Search-term review / guides

How this review decides which search terms become negatives

By Abhishek Pal, performance marketer (apships.com). Last updated 2026-11-02.

This negative keyword generator reviews your search terms in layers. It removes rows that should not be judged, protects certain terms, checks intent patterns, then tests the remaining terms against numbers from your file. The result is a set of suggestions for your review. It cannot decide whether a search matters to your business.

The rules below describe this tool. They are not Google's rules for deciding what you should exclude. The review handles Search and Shopping terms and sets Performance Max rows aside. All examples use a fictional lamp store and made-up data.

1. Check whether the row belongs in the review

The first layer is hygiene. A row is ignored when its cost is under 1 and its clicks are under 1. Both conditions must hold. Rows already excluded in the account are skipped. Rows with fewer than 10 impressions and no clicks are treated as noise, not judged.

For example, the fictional store's search for "amber desk lamp" has 7 impressions, no clicks and no cost. It is ignored under the cost-and-click check and also fits the noise definition. It is not evidence that amber lamps attract the wrong buyers.

2. Protect terms before looking for reasons to block them

Three walls come next. Brand terms are never negated. They are kept, or sent to a watchlist for a human. Terms with a recorded conversion are never negated. Keywords already live in the account, marked "Added", are never negated either.

For example, a search for the fictional store's own name, "Willow Lantern lamps", is never negated, even if it has spend and no recorded conversions.

These walls stop a later rule from overruling a recorded result, an existing keyword choice or a brand decision. A protected term may still deserve attention. Protection does not prove that its cost or business value is acceptable.

3. Look for intent patterns in a fixed order

After those checks, the engine looks for these patterns, in this order:

  1. DIY and free-content words, including "how to", "tutorial", "pdf" and "download".
  2. Wholesale and business-to-business words.
  3. Rental, repair and training words.
  4. Jobs and hiring words, in campaigns outside your brand campaign.
  5. Second-hand words, in campaigns outside your brand campaign.
  6. Competitor names and "free" searches.

Competitor-name and "free" suggestions use exact match. The other pattern suggestions use phrase match. The tool produces only exact or phrase negatives. The negative match types guide explains how to review that choice.

For example, "lamp repair tutorial" contains a free-content pattern and a repair pattern. The free-content layer comes first, so that is the earlier rule in the sequence.

A pattern is a reason to inspect the search. A shop that sells repair kits may want tutorial readers. A wholesaler may want wholesale searches. I would check the offer before accepting either suggestion. The industry guide gives more context for that decision.

4. Require evidence when no pattern applies

The numbers layer runs only when no pattern applies. For a term with zero conversions, one route checks whether its clicks would have predicted at least three conversions at its campaign group's conversion rate. Its spend must also pass an evidence gate based on the cost per thousand impressions of its campaign group in your own export.

There is another route: at least 100 clicks, zero conversions and at least 10 units of spend. The spend unit follows the file's currency. It is not a universal $10 threshold for every account.

A very low click-through rate can also qualify: the term must have at least 300 impressions, and its click-through rate must be below 30 percent of the group's median. The median is the middle value. This compares the term with its group instead of applying one click-through rate to every campaign.

For example, "adjustable reading light" has 60 clicks and zero conversions in a group converting at 5 percent. That predicts three conversions: 60 multiplied by 0.05. This route still requires the spend gate to pass. Reaching three expected conversions alone is insufficient.

There is also an account-health check. If a campaign group has at least 100 clicks but converts below 0.2 percent, its statistical negative decisions are held on the watchlist with an account-health warning. When the group barely converts, I would investigate the wider problem before treating individual searches as the explanation.

5. Build the benchmarks from the file

The comparison numbers come from the export: group, campaign and account medians and pooled rates. A pooled rate combines the underlying totals before dividing. There are no industry conversion rates in this review.

For example, suppose two rows in the fictional lamp store's group have 2 conversions from 100 clicks and 8 from 100. Their pooled conversion rate is 10 divided by 200, or 5 percent. That number describes these made-up rows. It says nothing about what lamp stores normally achieve.

This makes the export part of the decision. I would keep its date range and campaign selection with the review, so a later reader can tell which data supported the suggestions.

Use the watchlist to keep human decisions visible

The watchlist holds terms that need attention without turning them into negatives. Brand decisions and the account-health hold belong there. I would record what needs checking, who can answer it and when to return. A watchlist entry is not an instruction to block a term later.

What the engine cannot know

The file cannot explain your margins, catalogue changes or reasons for buying a particular kind of traffic. Recent zero-conversion rows also need care: Google explains that conversions can arrive after the click, making recent performance look worse while spend is already recorded. Check your conversion delay before judging those rows.

The engine also cannot recover hidden searches. Google omits some low-volume queries from the term-level report for privacy. An export review therefore covers the visible rows, not every search. See Google's search terms report documentation.

Use the free browser check for a summary and the three most expensive negatives, with nothing uploaded. Use the full review for an upload-based free summary or the $39 report with every suggested negative and reason, the full watchlist, a paste-ready bulk-upload file and a row-level CSV. Savings are estimates. Review every suggestion before applying it.

Questions

Does every search with zero conversions become a negative?
No. The review checks row quality and protected terms first. It then checks intent patterns. Statistical rules apply only when no pattern applies, and they require evidence. Every suggestion still needs your review.
Can the review suggest excluding my brand?
Brand terms are never negated by the engine. They are kept, or sent to a watchlist for a human decision. Terms with recorded conversions and keywords marked Added are also protected from automatic negative suggestions.
Does the engine use industry conversion rates?
No. It calculates group, campaign and account medians and pooled rates from the uploaded file. Its evidence comes from that export, without outside industry benchmarks.
Does a language model read my search terms?
No language model reads the data. The review uses the stated rules, requires no account access or login, and deletes the uploaded file after the run with deletion checked.

Sources

Have an export ready? The search-term review finds the wasted spend in it for free. The full report is $39. Run it on your file.

Guides