N-Gram Analysis for Amazon Ads: Step-by-Step Guide

N-gram analysis is the step that turns a search term report from a list of rows into a list of words you can act on. In the accounts we manage, it is where we find the wasted spend that a row-by-row review left untouched.

This guide covers what an n-gram is, how to build the analysis in a spreadsheet from a 2026 unified reporting export, a fully worked example with the arithmetic, how to act on each result, the pitfalls, and how often to run it.

What Is N-Gram Analysis on Amazon?

N-gram analysis on Amazon is a way of reading the search term report by word instead of by row. You split every customer search term into fragments of one, two and three words, then add up the clicks, spend, purchases and sales of every term that contains each fragment.

The name comes from linguistics, where an n-gram is a sequence of n adjacent words. In Amazon advertising, three sizes do almost all the work:

• 1-gram (unigram): one word, such as "travel" or "thick".

• 2-gram (bigram): two adjacent words, such as "extra thick" or "yoga mat".

• 3-gram (trigram): three adjacent words, such as "extra thick yoga".

A search term of six words gives six 1-grams, five 2-grams and four 3-grams. Each fragment inherits that term's full clicks, spend, purchases and sales.

Then you total each fragment across the whole report. "Extra thick" now carries the combined performance of every search that contained it, in every campaign.

If the difference between a keyword and a search term is still fuzzy, read our guide to Amazon keywords vs search terms first. N-grams only work on search terms, the words shoppers actually typed.

Why N-Grams Find Waste a Term-by-Term Review Misses

A term-by-term review judges each search term on its own clicks. Most accounts negate a relevant term only after it reaches a click threshold with no orders, and that is the right rule for single terms.

The problem is the long tail. Amazon's search term report lists every search with at least one click, and in a typical account most rows have under 15 clicks. None of them ever crosses the threshold.

A word like "travel" can sit in 40 small search terms at once. Each looks harmless. Together they can spend more than your best exact keyword, for nothing.

The same logic finds winners. A two-word phrase that converts across many low-volume searches never earns two orders in any one row, so a harvest rule never promotes it.

Amazon's own targeting guide suggests evaluating a keyword after at least 20 clicks before adding it as a negative. N-grams let you reach that sample size on a word, even when no single search has.

How to Do N-Gram Analysis Step by Step

This is the method we use when a client account has no tooling, using Excel for Microsoft 365. Google Sheets works too, with SPLIT in place of TEXTSPLIT.

1. Export the Search Term Report From Unified Reporting

In 2026 the report lives in unified reporting, which Amazon made generally available on June 8, 2026. In Amazon Ads Agent, open Reporting, then Create report.

Choose Sponsored Products and Sponsored Brands as ad products. Add the dimensions Search term, Campaign, Ad group and Targeting, with the time unit set to a summary for the date range.

Add the metrics Clicks, Total cost, Purchases, Sales and Purchase rate over clicks. Our search terms report guide walks through every card in the builder.

Use 30 days for a routine pass and 60 days for a monthly deep dive. The Sponsored Products search term report has a 65-day lookback, so 60 is about the most you can pull in one go.

Leave out the most recent 7 days. Sales keep attributing to clicks for up to 7 days for sellers and 14 for vendors, so recent fragments look worse than they are.

The old Sponsored ads reports page stops taking new reports on December 17 and shuts down on December 31, 2026, per Amazon's migration page. Build this export in unified reporting now.

2. Clean the Search Terms

Put the cleaned term in column A, with Clicks, Total cost, Purchases and Sales in B to E. Then clean in four passes:

• Lowercase and trim every term: `=LOWER(TRIM(A2))`. "Yoga Mat" and "yoga mat " must count as one string.

• Remove ASIN rows. Entries like "b0c1xyz123" are product pages from auto and product targeting, not words. Filter out anything that starts with "b0" and is 10 characters long.

• Tag brand terms. Add a column marking rows that contain your brand name or a competitor's. You will judge those separately.

• Strip punctuation that splits words oddly, such as quotes and slashes.

3. Split Each Term Into 1-, 2- and 3-Grams

Each formula sits in one row and spills to the right. Put the 1-grams in column G, the 2-grams in column AA and the 3-grams in column AU so the spills don't collide.

TEXTSPLIT breaks the term at each space. TAKE with a negative count takes words from the end, so the 2-gram formula pairs each word with the one after it.

Copy the formulas down every row. Then, on a new sheet, list each unique fragment with `=LET(x,TOCOL(Terms!G2:Z5000,3),UNIQUE(FILTER(x,x<>"")))`, and repeat for the 2-gram and 3-gram columns.

4. Total the Metrics for Each Fragment

For each fragment in column A of the new sheet, add up the metrics of every search term that contains it as whole words:

The spaces around both sides of the SEARCH stop "mat" from matching "mats" or "format". Each search term counts once per fragment, however many times the word repeats in it.

Never average the report's Purchase rate over clicks column across rows. Recalculate it from summed purchases and summed clicks, and do the same for ACOS.

For files of more than a few thousand terms, the SUMPRODUCT sheet gets slow. Power Query or a short script does the same aggregation in seconds.

5. Filter, Sort and Decide

Sort by spend, then apply two filters before reading a single row:

• Waste: clicks at or above three times your clicks per order, and zero purchases or an ACOS far above target.

• Winners: at least two purchases and an ACOS at or under target.

Clicks per order is 1 ÷ your purchase rate. At a 10% purchase rate, a fragment needs about 30 clicks with no order before it is a fair negative candidate. Our negative keywords guide shows where the two-to-three-times rule comes from.

A Worked N-Gram Example With the Arithmetic

Here is a small, illustrative account: a brand selling one 6 mm non-slip yoga mat at about $39, with a 30% target ACOS. The table is 30 days of search terms, excluding the most recent 7.

The account runs at 24.1% ACOS ($300.60 ÷ $1,248). Its purchase rate is 9.6% (32 ÷ 334), so it expects one order every 10.4 clicks.

What the Term-by-Term Review Says

Three times 10.4 is a threshold of about 31 clicks. Only two terms pass it, and both convert well under target.

Every term with zero purchases has 14 clicks or fewer. A by-the-book review negates nothing this month.

What the N-Grams Say

"Travel" is the waste. Its 42 clicks should have produced about 4 orders (42 ÷ 10.4). At a 9.6% purchase rate, the chance of 42 clicks and no order by bad luck is 0.904 to the power of 42, about 1.4%.

The fragment cost $37.80, 12.6% of all spend, and no term alone came close to the threshold. The brand doesn't sell a thin, foldable travel mat, so a negative phrase on "travel" closes it in one line.

"Extra thick" is the winner. Two terms, 34 clicks, 5 purchases: a 14.7% purchase rate against the account's 9.6%, and a 15.7% ACOS ($30.60 ÷ $195).

Neither term had the two orders most harvest rules require. The fragment tells you shoppers who say "extra thick" are your buyers.

"Towel" needs judgement, not arithmetic. Its 23 clicks are under the 31-click threshold, and it has one order. The numbers alone say wait.

But a yoga towel is a different product, and the one order may well be a shopper who bought a mat instead. If you don't sell towels, relevance decides: negate it.

"Yoga mat" is noise. It appears in 8 of 10 terms, so its 24.5% ACOS is just the account average. Fragments that appear in almost every term tell you nothing.

One last check: never add the rows of an n-gram table together. "Thick", "extra thick" and "yoga mat" overlap, so the same click is counted in several rows.

How to Act on N-Gram Results

Each result maps to one of four actions. Which one depends on relevance first and numbers second.

1. Negative Phrase for Words That Never Fit

Use negative phrase when a fragment is wrong in every combination: a product you don't sell, a size you don't carry, "free", "used", "replacement". Amazon's negative targeting page says a negative phrase blocks any search that contains the complete phrase or close variations, up to 4 words and 80 characters.

Add it at the campaign level in every campaign where the fragment shows up, usually auto, broad and phrase. Before you save, filter the report for that fragment and read every term that contains it.

2. Negative Exact for Single Losers Inside a Good Fragment

Some fragments are mixed. "Thin" might fail in "travel yoga mat thin" and convert in "thin yoga mat for knees".

Don't negate the word. Negate the specific losing terms as negative exact, which blocks only that search and its close variants, up to 10 words.

3. New Exact Keywords for Winning Fragments

Take every search term that contains the winning fragment and converted, and add it as an exact keyword in your performance campaign. Then add it as negative exact in the campaign that found it, so two campaigns don't bid on the same search.

For "extra thick", that means "extra thick yoga mat for women" and "extra thick exercise mat" in exact. We also add "extra thick yoga mat" as a phrase keyword at a modest bid, to find the long-tail versions.

Our keyword harvesting funnel guide covers the bids and campaign layers. The match types guide explains how phrase and exact reach different searches.

4. Listing Copy for Words That Sell

A converting fragment is also a listing brief. If "extra thick" sells, it belongs in your title, bullets or Item Highlights.

Space is tighter in 2026. Amazon announced that from July 27, 2026, US titles outside media categories must be 75 characters or less, and added a searchable 125-character Item Highlights field. Winning 1-grams and 2-grams that don't fit the title go in Item Highlights or the backend search terms.

Relevant fragments with plenty of clicks and a weak purchase rate point the other way. When a "6 mm" fragment underperforms on a 6 mm mat, the listing probably doesn't make the thickness obvious in the main image or title.

Fragments Above Target but Converting

You can't bid on a fragment. If "for women" converts at 45% ACOS against a 30% target, lower bids on the keywords and targets that keep matching it, or move its best terms to exact with a lower bid. Our guide to reducing ACOS covers bid changes in detail.

N-Gram Analysis Pitfalls We See in Real Accounts

Most damage from n-grams comes from acting on a fragment without reading the terms behind it. These are the five we check every time.

1. A Bad Fragment Inside a Converting Phrase

A 1-gram can look terrible overall while one phrase containing it sells well. "Kids" might waste money in "kids yoga mat" and still convert in "yoga mat for kids and adults".

A negative phrase on the word kills both. Always open the full list of terms behind a fragment, and check its 2-grams and 3-grams before negating a 1-gram.

2. Brand Terms

Your own brand name converts cheaply and makes any fragment it travels with look better than it is. A competitor's brand name usually does the opposite.

Exclude rows containing your brand before judging generic fragments, and judge competitor brand fragments as their own group. Never negative-phrase your own brand on the strength of an n-gram.

3. Plural and Singular Forms

Your spreadsheet treats "towel" and "towels" as two fragments, each with half the clicks. Neither may reach the threshold alone.

Merge obvious plurals before judging. Amazon's negatives already match close variations such as plurals, so one negative phrase on "towel" also covers "towels".

4. Inferred Search Terms

Amazon's search term report page now says the report includes terms "inferred as best match to your product in non search context (such as off Amazon placements on social sites)." The change was in Amazon's documentation by May 2026, and inferred rows are not labeled.

So some fragments come from words no shopper typed. A fragment whose waste sits mostly in odd-looking rows under one well-performing target is a weak case for a negative phrase, which acts on every typed search that contains it.

5. Filler Words and Tiny Samples

"For", "with", "and" and "the" appear everywhere and mean nothing alone. Skip them as 1-grams, and read them inside 2-grams and 3-grams, where "for women" or "with strap" carry intent.

Fragments with a handful of clicks are noise too. Apply the same click threshold to a fragment that you would to a search term.

How Often to Run N-Gram Analysis

Run it on the same rhythm as your search term review, scaled to how many terms you have.

• Weekly, as part of the search term routine, for accounts with thousands of search terms or more than a few thousand dollars of monthly spend.

• Monthly for every other account, on 60 days of data, across all research campaigns.

• Daily during events such as Prime Day and Black Friday week, looking only for new irrelevant words. Event traffic brings odd searches fast.

• After a launch or a listing change, once the new product has a month of clicks.

Negatives can take up to 72 hours to apply, so judge the effect of a new negative phrase two weeks later, not the next morning.

N-grams also belong in every Amazon PPC audit. The first time we run them on a new account, the 1-gram list usually explains a large share of its wasted spend.

N-Grams on Walmart

Walmart Connect added negative keywords for Sponsored Products in July 2026, for manual campaigns and automatic campaigns using fixed bidding, with negative exact and negative phrase match. The same analysis now has a lever on Walmart too. Our Walmart keyword research guide covers the search side there.

How We Run N-Gram Analysis in Adbrew

The spreadsheet method works. On one account it takes an hour; across twenty accounts it is the first job that slips.

Adbrew is an agentic retail media management platform for Amazon and Walmart, and n-gram analysis runs inside it on your live search term data:

• N-grams in the search term workspace. See 1-, 2- and 3-grams across every campaign in Sponsored Ads management, filter by clicks, purchases and ACOS, and open the terms behind any fragment.

• Act from the same view. Add negatives or new keywords in bulk without exporting a file.

• Rules you set for harvesting and negation. Choose the lookback, click threshold and target ACOS, with brand terms excluded, and every change lands in Change History.

• Agents for the reading. Ask Adbrew Intelligence for "two-word n-grams with over 30 clicks and no purchases in the last 60 days, excluding brand terms". The same works from Claude or ChatGPT through the Adbrew MCP server, and nothing applies until you approve it.

Brands on Adbrew cut ACOS by 46% on average, and Seller Presto saves more than 250 hours of manual optimization a month.

Read Your Search Terms by Word

The search term report tells you what shoppers typed. N-grams tell you which words in those searches cost you money and which ones make it.

Run the five steps on your last 60 days this week, and start with the 1-grams sorted by spend. Want n-grams, harvesting and negation running on rules you control? Book a demo and we'll run an n-gram pass on your account with you.