages fill with synonyms. Step 4 decides asset or document — separate sources diverge within a quarter.

Put it on the page: one name per position

Aliases must appear; they must not pile up. Five synonyms in a title and a first paragraph in three languages read as several attributes. Four placements:

• Title: canonical name plus model string. Vernacular never enters the H1, or the primary name drifts.
• Body: one sentence after the canonical name — “also sold as a solar generator” — in the buyer’s words.
• FAQ: the buyer’s question as the question, with name, model, and boundary in the answer. Easiest to extract.
• Structured data: alternateName, sku, gtin, matching the visible text; see Product schema in 2026.

The marketplace category name stays in the listing and feed; your site need not adopt it. One term, one position, one job.

Example: an alias table for one power station

Example — fictional brand, model naming, and scenario; structure only.

Two rows. Row one: Portable Power Station 1000W, model ES-1000 Pro, aliases “solar generator,” “camping battery,” “backup battery for CPAP,” category “outdoor power station.” Row two: the ES-900, own page, discontinued, successor named. Rejected: “ES1000Pro,” and “mobile power station,” which reads as industrial equipment.

Three edits. Title: “Portable Power Station 1000W | ES-1000 Pro.” Second paragraph, one sentence: “Also sold as a solar generator or camping battery; model ES-1000 Pro.” One spec bullet becomes an FAQ — “Can it run a CPAP machine overnight?” — answered with model, watt-hours, and the low-temperature runtime limit.

Acceptance is not a ranking: when asked “portable power station for CPAP,” does an assistant name the ES-1000 Pro and leave the 1,500 W runtime out? If not, check whether the buyer’s phrase appears on the page before rewriting copy.

Where this breaks

First, commodity products with one settled name. Cables and consumables are searched by spec and compatibility; listing compatible models beats aliases.

Second, regulated categories. Where the model string is part of a certification, it cannot bend for vernacular; aliases live in explanatory text cleared by compliance.

Third, marketplace-only sales with locked fields. You can edit the listing title and A+ content; a site table the platform cannot read is unpaid work.

Fourth, a brand no third party has mentioned. Aliases fix retrieval, not credibility; check whether the brand entity exists first.

Fifth, a synonym dump. Ten names teach a model that synonyms are attributes, and variants contaminate each other — worse than silence.

Recall fails differently by entry point — search query, chat question, price-comparing agent — as mapped in the four AI search entry points; the rest sits under content and semantics.

This month, copy 30 days of support messages about one hero SKU into a sheet, verbatim, and mark which phrases are the same entity, which another model. Then change three places: the title gets the standard model string, the body one “also called” sentence in a buyer’s words, one FAQ a real question. Ask an assistant and see which model comes back.