Amazon publishes an example of broad match in its own targeting guide. The keyword is sneakers. Amazon says it can match:
“canvas sneakers,” “sneaker,” “basketball shoes,” “athletic shoes,” “cleats,” “trainers,” or “foam runners”
Read that list again. “Cleats” and “foam runners” share no words at all with “sneakers.” That’s not a bug or an edge case — it’s Amazon documenting its own behaviour. Broad match treats your keyword as a topic, and it decides what else belongs to that topic.
If you sell premium leather trainers, you are now bidding on football boots.
How each match type actually behaves
Straight from Amazon’s documentation:
| Match type | Amazon’s own wording |
|---|---|
| Broad | “offers your ad broad exposure”; a query “can contain the keyword terms in any order“; covers “singulars, plurals, variations, synonyms, and related terms“ |
| Phrase | “search terms matched will contain all components of the targeted keyword in the same order“ |
| Exact | “search terms are matched word for word (same words and in the same order)” |
And underneath all three sits the close-variant layer. Amazon’s wording: “With manual targeting campaigns you have three options. They all account for plurals, misspellings, and translations.“
That’s the whole documented close-variant surface on Amazon: plurals, misspellings, translations — applied to every match type and not switchable off. It’s much narrower than Google’s, which matters more than you’d think.
One documented exception, and it’s Sponsored Brands only. Amazon states that in Sponsored Brands campaigns, phrase match “also allows your ad to show on searches that have similar meaning or intent as your keyword, even if the exact wording differs.” Their example: phrase match curtain rods triggering on “pole for curtains for bedroom.” Zero token overlap, on phrase match.
So Sponsored Brands phrase match is semantically expanded and Sponsored Products phrase match is not. If you have been treating them as the same control, they aren’t.
The 2026 problem: placements you were enrolled in automatically
Match types haven’t changed. I read Amazon’s Ads API release notes through to 28 August 2026 and there is no release note altering broad, phrase or exact matching logic in 2025 or 2026. Any post telling you “broad match got smarter in 2026” is asserting something with no source behind it.
What did change is where your ads run, and you weren’t asked.
Off-Amazon creator placements, 10 August 2026. Sponsored Products campaigns began serving on creator placements outside Amazon, with existing campaigns enrolled automatically at current bids and budgets and no advertiser action required. The campaign setting is “Settings for ads served off Amazon”, where the default is “Increase reach” and the alternative is “Limit off-Amazon spend.” Top of Search and Product Pages bid adjustments don’t apply to these placements, though dynamic bidding does. Markets include the US, Canada, Mexico, Brazil, India, Turkey and parts of the EU — the UK is not on the list, which matters if you’re reading US commentary about it.
Here’s the part that belongs in a post about drift: the search terms on those placements are inferred by Amazon, not typed by a shopper. Amazon’s own documentation acknowledges this generally — “When a customer search term is not available, Amazon will infer a search term from the context that best matches your keywords.” You can negate against them, but you’re negating against Amazon’s guess at intent rather than a real query.
Prompt Ad Extensions. Amazon’s documentation: “Prompts are a new ad format that integrates into your existing Sponsored Products and Sponsored Brands campaigns with zero additional setup required.” Reporting exists — spPromptAdExtension and sbPromptAdExtension report types, 90-day maximum range, and individual prompts can be paused in the console. But it’s another surface you were opted into.
And one you cannot see at all. A September 2024 release note states that ads may appear in Rufus-related placements, and then this: “Your campaign reports will not include Rufus metrics.” I found no later note reversing it. Amazon has since folded Rufus into Alexa for Shopping, but if that reporting exclusion still holds, it is a category of spend that appears in no report you can pull.
The negative keyword limits everyone quotes are wrong
Widely cited guidance says 1,000 negative keywords per campaign and per ad group. Amazon’s live limits page says 10,000 for both. That’s a tenfold difference, and most published advice is running on a figure from 2023.
The full picture, from Amazon’s campaign limits page:
| Limit | Value |
|---|---|
| Campaigns per account | 10,000 |
| Ad groups per campaign | 20,000 |
| Keywords per ad group (manual, excluding negatives) | 1,000 |
| Campaign negative keywords | 10,000 |
| Ad group negative keywords | 10,000 |
| Negative phrase length | 4 words or 80 characters |
| Negative exact length | 10 words or 80 characters |
Two things follow from this. First, negative keyword capacity is not your constraint — if your list is short, that’s a process problem, not a platform one. Second, there is no negative broad match. Only negativeExact and negativePhrase exist. That asymmetry is the whole problem: broad match can expand into synonyms and related terms, but your negatives cannot.
And negatives are themselves subject to close variants. Amazon says negative phrase blocks queries containing “the complete phrase or close variations“, negative exact likewise. So a negative can over-block as well as under-block — worth knowing before you negate a term that shares a plural with something you want.
Working the search term report
Reports → Advertising Reports → Search Term. Before you rely on it, understand three constraints Amazon publishes:
“Search term reports only include impressions that resulted in at least one ad click.” Impressions with no clicks are invisible. You’re seeing the spending part of drift, not the full extent of it.
Product page placements show an asterisk. “If a placement does not have a search keyword associated with it on a product detail page, the search term in the report will be an asterisk *.” Those rows are real spend you cannot attribute to a query.
There is no Sponsored Display search term report at all. SD doesn’t appear in the report-type matrix.
Then the timing. Data retention is 65 days for Sponsored Products and 60 for Sponsored Brands, with a 31-day maximum range per report. Amazon’s published P99 latencies: impression and click events available within 12 hours, but invalidation takes up to 72 hours. Amazon reserves three days to strip out invalid traffic.
That’s the hard number behind “don’t optimise on fresh data.” Yesterday’s report is not yesterday’s truth, and because sales are attributed to click date rather than purchase date, rows keep backfilling after you’ve looked at them. Every fresh report is a pessimistic version of your own performance.
Work on data at least 72 hours old, and preferably a fortnight, so the attribution window has closed.
N-gram analysis: finding the waste no single term reveals
The reason drift is hard to catch is that it rarely concentrates. Most search terms get fewer than five clicks. None of them individually crosses a threshold that would make you negate it. Collectively they’re expensive.
N-gram analysis fixes that by aggregating on the shared words rather than the whole query:
- Download the search term report
- Split every search term into monograms, bigrams and trigrams
- Aggregate impressions, clicks, cost, orders and sales across every search term containing that gram
- Rank grams by spend with zero or negligible conversions
- Negate at gram level
Suddenly “kids”, “toddler” or “refill” shows up as £400 of spend across ninety search terms that individually looked like noise.
Run it monthly, not weekly. It needs accumulation to be meaningful, and the 72-hour invalidation window makes fresh data unreliable anyway.
Setting a negation threshold that isn’t arbitrary
You’ll see flat rules published — negate at $35 spend with no conversions, or 20 to 30 clicks with no conversions. They’re fine as a starting point but they ignore your conversion rate, which is the only thing that determines how many clicks a converting term should need.
The defensible version:
Expected clicks per order = 1 ÷ conversion rate. Allow 2–3× that before negating.
At a 10% conversion rate, one order should take about ten clicks, so negate somewhere between 20 and 30. At 4%, one order takes 25 clicks, and negating at 20 would be cutting terms before they’ve had a fair test. At 20%, waiting for 30 clicks wastes money on terms you could have called at 12.
Same logic, different answer per product. That’s the point.
The 28–40% figure: don’t quote it
You’ll see it stated as fact that the average Amazon seller wastes 28–40% of monthly ad spend on search terms that click but never convert.
I traced it. One agency cites another; that one attributes it to “PPC audit data from industry analysts” with no study, no sample size, no methodology and no link. Everyone downstream cites the second agency.
There is no published study with a methodology quantifying the share of Amazon ad spend going to irrelevant terms. That’s a genuine gap. Run your own n-gram analysis and you’ll have a number for your account that’s worth more than a figure nobody can source.
The routine
Monthly, on data at least 72 hours old:
- Pull the search term report for the last 30 days
- Run n-gram aggregation across mono-, bi- and trigrams
- Negate grams above 2–3× your expected-clicks-per-order with no conversions, using negative phrase where a whole concept is wrong and negative exact where a single query is
- Promote search terms with two or more conversions below break-even ACoS into their own exact-match targeting
- Check your off-Amazon and Prompt placement settings — you were enrolled by default
- Recheck negatives for over-blocking, since close variants apply to them too
Drift isn’t a fault you fix once. It’s a property of how broad match works, and your campaign architecture decides how much of it you’re exposed to, and containment is a routine rather than a project.
FAQ
Why are my Amazon ads showing on irrelevant search terms? Because broad match treats your keyword as a topic rather than a directive. Amazon’s own documentation shows sneakers matching “cleats” and “foam runners” — terms sharing no words with the keyword. Phrase and exact are tighter, but all three include plurals, misspellings and translations.
How many negative keywords can I add on Amazon? 10,000 at campaign level and 10,000 at ad group level. Much published guidance still says 1,000, which is out of date.
Is there a negative broad match on Amazon? No. Only negative exact and negative phrase exist. Negative phrase allows up to 4 words or 80 characters; negative exact up to 10 words or 80 characters.
How fresh is Amazon search term data? Amazon publishes a 12-hour P99 for impressions and clicks to become available, but up to 72 hours for invalidation of invalid traffic. Data retention is 65 days for Sponsored Products, 60 for Sponsored Brands, with a 31-day maximum report range.
How many clicks before I negate a keyword? Divide 1 by your conversion rate to get expected clicks per order, then allow two to three times that. At 10% conversion, negate at 20–30 clicks with no orders. Flat thresholds ignore your conversion rate.
Sami Sultan runs Ecom Enable, an Amazon growth partner for DTC brands. Match type definitions, campaign limits and reporting constraints verified against Amazon Ads documentation on 27 August 2026.
Sources: Amazon Ads targeting guide; Set up keyword targeting; Sponsored ads campaign limits; Amazon Ads API search term report documentation and release notes.