A Directory of Other Businesses Made LLMs Credit a Spice Shop With Denim Repair

Non-core directories hosted on a primary domain muddy brand entity salience in LLM training data.

Crystal Carter describes working with a spice merchant that LLMs repeatedly claimed offered denim refurbishment, event planning, and other unrelated services.

Digging into the source of the hallucinations — which rarely come from thin air and normally stem from confused inputs — uncovered a COVID-era marketplace the small business had built on its website to support other companies.

The showcase hosted profiles for those outside businesses and earned press coverage, causing LLMs to ingest the entire directory as a single business rather than a third-party showcase.

To launch a brand pivot or keep a partner showcase live without polluting the core entity, spin the project off onto a dedicated domain to isolate its entity relationships.

Alternatively, apply a noindex directive so users can still access the directory while signaling to LLMs that the pages sit outside the core business, or block the showcase pages in robots.txt.

Separately, Gianluca Fiorelli reports testing a client's informational health microsite that remained locked out of LLM headache answers due to an expired domain's entity history.

When buying an expired domain, audit its historical entity associations alongside its backlink profile.

The client had built the microsite around headaches — including the term "aura headache" — on an acquired domain, but probably never checked that a new-age therapist had previously used the domain for personal "aura" color readings.

While live grounded data reflects the current medical content, LLM training data still classifies the domain as a new-age therapy website.

Because headache queries often return parametric answers rather than grounded results, the microsite never surfaces in those responses across repeated tests.