Copyright applies to the prose itself, though—the word choice and ordering.
A site like the one I'm describing doesn't have to use the same words the original site uses. In fact, I'd really hope it wouldn't, since these sites almost never use a standardized vocabulary, and in fact try as hard as they can to make their recipes into interesting prose (i.e. using a different word in reference to something each time they reference it) at the expense of clarity.
The "boiled-down form" of these recipes, would look less like prose text, and more like an RDF graph linking ingredient-noun vertices together using action-verb edges.
In fact, now that I think about it, it could probably be stored in a format machine-readable enough to function equally well as machinima choreography for a "cooking simulator", i.e. an HTML5 canvas app with an engine approximating a cooking game (e.g. Cooking Mama) but without the "game" part.
Such a "timeline view" of the recipe could actually keep track of the collection of pots/pans/bowls/etc. required in the recipe, and then visually display the current state of what's in each of them at each step, as you scroll through each step in the recipe's timeline. Like a visual debugger, for food.
The textual representation of the recipe would just be generated from that same non-textual data format, via a simple structured NLP rule-set. (Yes, you could instead throw the sort of algorithms they're using to generate baseball news coverage[1] at the problem, but we want standardized text, not stylized text.) Bonus: i18n for free. The normalized recipe data isn't "in" any language, so it's equally easy to write a ruleset for generating recipe-step prose that is in any language.
The resultant text of such a generative process shouldn't be copyright-infringing, even if it matched some other recipe word-for-word by coincidence. When you "extract" only the bare facts from a work and discard the rest, you're not making a derivative work; a work of yours that employs those "extracted" facts is not a derivative of the source of those facts. (If it were, then every summary or review of a work, would be in violation of the copyright of the work it's summarizing or reviewing, since the bare facts referenced were gleaned from the original work.)
[1] https://venturebeat.com/2016/07/01/associated-press-expands-...