^{(Disclaimer: I am in the anti-spam industry but I am not officially representing my employer.)}

There are two types of spam in this question.

The first two examples ("arugula" and "parroted comments") are Bayesian poisoning.

Bayesian poisoning is an attempt to hide spam content among ham content, which aspires to confuse machine learning spam filters.

 It does not actually work but spammers keep trying it, which is great: Bayes poisoning is actually easier to catch as spam with Bayes!

The third example ("kitchen units") has nothing off-topic (e.g. random quotes like the first two examples), and is quite brief. Bayes poisoning is defined by its off-topic or non-sequitur content and is almost always quite verbose, so this is not Bayes poisoning.

Kitchen Units For Sale. Thirty Ex Display Kitchens To Clear. www. e x d i s p l a y k i t c h e n s 1 .co.uk £ 595 Each with appliances.

This is snowshoe spam, which is named after the giant basket-like shoes that distribute your weight across the snow and thus prevent sinking into the snow with each step. This leaves a lighter footprint and is therefore harder to track. As I'll demonstrate below, snowshoe spam aspires to similarly tread lightly and be harder to notice.

(URL slightly corrupted so as to not encourage these folks.)

That caveat is actually really important. Snowshoe tends not to obfuscate its links much (since that makes victims less likely to click). Instead, the domain is used so briefly that the spam has already arrived in your inbox by the time URI DNSBLs can blacklist it.

Snowshoe generally has a short body, is selling something, and pretends to be a somewhat legitimate marketer. The current generation of snowshoe is limited to morally clean items (like kitchen units or garden hoses) rather than morally questionable items (like porn or drugs), but this could easily change. (The previous generation also included subscription management to look more like legitimate marketing, but this has basically gone away.)

Bayesian spam filtering is actually among the best tools to combat Snowshoe spam because it is so robust against morphing text.

^{(Disclaimer: I am in the anti-spam industry but I am not officially representing my employer.)}

There are two types of spam in this question.

The first two examples ("arugula" and "parroted comments") are Bayesian poisoning.

Bayesian poisoning is an attempt to hide spam content among ham content, which aspires to confuse machine learning spam filters.  It does not actually work.

The third example ("kitchen units") has nothing off-topic (e.g. random quotes like the first two examples), and is quite brief. Bayes poisoning is defined by its off-topic or non-sequitur content and is almost always quite verbose, so this is not Bayes poisoning.

Kitchen Units For Sale. Thirty Ex Display Kitchens To Clear. www. e x d i s p l a y k i t c h e n s 1 .co.uk £ 595 Each with appliances.

This is snowshoe spam, which is named after the giant basket-like shoes that distribute your weight across the snow and thus prevent sinking into the snow with each step. This leaves a lighter footprint and is therefore harder to track. Snowshoe spam aspires to similarly tread lightly and be harder to notice.

(URL slightly corrupted so as to not encourage these folks.)

That caveat is actually important. Snowshoe tends not to obfuscate its links much (since that makes victims less likely to click). Instead, the domain is used so briefly that the spam has already arrived in your inbox by the time URI DNSBLs can blacklist it.

Snowshoe spam generally has a short body, is selling something, and pretends to be a somewhat legitimate marketer. The current generation of snowshoe is limited to morally clean items (like kitchen units or garden hoses) rather than morally questionable items (like porn or drugs), but this could easily change.

Originally, snowshoe spam was very low volume in order to evade notice from spam traps, but spammers have learned that because trap-fed filters (such as DNSBLs) take a few minutes to propagate their knowledge, very high volume would work just fine if the entire spam campaign completed first. This fits the "tread lightly" principle that got this class of spam named even though it's less applicable nowadays.