N-gram analysis takes every search term your account has paid for and regroups them by the words they share, splitting each into 1-, 2-, and 3-word patterns and totalling spend, clicks, orders, and sales across every term a pattern appears in. Instead of judging thousands of near-unique search terms one row at a time, you judge the words they are built from, and words repeat often enough to give you real evidence. In the accounts I audit, this is the fastest way to find waste, because most bleed does not arrive as one expensive search term. It arrives as one bad word spread thin across a hundred cheap ones. This guide covers what an n-gram is, why the word-level read beats term-by-term optimization, the six steps I run it in, and how the output turns into phrase negatives on the losing patterns and harvests on the winning ones.
What an n-gram is: 1-, 2-, and 3-word patterns
An n-gram is a sequence of n consecutive words taken from a search term. A 1-gram (unigram) is a single word like "orthopedic." A 2-gram (bigram) is a word pair like "dog bed." A 3-gram (trigram) is a triple like "orthopedic dog bed." N-gram analysis splits every search term into these fragments and totals performance across each one.
Split one term and the mechanics are obvious. Take "orthopedic dog bed for large dogs":
| Pattern size | What "orthopedic dog bed for large dogs" produces |
|---|---|
| Unigrams (1-word) | orthopedic, dog, bed, for, large, dogs |
| Bigrams (2-word) | orthopedic dog, dog bed, bed for, for large, large dogs |
| Trigrams (3-word) | orthopedic dog bed, dog bed for, bed for large, for large dogs |
Every fragment inherits its parent term's metrics, and when the same fragment appears in other search terms, the metrics add. The row for "bed" carries the combined spend, clicks, and orders of every search in the account containing the word: one row per pattern, the whole account's evidence behind it.
Why word patterns beat term-by-term optimization
Search terms are near-unique; the words inside them repeat. A single search term rarely collects enough clicks for an honest verdict, but a word that appears in two hundred terms carries the combined evidence of all of them. N-gram analysis moves the decision to the level where the statistics actually exist.
Term-by-term optimization loses to the long tail by design: broad match, phrase match, and auto targeting mint new variants continuously, so negating individual terms is a subscription. The sort-by-spend review most sellers run misses the leak too, because the waste is not concentrated. If a word shows up across fourteen non-converting search terms and zero converters, no individual row ever looks bad enough to act on. The word is the problem, and it never gets its own row in any report Amazon gives you. N-gram analysis writes that row.
The aggregation also crosses campaign boundaries. The search term report arrives grouped by campaign, so a bad word can bleed a little in six campaigns and a lot in none of them; pooled into one pattern table, its full bill finally appears on one line. The report itself is the raw material; my search term report guide covers reading it row by row, and n-gram analysis is how I process it at scale.
How to run n-gram analysis, step by step
The workflow is six steps: pull a 60-day search term report, merge duplicates, strip what does not belong, split every term into 1-, 2-, and 3-word patterns, aggregate the metrics per pattern, then sort by spend and drill down before acting. My free N-Gram Analyzer runs the middle steps in your browser in about 60 seconds.
- Pull the 60-day search term report. Amazon Ads > Bulk Operations > Download, with Sponsored Products Search Term Data ticked (add Sponsored Brands if you run it). Sixty days, not thirty, so slow converters get a fair trial.
- Merge duplicate search terms across campaigns. If several campaigns bought the same search, aggregate its rows into one first; fragmented rows understate every pattern they feed. Heavy duplication is itself a finding: campaigns buying each other's traffic, the subject of my search term bleed guide.
- Strip what does not belong. Your own brand terms come out (loyalty traffic pollutes every winners list), ASIN rows from product targeting come out (they are not words), and rows with zero clicks can go too; they add noise, not insight.
- Split every term into its n-grams. Each remaining search term decomposes into unigrams, bigrams, and trigrams, and each fragment inherits the term's full metrics.
- Aggregate per pattern. For every unique pattern, sum spend, clicks, orders, and sales across all the terms containing it, count how many terms it appears in, and compute ACOS.
- Sort by spend, filter orders to zero, and read. The verdicts are the next section.
You can build this in a spreadsheet with SPLIT and SUMPRODUCT for a small account. The N-Gram Analyzer exists because most accounts are not small: upload the same bulk file and it returns 1-, 2-, and 3-word tabs with a non-converting filter and CSV export, parsed in your browser, no email, no account. The same engine runs the n-gram panel inside my free Audit Dashboard, one of the nine checks it performs on a bulk file.
Reading the output: four verdicts, one human question
Every pattern in the output lands in one of four buckets: irrelevant losers to negate, relevant but expensive patterns to reprice, proven winners to harvest, and low-data patterns to leave alone. Spend and orders make the first cut. Relevance decides the verdict, and relevance is a judgment no spreadsheet can compute.
| What the pattern shows | Relevant to your product? | The verdict |
|---|---|---|
| Real spend, zero or near-zero orders | No | Negative phrase match on the pattern |
| Real spend, orders above your target ACOS | Yes | Bid cuts on the targets buying it. Never negation |
| Strong orders at healthy efficiency | Yes | Harvest exact keywords; carry the words into your listing |
| Spend below the fair-judgment threshold | Either | Leave it alone; it has not had its trial yet |
The threshold in that last row is aCTC times CPC: average clicks to conversion (one divided by your conversion rate) multiplied by your average cost per click. That is what one average order's worth of clicks costs in your account, and I treat it as the price of a fair trial. A zero-order pattern under it has not earned a verdict; a zero-order pattern past it has had its trial and failed.
The cardinal sin is negating on ACOS alone. A word with an ugly aggregate ACOS usually mixes relevant and irrelevant searches, and negating the word kills both. So before any pattern becomes a negative, I read the actual search terms containing it. Categorically wrong for the product (wrong material, wrong species, wrong audience): negate. A meaningful share converts: the word stays, and either the bids come down using the repricing logic from my guide on lowering ACOS without killing sales, or I step down to a narrower pattern.
Losing patterns become phrase negatives (mind the blast radius)
A losing pattern becomes one negative phrase match, which blocks every current and future search term containing it. That is the leverage and the risk in the same move. Phrase negatives fail silently, so a pattern earns one only when it fails across many terms and appears inside zero converters.
Blast radius scales inversely with pattern length. A unigram negative is the widest net you can throw: one word, and every search containing it is gone, with no report ever showing what the block cost. So I read the table starting at unigrams, where the big totals live, but I negate at the narrowest width that still covers the waste. When a single word is too broad because one combination of it converts, I layer bigrams instead: a bakeware seller whose report is full of "dish rack," "wine rack," and "shoe rack" searches can negate those three word pairs and leave "cooling rack" alive.
Placement follows the same logic as any negative: campaign level, in the campaigns that bought the traffic, with account-level lists reserved for words wrong for everything you sell. Amazon offers exactly two negative match types, exact and phrase, no negative broad; my negative keywords guide covers the decision rule in full, and the free Negative Keyword Finder builds the ready-to-upload list from your own file.
Winning patterns are the other half of the read
The same aggregation that finds waste also finds strength. Patterns with strong orders at healthy efficiency are the words shoppers actually use when they buy your product. Harvest exact match keywords around them, check that your title and bullets carry them, and point new keyword research at them first.
A 2-word pattern converting across dozens of different long-tail searches means you under-own that theme at exact match, where you control the bid and placement directly instead of renting the traffic through discovery settings. It is also the cheapest listing research you will ever get: these are the words your own customers paid to prove. Concentrating spend on proven words also pushes organic rank, the mechanic my ACOS vs TACOS guide walks through on a real account.
A worked example on real numbers
Take the dataset I grade across these guides: $36,303 in ad spend, $111,058 in ad sales, 4,140 orders from 45,672 clicks. That computes to a 9.1% conversion rate, a $0.79 average CPC, and a 32.7% ACOS, and those averages set the whole n-gram read before you open a single tab.
At a 9.1% conversion rate, an average order takes roughly eleven clicks, and eleven clicks at $0.79 is $8.69 of spend. That is this account's fair-judgment threshold. Any pattern that spent past $8.69 with zero orders across every search term containing it has had a full, fair trial and failed it; anything under that line gets left alone. The N-Gram Analyzer computes the same threshold from your own file and flags every pattern past it.
The winning side of the read is worth seeing on real data too. The sample report built into the tool runs on real, anonymized numbers from a $110K/mo pet-bed brand, and its strongest 2-word patterns look like this:
| 2-word pattern | Spend | Orders |
|---|---|---|
| pet bed | $5,811 | 753 |
| anti slip | $2,810 | 325 |
| bed cover | $2,189 | 309 |
| donut bed | $1,569 | 182 |
Four patterns, each aggregating every search term that contains it, each hundreds of orders deep or close to it. That is a harvest list and half a product title writing itself. The losing side of the same report is the negation list, and I am not going to print invented losers to make the point prettier: the free audit demo renders the full read, winners and losers, on real rows. For scale, my benchmarks guide puts tolerable zero-order spend at 50–60% during launch, tightening to 25–30% at maturity, and the free Wasted Spend Finder sizes your leak in dollars while the n-gram read names the words responsible for it.
Frequently asked questions
What is an n-gram in Amazon PPC?
An n-gram is a sequence of n consecutive words pulled from a customer search term: 1-word (unigram), 2-word (bigram), and 3-word (trigram) patterns. N-gram analysis splits every search term in the account into these fragments and aggregates spend, clicks, orders, and sales across every term containing each one, so you judge word patterns instead of individual searches.
What is the difference between a unigram, a bigram, and a trigram?
Length, and therefore blast radius. A unigram is one word, a bigram two consecutive words, a trigram three. Unigrams aggregate the most data and surface the broadest themes, but they are the riskiest to negate because one word can live inside both junk and converters. Bigrams add context with fewer false positives. Trigrams mostly confirm what the shorter patterns suggest.
How much data do I need before n-gram analysis is worth running?
A 60-day search term report with enough traffic for words to repeat across many terms. If your account only has a few hundred search terms, read the report directly; n-grams add little. The pattern-level read starts paying once the report has too many rows to judge honestly one at a time.
How do I run n-gram analysis without Excel or Python?
Upload your Amazon Ads bulk file to my free N-Gram Analyzer. It parses in your browser, groups every search term into 1-, 2-, and 3-word patterns, aggregates the metrics, flags non-converting patterns against a spend threshold computed from your own conversion rate and CPC, and exports each tab to CSV. No email, no account. The same engine runs the n-gram panel in the free Audit Dashboard.
Can n-gram analysis find keywords to target, or only negatives?
Both. The negation side gets the attention, but patterns that convert strongly across many different search terms are proof of what shoppers actually call your product. Those words are your best candidates for exact match harvests, for new keyword research, and for your title and bullets. It is the same table, read from the other end.
How often should I run n-gram analysis on Amazon?
Immediately on any account you take over or audit, because the first pass on untouched data finds patterns that have leaked for months. After that, monthly while campaigns are changing or discovery is running hot, and quarterly once the account is stable. Always over a 60-day window so slow converters get a fair trial before their words are judged.
See which words your budget is actually buying
The word-level read takes about 60 seconds to start: upload your bulk file to the free N-Gram Analyzer, or run the free Account Health Snapshot first to grade all nine metrics at once. Both parse in your browser, no email, no account. And when the table hands you the calls no tool should make for you (negate or reprice, unigram or bigram, which campaign), the free 30-minute diagnosis call is where I make them with you, pattern by pattern.