A negative keyword you added last week is not blocking the traffic you built it for — and the search terms report is telling you it worked. This is the quiet failure mode that arrived when Google changed what the report shows. As the trade press reported, Google’s search terms report now shows AI intent, not what users typed for a growing share of queries. That sounds like a reporting cosmetic. It is not. It breaks a hidden assumption underneath every negative keyword workflow: that the text in the report is the text your negative will match against. It is not, and it never was — the gap simply did not matter until the report started showing you paraphrase instead of the raw query.
The consequence is a negative that fires on a description of the search rather than the search itself. You read “affordable running shoes” in the report, decide you do not want bargain-hunters, and add “affordable” as a negative. But the user typed “cheap trainers under 20”; the word “affordable” was Google’s interpretation, never in the query. Your negative matches nothing. The row may even vanish from next week’s report — re-paraphrased or re-bucketed — which reads like success while the spend continues untouched. This post is about spotting that trap and building negatives that survive it.
What changed: the report shows intent, not the query
The search terms report used to be, functionally, a transcript: each row was a real string a user typed, matched against your keywords, with its own clicks and cost. Google has been steadily moving it away from that. First came aggressive aggregation and the withholding of low-volume terms; now, for a growing set of queries, the report shows an AI-interpreted rendering of what Google believes the searcher meantrather than the exact words they entered. Search Engine Journal documented the shift, noting Google “quietly changed how search terms are reported for some AI queries”, and Adalysis has catalogued the reporting oddities this produces inside AI Max.
For most reporting purposes this is a nuisance you can live with — a slightly fuzzier picture of demand. For negative keyword hygiene it is a structural problem, because the entire workflow is built on copying real strings out of the report and into a negative list. When the report stops being a transcript and becomes a paraphrase, the copy-paste step silently corrupts: you are transcribing Google’s summary of the query, then asking the match engine to block traffic on the basis of that summary. The two are related but not identical, and the difference is exactly the set of queries that slip through. It sits alongside the older visibility problem — terms Google hides entirely — covered in the hidden and redacted terms breakdown; paraphrasing is the same erosion of trust in the report by a different mechanism.
Why this breaks negatives: they match the query, not the label
Negative keywords are evaluated against the actual query string the user typed, under the ordinary match-type rules. A broad-match negative blocks a query when the query contains every word of the negative in any order; a phrase negative needs the words in sequence; an exact negative needs the whole query to match. None of these rules ever looked at the search terms report — the report is a downstream display, and the matching happens live at auction time against the real string. So a negative and the report label were always two different objects that usually happened to agree. Paraphrasing is what pulls them apart.
Once they diverge, the failure is specific and predictable. If the report’s paraphrase introduces a word the user never typed, a negative built on that word matches nothing, and the real query keeps spending. If the paraphrase drops a word the user did type — the actual token you would have wanted to negate — you never see it to negate it. And because the paraphrase can change between reporting periods, the same underlying query can appear under different labels week to week, so even a negative that worked can look like it stopped working, or a row you targeted can disappear while the traffic persists. The negative did exactly what it was told; you told it about the label, not the query. The strategic side of negating these intent-rendered terms is covered in negating AI-interpreted search terms — this post is the mechanical failure underneath that strategy.
How to spot a negative that never fired
Stop trusting the disappearance of a report row as proof a negative worked, because that is precisely the signal paraphrasing corrupts. The row can vanish for three different reasons: the negative genuinely blocked the queries, the queries fell below the reporting threshold, or Google re-paraphrased them under a new label. Only the first is success, and the row disappearing looks identical in all three. The reliable signal lives one level down, in spend and impressions on the theme you meant to block, not the specific row you targeted.
The check is concrete. Before you add a negative, note the campaign’s spend on the theme — the cluster of queries around the intent you are trying to exclude, however the report labels them. Add the negative, then over the following days watch whether that thematic spend actually falls. If total spend on the theme holds steady while the specific row you targeted vanished, your negative matched the paraphrase and missed the traffic, which has simply resurfaced under a different label or slid into the anonymous “other search terms” aggregate. Reading spend against how much of it you can even see is the subject of the visibility ratio audit, and it is the habit that keeps a vanished row from being mistaken for a solved problem.
Building negatives that match the query, not the paraphrase
The fix is to stop negating rows and start negating patterns. A single copied row is exactly the artefact paraphrasing corrupts; a pattern — a token or small n-gram that the real queries almost certainly contain regardless of how Google labels them — survives the paraphrase because it targets the underlying string. If the intent you want gone is job-seekers, the real queries overwhelmingly contain “job”, “jobs”, “career”, “salary”, “hiring”; negate those as broad-match negatives and you catch the traffic whether the report calls it “employment enquiry” or anything else. You are negating what people type, which the paraphrase cannot hide, rather than what Google says they meant, which it rewrites at will.
This is where n-gram analysis stops being a nice-to-have and becomes the primary method. Instead of reading the report row by row — where every row might be a paraphrase — you break the visible queries into their component words and phrases and look for tokens that recur across losing spend, because a token that appears in many real queries is far less likely to be an artefact of any single paraphrase. The full technique is in n-gram search-term analysis. Pair it with deliberate match-type choice on the negatives themselves — broad negatives to catch the token in any phrasing, phrase or exact only when you need to protect a legitimate adjacent query — which is worked through in negative keyword match types, plurals and close variants. Patterns plus the right negative match type are what make a negative fire on traffic the report no longer shows you verbatim.
A confirmation routine before you trust a negative
Treat every negative as unverified until spend proves otherwise. The routine is short and it closes the loop that paraphrasing opened. One: define the theme and its current spend before you touch anything, so you have a baseline that is not tied to a single report row. Two: add the negative as a broad-match pattern built on tokens the real queries contain, not on the copied paraphrase. Three: wait past your conversion lag and re-measure thematic spend, not the row. Four: if the spend did not move, the negative missed — widen the pattern, check for a token the paraphrase was hiding, and try again rather than assuming the job is done.
Building this loop into the weekly pass is what separates an account that is actually clean from one that merely looks clean in a report Google is increasingly rewriting. The paraphrasing change does not mean the search terms report is useless; it means you can no longer treat it as a transcript to copy from, and every negative needs a spend-level confirmation before you count it as done. Fold that confirmation into the standing weekly search-terms-report routine so it is a habit and not a special project. The report will keep getting fuzzier; a workflow that verifies negatives against spend rather than against report rows is what keeps your hygiene honest as it does.