Menu descriptions are now the discovery layer.

An observation crossed my desk this week that is worth naming plainly. The systems that recommend restaurants — ChatGPT, Perplexity, Google AI Overviews, the delivery-app agents — have quietly reclassified the menu. It is no longer a list of things you sell. It is the primary surface those systems read to decide what kind of restaurant you are.

\u201cChicken sandwich\u201d tells a machine nothing. It resolves to the median of every chicken sandwich in its training data — which is to say, a fast-food default.

"Buttermilk-brined thigh, pickled fresno, soft brioche, served on butcher paper" tells a machine everything. Preparation. Texture. Heat. Service register. It cannot be confused with a drive-through item. It maps cleanly onto occasion queries — "casual dinner with friends," "elevated comfort food," "somewhere I can bring my in-laws."

This is not a new argument.

It is the thesis behind the Vocabulary Translator, the 22 Swaps, and every Guest Read run through this system. Sensory specificity is the unit AI uses to decide whether to name you. Generic category words collapse you into the middle of the category. Specific language pulls you out of it.

What is new is the external validation. The platforms that once rewarded keyword density are now rewarding descriptive precision. The direction of travel is settled.

A note on schema markup.

Alongside menu language, structured data — Restaurant, Menu, and MenuItem schema — has become a meaningful lever for how AI parses a site. It is the structural clarity that lets a machine read the brand without ambiguity: a machine-readability layer that works with menu language, not a P.E.D.A.L. pillar.

Every Signal Integrity Audit™ addresses both surfaces — the language a guest reads and the structure a machine parses — because they are the same problem stated twice.

What to do with this.

Open your menu. Read the first five items out loud. If a machine could not tell your chicken sandwich apart from any other chicken sandwich in its training data, the machine is not going to recommend yours.

That is a fixable problem. It is also the fastest correction available to most operators — and the one that tends to move an interpretation more than any other single change.