# N-Gram Analysis: Finding Patterns in Search Terms

What n-gram analysis is in PPC, why fragments beat whole queries for finding waste, and how AI turned a script ritual into a conversation.

Glossary · Published 2026-08-10.

> N-gram analysis breaks every search term in an account into overlapping word fragments — unigrams (one word), bigrams (two), trigrams (three) — and aggregates spend and conversions by fragment instead of by query. One bad query is an anecdote; the word 'free' carrying spend across four hundred losing queries is a diagnosis. It finds the patterns that row-by-row search term review structurally misses.

N-gram analysis is a way of reading the [search terms report](https://adcopilot.cloud/glossary/search-terms-report) by fragment instead of by row: every query is split into its overlapping word sequences — unigrams (single words), bigrams (pairs), trigrams (triples) — and performance is summed per fragment across the whole account. The query "free crm software download" contributes to *free*, *crm*, *software*, *download*, *free crm*, *crm software* and so on; each fragment accumulates the spend, clicks and conversions of every query containing it. Sort fragments by unconverting spend, and the account's vocabulary problems introduce themselves.

## How n-gram analysis works

The mechanism is aggregation attacking sparsity. A busy account's search terms table runs to thousands of rows, most individually too small to judge — three clicks here, one there, nothing significant anywhere. But the same words recur across those rows, and at fragment level the sample sizes become honest: queries containing *manual*, individually negligible, may jointly represent real money and no conversions. Fragment-level review finds what query-level review structurally cannot, because the pattern never sits inside any single row.

Why fragments beat queries is worth stating plainly: a query is an event, a fragment is a theme. Decisions about negatives, bids and copy are decisions about themes — and n-grams are simply performance reporting at the theme level.

The classic implementation was a script: export the terms, tokenise, pivot, sort. It worked, and it had a ceiling — static output, no follow-up questions, analysis divorced from action. You learned that *free* was bleeding; the negatives still had to be designed, checked against converting terms, and added by hand. Next quarter, if the ritual survived staff turnover, you learned it again.

## From script ritual to conversation

With an AI agent connected to the account, n-grams stop being a script you run and become a sentence you speak: "pull ninety days of search terms, break them into one-to-three-word fragments, and rank fragments by spend without conversions." The follow-ups — the part spreadsheets never did — happen in the same conversation. Does *free* ever convert anywhere? Which campaigns does it bleed in? What would a phrase negative block that an exact negative would not?

Then fragment becomes action in seconds, with receipts. Losing fragments become staged [negative keywords](https://adcopilot.cloud/negative-keywords-automation) you approve, at the right level and match type. Converting fragments feed bids and copy — if *same day* converts, say *same day* in the headline. Next month's rerun shows whether the fence held. The full worked pattern, prompts included, is in the [n-gram walkthrough](https://adcopilot.cloud/blog/search-terms-ngram-analysis-ai).

The quiet upgrade is evidentiary. "This query looks bad" starts an argument; "this word carried this much spend to zero conversions across hundreds of queries" ends one. N-grams were always the cheapest way to manufacture that class of evidence. AI removed the last excuse for not running them.

## Questions

### How many words should an n-gram be?

One to three covers nearly all the value. Unigrams surface the big thematic offenders, bigrams add the context that changes meaning — 'free shipping' is a different signal from 'free download' — and trigrams occasionally catch a phrase pattern the shorter cuts blur. Beyond three, fragments approach the original queries and the aggregation stops aggregating; the sample per fragment thins back toward anecdote.

### Does n-gram analysis work for Performance Max?

Partially, and honesty matters here. PMax exposes search themes and category-level insights rather than a complete query-by-query report, so fragment analysis runs only on what is visible — useful for spotting thematic drift, incomplete by construction. Where search campaigns overlap PMax, the search campaigns' terms often carry the readable signal for both. Treat PMax n-grams as directional, not exhaustive.

---

This is the markdown copy of https://adcopilot.cloud/glossary/n-gram-analysis. The index of the site's pages is https://adcopilot.cloud/llms.txt.
