Amazon Stops Reading Your Search Terms at Byte 250
We ran 1,238 searches against 117 live listings in 13 Amazon marketplaces to find out what happens to backend keywords past 250 bytes. Amazon reads the first 250 bytes and ignores the rest.
Every Amazon listing has a field shoppers never see. Sellers call it backend keywords and Amazon calls it search terms. It's where the synonyms, misspellings and other-language words go that would ruin a title, and it comes with a rule everyone quotes and nobody agrees on: 250 bytes.
The field doesn't enforce the rule. Across 47,257 listings in our data, one search-terms field in seven runs past 250 bytes, and nine run past 2,000. So what happens to the extra words? Ask around and you'll get three answers:
- Amazon reads all of it.
- Amazon reads up to 500 bytes.
- Go one byte over and Amazon throws the whole field away.
None of those answers come with evidence. So we went and got some: 1,238 searches against 117 live listings in 13 Amazon marketplaces, each asking Amazon a single yes-or-no question about a single word.
The short version
Amazon reads the first 250 bytes of the field and nothing after. A word written past that point is found about as often as a word that isn't in the listing at all.
Picture a typical over-limit field. Amazon reads up to byte 250, which here lands in the middle of a word, and ignores everything after it.
How do you ask Amazon what it read?
Amazon doesn't publish its search index, but the catalog search in its Selling Partner API will answer a narrow question if you phrase it right. Search a marketplace for a product's ASIN plus one word, and the product comes back only if Amazon matches that word to it. The ASIN on its own found its product every time (117 of 117). A made-up word next to the ASIN never did (0 of 84).
There's a catch, and it's why most home-made tests of this question go wrong. A word can bring a product back from anywhere in the listing. If "stainless" sits in your title and again past byte 250 of your search terms, the product comes back for it, and you've learned nothing about the search terms.
So we only tested words that live in one place. Before every search we pulled the listing's current content from Amazon and checked each word against the title, bullets, description, brand, every other attribute, the browse category and the seller's other listings of the same product. We caught plurals, inflected forms, close misspellings and words buried inside compounds, in each marketplace's own language. When a word like that comes back, only the search terms can have put it there.
Every listing also carried four controls:
- Its ASIN alone. If Amazon doesn't return the product for its own ASIN, the listing is out.
- A title word. It should almost always come back.
- A made-up word. It should never come back.
- Two words the same seller uses in other listings, but never in this one. These turned out to be the most useful number in the study. They show how often Amazon returns a product for a word it has no business matching, the noise floor every result has to clear.
The first surprise
Before we could find where Amazon stops reading, something got in the way. About two fields in five never came back for any of their words, not even the first few.
It had nothing to do with length. Fields under the limit went quiet as often as fields over it (40% against 38%), and those listings came back less often for their own title words too: 8 times out of 17, against 25 out of 26 everywhere else. Something about those listings keeps their search terms out of search, and whatever it is, it isn't the 250-byte rule.
A field Amazon isn't searching can't tell you where its limit is, so we set those aside and kept the fields that clearly were being searched, where at least two words near the start came back. To keep that rule honest, each early word was judged by its field's other words, never by itself.
The wall at byte 250
In the fields Amazon searches, the answer isn't subtle:
- Words in the first 250 bytes were found 82 times out of 88.
- Words that start after byte 250 were found once in 59.
- The seller's own words from other listings, which have no business matching at all, were found at the same rate: once in 42.
The drop doesn't fade in gradually. It happens all at once, right at the line. Words that end in the last ten bytes before byte 250 were found 12 times out of 14, and words that start in the first ten bytes after it were found 0 times out of 13.
It isn't a quirk of a few listings either. Resample the listings and the found rate stays between 86% and 99% before the line and between 0% and 5% after it. In 15 of the 16 listings where we tested both sides, more words were found before the line than after it, and in none was it the other way round. The line sits in the same place in every marketplace group we tested, and in fields of every size up to 500 bytes.
Bytes, not characters
The rule says bytes, and it means bytes. In UTF-8, the encoding the limit is counted in, a plain English letter takes one byte. Accented letters take two, and a euro sign or a Chinese or Korean character takes three.
So in a French, German, Polish or Turkish field, the 250th byte arrives before the 250th character. That gave us a clean test: words that fit within 250 characters but not within 250 bytes. If Amazon counted characters, they'd be found. None of the 8 we tested were. When we fitted the cut-off under six different ways of measuring a field, counting bytes beat counting characters by a likelihood ratio of about 800,000 to one.
Two more pieces of folklore fell the same way:
- "Spaces don't count." If that were true, words that fit within 250 bytes of non-space text would be found. None of 25 were. Spaces count.
- "Repeating a word is free." If that were true, words that fit once the repeats are taken out would be found. One of 32 was, which is the noise floor. Repeats count too.
The one thing the data can't settle is whether the 250th byte itself makes the cut. No word in our sample ended exactly on it.
Going over doesn't break the field
The scariest version of the rule says one byte over and Amazon ignores the whole field. It doesn't. 34 of the 55 over-limit fields we could classify were clearly being searched, and their early words were found 93% of the time. Fields went unsearched just as often under the limit as over it. Amazon keeps the first 250 bytes and drops the rest.
Why the myth survives
If words past byte 250 do nothing, why are so many sellers sure they work? Because they've checked, and the check said yes.
Most of the words sellers write past byte 250, 57% of them, also appear somewhere else in the listing. Search for the product with one of those words and it often comes back, through the title 19 times out of 21 and through the bullets 8 times out of 17. Test every word past byte 250 that way and 39% of them look indexed. Almost none of them are.
The search terms didn't earn those hits. The title did.
A loose thread
Two fields in our sample mixed Chinese and Korean words into English listings on Amazon.com, and in both the wall moved. It fell later than byte 250, by about as many bytes as those words take. Five of the six words that sit past byte 250, but inside it once the Chinese and Korean words are skipped, were found. None of the four words after that point were. It looks as though Amazon drops words it doesn't index on a marketplace before it starts counting. Two fields from one seller make a lead, not a finding, so we'll keep watching this one.
What to do with your search terms
- Put everything that matters in the first 250 bytes. Treat the rest as invisible.
- Count bytes, not characters. Every accented letter costs an extra byte and a Chinese or Korean character costs three, so 250 characters of French or German is almost certainly over the limit.
- Don't spend bytes on spaces or repeats. One space between terms, and each word once.
- Leave out what the title already says. Title words were found 25 times out of 26 without any help from the search terms.
- Test with words that live only in the field. A word that also appears in your title will tell you everything works.
Questions or feedback? Reach out to us at [email protected].
