How to analyse the search terms report on commercial grounds

The minimum amount of data is a sum of money

Zero orders only mean something once the cost has reached around three times the contribution margin per order. Before that, even a search term that exactly pays for itself may simply have been unlucky.

A group of terms breaks even when its conversion rate is at least CPC ÷ contribution margin. The rule of three says: if an event never occurs in n trials, the 95% interval for its rate runs from 0 to 3/n, a good approximation from about 30 trials. The signal is therefore reliable as soon as 3/n falls below CPC ÷ contribution margin. Rearrange that and the cost per click cancels out:

Exclusion signal: cost without an order ≥ 3 × contribution margin per order

That makes the click the wrong unit. Rules you will come across include 30 clicks in 60 days, or the filter "zero conversions, cost above target CPA". Both are useful as a checklist. As a rule for exclusions, they kick in too early or too late depending on margin and cost per click, as the worked example with invented, round numbers shows:

The last column is my approximation using e to the power of minus cost ÷ contribution margin. If your target CPA equals your contribution margin, the CPA filter throws away more than one in three break-even terms. The 30-click rule only hits the threshold when a click costs a tenth of the contribution margin. With cheaper clicks it judges too early, with more expensive ones too late.

Which amount belongs in the calculation is explained in contribution margin for advertising decisions: after goods, shipping, fees and returns, before advertising. If you expect more than break-even, use the allowable ad spend per order instead of the contribution margin. The threshold then falls accordingly.

Why you group terms instead of reading rows

Only the group ever reaches the threshold. Suppose 40 terms with the same intent each cost €5, with a contribution margin of €30. Each one stays far below one contribution margin and could never be decided on its own. Together, the report shows €200, almost seven times as much. Without an order, that is a clear statement about the intent that no single row can give you.

This only works if the intent really is the same. Suppose a running shoe shop throws "running shoes wide feet" and "how to lace running shoes" into one pot because both contain "running shoes": it will then make the group look better or worse than it is. I would rather work with five to ten clusters that you can explain to someone in one sentence than with forty.

Four steps from export to clusters

First measure how much you can see. Google omits terms with too little search activity from the report. You measure the visible share with Seer's formula: cost in the search terms report ÷ cost on the Search Network, for the same period. The figure tells you how much money your judgement rests on. If it is low, compare clusters with each other, but do not extrapolate the result to the whole account.

Then count fragments. An n-gram analysis breaks the terms down into words and word pairs and adds up cost and conversions per fragment. That way, recurring patterns such as "used" or "instructions" stand out, which get lost when you read row by row.

Condense fragments into intents. A fragment is not yet an intent: "used" is one, "black" is not. Google's search terms insights help as a starting framework. The categories are created automatically based on customer intent and product attributes, and they also take into account the terms missing from the report, for Search, Performance Max and Shopping. They show clicks, conversions and value, but no cost (as of November 2026), so they tell you where demand lies. Profitability comes from your export. The catch-all "other search terms" contains terms with lower spend or without an identifiable category, and I would not invest time there.

Calculate the contribution per cluster. Orders × contribution margin minus cost. If your account sends revenue as the value, derive the contribution margin from conversion value and margin, subtract the order costs and only then the ad spend. If you send margin as the conversion value, the answer is almost directly in the table. Sort by this amount. The clusters that deserve attention are those with the largest negative and the largest positive amounts, regardless of their ROAS and their click count.

What you decide for each cluster

Unrelated intent needs no minimum data; unproven intent does. Google itself recommends negatives for search terms that are similar to your keywords but are aimed at a different product. If the shop only sells new goods, it does not need to wait and see whether "used running shoes" ever converts. "Running shoes for a half marathon", on the other hand, is unproven: the intent could fit, and the money has to show it.

The values for one and two orders follow the same logic, which I calculated using the Poisson distribution: from there on, the probability that a cluster that exactly breaks even shows so few orders is around 5%. A single order therefore does not save a cluster. At 3 ×, though, it is not yet a verdict either: that happens to a break-even cluster in around 20% of cases.

A cluster with matching intent and no orders is also not automatically a case for exclusion. I would first ask whether the ad leads to a page that serves this intent. The report for Performance Max shows the ad format and landing page for each search term. If you exclude too early, you hide a range problem or a page problem behind a tidy list.

How the negative hits the right volume

A negative keyword reaches further than the report shows. A negative broad match keyword blocks every search that contains all of its terms in any order, including searches that never appear in the report. That is your lever against the hidden part, but also the risk.

So every exclusion needs a cross-check first: filter the export for all terms that contain the fragment and add up their contribution. If it is positive, the fragment is too broad and you need a narrower combination. According to Google, negatives also do not match close variants. You add synonyms and singular and plural forms separately, while Google automatically accounts for casing and misspellings. If a misspelling still shows up in the report, exclude it as well.

For permanent exclusions, according to Google, a single global list at account level is enough, and it applies to the eligible Search and Shopping inventory of the relevant campaign types. In Performance Max, negatives for this inventory work at account or campaign level. There is more on this in lists that do not cost you revenue by accident and negative keywords in Performance Max.

Five cases against the three-times rule

The rule is my assessment based on an approximation, and it assumes a constant order rate per click. In these cases, I would not apply it mechanically:

• Long purchase cycles: if weeks pass between click and order, leave the most recent days out of the analysis period, otherwise orders that are still to come will be missing.
• Seasonality: cost from the off-season and orders from the peak season do not belong in the same calculation, and a cluster can look weak in the wrong month.
• Very expensive clicks: below about 30 clicks, the approximation is rough. Judge expensive individual terms by their intent rather than by the statistics.
• New customers and later purchases: generic terms that bring first contacts or new customers show too little value in the report. Be careful here and, if you can measure new-customer value, judge them by that value.
• Automatically expanded reach: if the volume comes from broad match, the same cluster can reappear in new terms after an exclusion. That is then a question of keyword strategy, not of the negative keyword list.