Amazon prices by exit IP, and the parser everyone writes deletes the only clue

The two currency mixups in this log are not equally survivable. A Vietnamese row landing in a euro column reads 390,037 against 8.63, a factor of roughly 45,000 ^[2][1], and any range check ever written will stop it. A dollar row landing in a euro column reads 16.8 percent high, which is what the cookie ratios of 1.168, 1.167 and 1.168 in the dev.to write-up work out to ^[5][4]. Sixteen percent is a promotion ending. It clears every plausibility test and moves the average instead of breaking the job.

The same invisibility applies to the number itself. 1,234.56 and 1.234,56 are the same amount, and the usual cleanup of deleting commas turns 8,63 into 863 ^[13], which is a believable price for something. The rule that survives is positional: the decimal separator is whichever of the dot and the comma comes last, and only when exactly two digits follow it ^[14]. A regex anchored to a leading symbol, meanwhile, returns nothing at all in Poland (123,45 zl) or Sweden (1 299 kr) ^[15], which the author hit while testing all 18 marketplaces ^[16].

Byte count is the part worth stealing. Amazon's anti-bot interstitial arrives as HTTP 200 with Content-Type text/html at about 2,265 bytes, carrying none of the classic challenge strings, and on amazon.fr the same thing comes back as 202 ^[11]. Real search pages run 300 KB to 1.6 MB ^[12]. That puts the challenge page at roughly one one-hundred-and-thirtieth of the smallest legitimate response ^[2], and the suggested 50 KB floor sits about 22 times above the interstitial while staying a sixth below the smallest real page ^[3]. Margins that wide are why size works where the status line does not.

The rate numbers are one developer's, from a week of logging, and he says the single-IP burn-out figure is not published anywhere else ^[18]. Hourly probing from one datacenter IP at two requests per check produced two blocks in fourteen requests, and the baseline decayed across the day from two working methods to none ^[8]. Two requests seven minutes apart were both blocked and cost that address hours ^[9]. His reading is that Amazon weighs density per address far above total volume, so the unit should be a fresh exit per request rather than a session per run ^[10].

That prescription and the currency problem pull against each other. Maximum rotation is precisely the condition under which every request can land in a different pricing country ^[1], so the correctness work has to move into the parser: keep the a-offscreen symbol, resolve it to an ISO code, and compare it with the currency the requested marketplace should have returned ^[3][17]. Pinning the pool per marketplace gets you that comparison cheaply ^[17]. Either way the mismatch has to be flagged rather than dropped, because nothing downstream is ever going to notice it.