An irrelevant-looking search query is not automatically a negative keyword

Reviewing search terms is one of the most useful habits in paid search. It shows what people actually typed before they saw an ad, and it is where wasted spend becomes visible. It is also where an account can lose valuable demand without anyone noticing, because a negative keyword that blocks good traffic produces no error. The blocked searches simply stop appearing in the report.

That asymmetry matters. A missed negative usually costs a little each day and stays visible, so it can be corrected later. A wrong negative removes the evidence that would show it was wrong. When in doubt, the safer default is to keep the query and keep watching it.

Two judgements tend to get merged. One is relevance: does this query mean what the business sells? The other is performance: has the query, or the theme it belongs to, spent enough to matter, and what has it produced? They need different evidence, and a query should be excluded only when both point the same way.

• Irrelevant and wasting money: a strong candidate for a negative keyword, added at the narrowest match type and scope that solves the problem.
• Irrelevant but converting: a sign that the relevance judgement is wrong, or that customers describe the product differently from the business. Investigate before excluding.
• Relevant but wasting money: not a negative keyword problem. Look at the landing page, the ad, the bid or the offer instead.
• Too little data to judge: leave it alone, and let match types and account structure manage the long tail.

Some queries look odd to someone who knows the product well and still come from buyers. Customers use trade slang, old model names, a competitor's product name as a generic term, or a description of the problem rather than the solution. A query can be informational today and commercial next month; excluding every search that starts with how or what can cut off people at the beginning of a considered purchase.

The reverse happens too. A query can contain exactly the right words and still come from the wrong buyer: someone looking for a job in the category, or for a do-it-yourself version of a service the business only delivers in full. Reading the words alone will not separate these cases. The conversion data, the landing page and the business's own knowledge of its customers all help.

Negative keywords do not behave like the keywords you target. Google's documentation says negative keywords do not match close variants or other expansions, so excluding one word will not exclude its plural or its synonyms; you need to add those yourself, although casing and misspellings are accounted for automatically. A negative broad match keyword blocks a search only when the search contains all of its terms, in any order.

Scope matters as much as match type. An account-level negative keyword list applies across all eligible search and shopping inventory, which suits themes that are never wanted and is risky for anything that might be relevant to one product line. Campaign-level negatives are better for keeping campaigns from competing with each other.

Coverage is also incomplete. Google omits some low-volume search terms from the search terms report to protect privacy. Decisions are made on the visible share of searches, so it is worth knowing how much of your spend that visible share represents.

• Sort the report by cost, so the review starts with the spend that matters.
• Set a materiality threshold that suits your account, and do not judge queries below it one by one.
• For each candidate, record the relevance judgement and the performance evidence separately.
• Look at conversions over a long enough window before calling a query a waste; a handful of clicks cannot support a conclusion either way.
• Choose the narrowest match type and scope that solves the problem, and prefer exact match negatives for single unwanted queries.
• Protect brand and product range terms explicitly, so that they cannot be excluded by accident.
• Record why each negative was added, and review old exclusions periodically, especially after the product range or the offer changes.

A hypothetical example

This example is hypothetical. An online furniture retailer sees the query second hand oak table in a campaign for new tables. It looks irrelevant, because the retailer does not sell used furniture. Over the review window, though, it has spent a modest amount and produced several sales. On inspection, the landing page features a clearance range. The query is not a negative keyword; it is a signal that some buyers are price-led, and that the clearance range deserves its own ads.

A second query, oak table repair, has spent more with no sales, and repairs are not something the retailer offers. That one is a sound exclusion, added as an exact or phrase match negative rather than a broad exclusion of everything containing oak table.

Language models can classify large volumes of queries by likely intent, which saves a great deal of reading. They should be grounded in the business's own product and service definitions, their uncertain verdicts should go to a person, and their judgements should be checked against the conversion data rather than replacing it. A confident classification is still a hypothesis until the evidence agrees.