the numbers the product will not invent

Sample content has a way of surviving into production. A review section carried a demonstration fallback: a few plausible reviews and an average just under five stars, rendered when the real fetch failed. It existed to make review builds look complete. What it actually did was make a failed fetch look like a well-reviewed business.

The fix inverted the default. The fallback is off unless a business opts in through its own configuration, and even then the preview renders under a clearly labelled banner so it cannot be taken for customer content. The gate reads configuration reactively, so a late-arriving configuration is respected rather than cached over. The placeholder that replaces it is deliberately plainer: a header, honest copy, no stars, no numbers, and a configuration switch that can hide the section rather than dress it up.

The rule was then extended to machine-readable output. A structured rating block is emitted only when a real, valid aggregate exists: a score in range and a count above zero. Missing, malformed or out-of-range values omit the key entirely. Nothing is defaulted, because a defaulted rating in machine-readable markup is a claim made to a search engine on the business's behalf.

Cost: the failure state looks worse than the fabricated one, and somebody had to accept that. A business that wants a filled-in review section now has to earn one.

What it proves: the surface is the evidence, and a convincing placeholder is treated as a defect rather than a convenience.