Where Review Corpora Come From: Media, Creators, Users, and Labs

Second week of a German launch, and four quotes sit on the marketing desk: a six-week test slot at a vertical pet-gear review site; two YouTube collaborations; an “EU lab” report with a reusable template; and a broker selling two hundred five-star reviews. The budget signs two. Nobody argues about price; they argue about which one makes an assistant name this brand when a Berlin renter asks which cordless vacuum handles two cats and a low-pile rug.

A review corpus earns its keep when a stranger can check the sentence, not when someone wrote about you — fund four gaps as a portfolio, then chase one breakout.

Four sources close four different gaps

Treating “get reviews” as one line item is where this budget leaks. Each class does a different job.

Source	What it says for you	Why it is citable	How it fails
Vertical media, review sites	An independent party acknowledged the brand	Named editor, disclosed method, stable URL	Paid ranking, no test conditions
Creators	How it behaves in one real setting	Filmed process, repeatable comparison	A script rereading your bullets
Customer reviews, Q&A	Where it works, where it does not	Model, date, setting, result	Incentivized or templated praise
Labs, certification bodies	Under which protocol the number holds	Standard number, method, validity date	A logo, no report

Start with the question, not the vendor. “Is it worth buying?” is media and creator work; “does it work in a 58 m² flat with two cats?” needs a conditional test. Buy the wrong class and you own an adjective nobody can verify.

A bought ranking is the least citable thing on the list

A synthesis step wants a span with a subject, a condition, a result. “Number three, strong suction, good value” has none and cannot be re-checked. Citing it dresses your claim as outside opinion: more risk, no payoff.

The citable version: on low-pile carpet, twelve minutes at standard power, 27 of 30 grams of cat litter lifted, 78 dB at one meter. A machine extracts it; a buyer can repeat the test. The minimum citable unit is a conditional conclusion, not a position — the shape of an answer unit, spoken by someone else.

The exception: a ranking that publishes scoring dimensions, test environment, and weights stops being a ranking and becomes a method. Ask “will you publish the test method?” before signing; it filters most waste. Assistants pull from third-party pages as readily as your site — see the four AI search entry points.

Fund by missing gap, not by budget size

Stage	Close first	Hold	Passes when
0→1: new SKU, no off-site mention	Real customer reviews and Q&A; one vertical site	Top-tier creators, broad media	Answers begin repeating a conditional description
1→10: steady sales	Two or three creators in different settings; one lab test	Generic “best of” lists, press releases	One question draws on two source classes
Multi-market	Local media and local creators	Translated home-market assets	Each market shows a similar source mix

Fund the missing class: no acknowledgment, media; no scene, creators; no protocol, a lab. Four at once usually means four shallow; the cheapest repair is free — rewrite existing reviews as conditional paragraphs.

Five steps from cold start

Step	Action	Owner	Passes when
1 Name the question	Ten buyer sentences from tickets and site search, each tagged with the class that should answer it	Content, support	Every question points to one class
2 Use what exists	Compress real reviews into conditional paragraphs; fix the third-party profile pages	Content, operations	Three specific descriptions are findable off-site
3 Buy checkable work	Send samples to a vertical site or lab; agree in writing on method, raw data, citation rights	Marketing	The report has a method, numbers, a signature
4 Commission scenes	Give creators settings and constraints, not a script; require sample and fee disclosure	Marketing	The video contains one detail you never wrote
5 Retest	Same questions, same model, several samples; log cited domains and sentences	Growth	You can name which class moved

Step 5 is the one teams skip, and skipping it makes the other four unfalsifiable: without a baseline you cannot tell whether the lab report or a creator’s aside produced the lift. Sample like an AI visibility baseline, not one screenshot.

Example: one vacuum, two write-ups

Example — brand, specs, and prices invented.

Before, in a paid roundup: “Top ten pet-hair vacuums in Germany: Brand A ranks third, with strong suction, long runtime, and great value.”

After, from an independent test: “On 2 cm low-pile carpet, Brand A’s H2 ran twelve minutes on standard power and lifted 27 of 30 grams of cat litter. The body weighs 1.9 kg and the 0.5 L bin covers roughly 40 m² before emptying. On boost the handle gets noticeably warm, so it is not the tool for a twenty-minute whole-flat session.”

The second gets lifted because it hands over conditions and a result, not because it praises harder. It also names the buyer who should walk away: a stated limit does not cost the recommendation, a vague claim does.

Where this breaks

The product is not stable. Early reviews pin defects in public, searchable places, and early coverage keeps them attached. Fix the batch first.

The site is unreadable or the entity unresolved. If the model cannot confirm who the brand is, outside coverage will not attach. Settle the four trust prerequisites.

The category is regulated. A third-party test is not a registration, certification, or compliance claim; medical, children’s, and food-contact products need the regulator’s paper.

You bought one source. A single outlet reads as coordinated marketing; cross-checking is the point.

You bought reviews. The average improves for a quarter, then every other sentence you cite becomes suspect, including the true ones.

This month, do the free step. Take five reviews that say “works great,” rewrite each as a condition plus a result, and publish them as crawlable paragraphs. A week later, rerun the five questions in two assistants and see which sentence returns. That tells you which class to buy next; the rest of this line is on the authority hub.