For years the search terms report had one job you could rely on completely: it told you the actual words a person typed before your ad showed. That literal fidelity is what made negative keywords work — you read the exact string, decided it was waste, and blocked that string. In May 2026 Google quietly changed the contract. Its documentation now says that terms sourced from AI Mode, AI Overviews, Google Lens and autocomplete may appear as an AI-interpreted version of the query rather than the words the user actually typed. A row you read as a verbatim search string may now be Google's summary of what it thinks the person meant.
This is a different problem from the one the hidden wasted spend post covers. That piece is about omission — the low-volume terms Google withholds entirely. This is about distortion: the terms you can see may no longer be the exact text that triggered your ad. Both push you toward the same conclusion, but for different reasons, and if you keep writing negatives the old way you will block paraphrases that never run again while the real queries keep spending. Here is how to adjust the workflow.
What actually changed in the report
The change is documentation, not a visible UI banner, which is why most accounts have not noticed it. Google clarified that for queries arriving through its AI surfaces, the string in the search terms report can be an interpreted rendering — a normalised, sometimes reworded version of the user's input — rather than the raw text. The stated reason is that those surfaces do not always produce a single clean query the way a search box does; AI Mode and Lens in particular take conversational or visual input and model an intent from it. What lands in your report is Google's read of that intent.
The practical consequence is that the report has shifted from a record of queries toward a summary of intent, and it did so without changing shape. The columns look identical, the rows still read like search strings, and nothing flags which rows are literal and which are interpreted. That is the trap: the report looks exactly as trustworthy as it did in 2024, but a slice of it is now modeled. Practical Ecommerce documented the parallel trend of Google decreasing search-terms visibility over several years; the interpreted-query change is the same drift taken one step further — from hiding terms to paraphrasing them.
Why this breaks string-level negatives
A negative keyword blocks the text that triggers the ad, not the text printed in your report. When those two are the same — as they always used to be — reading a wasteful row and adding it as an exact negative works perfectly. When the report shows an interpreted paraphrase, the exact string that actually triggered the ad may differ word for word, and an exact-match negative on the paraphrase blocks nothing because no user ever typed the paraphrase. You have added a negative that feels like action and changes nothing.
This failure is silent, which makes it worse than an obvious error. You block the interpreted row, the row stops appearing because it was a one-off rendering anyway, and you conclude the negative worked. Meanwhile the underlying queries keep triggering under slightly different interpreted labels, and your spend on the theme does not move. If you have ever added a batch of exact negatives and watched the wasted spend on a theme barely budge, interpreted terms are one reason. The fix is not to distrust the report — it is to stop treating individual rows as exact strings you can surgically block.
Move from strings to themes
The durable response is to write negatives against the wasteful concept, not the specific row. Phrase and broad-match negatives block every query that contains the term in any wording, so a negative on free, jobs, salary or a competitor's name catches the verbatim queries and the interpreted paraphrases alike. Thematic negatives were already the only tool that reached the withheld low-volume layer; interpreted terms make them the only tool that reliably reaches the visible layer too. The theme is stable even when the exact string Google prints is not.
Concretely, this means grouping the report by the word or intent that makes a row wasteful rather than scanning for individual strings to exact-negate. Run the same n-gram pass the wasted-spend work already relies on: split the query column into words, total cost and conversions by word, and target the high-cost zero-conversion words with phrase negatives. Because you are matching on the concept, it does not matter whether a given row was literal or modeled — the negative reaches both. This is the same discipline the Performance Max negatives workflow uses, applied now because the report itself has become less literal.
Verify the theme against data the report cannot fake
Because a row may be interpreted, confirm a wasteful theme against signals that do not depend on the report's wording before you block at scale. Conversion data is the strongest: a theme with real spend and zero conversions over a window longer than your conversion lag is provable waste regardless of how each row is labelled. Landing-page and analytics behaviour — bounce, time on page, whether the traffic ever reaches a conversion step — corroborates it. So does first-party data, if the clicks that do convert on a theme turn into low-quality leads in your CRM.
The reason this matters more now is that you have lost the ability to sanity-check a negative by re-reading the exact query. Previously, if a negative looked risky you could eyeball the literal strings it would block. With interpreted terms you cannot fully trust that preview, so the check moves downstream to outcomes. Treat the search terms report as the place you form a hypothesis about which themes waste money, and treat conversions and first-party signals as the place you confirm it. That division of labour is the whole adjustment: the report proposes, your outcome data disposes.
A negative-keyword routine built for a modeled report
Rebuild the routine around themes and verification rather than string-by-string blocking. First, on your normal cadence — weekly for heavy spenders, monthly for smaller accounts — pull the report and run the n-gram rollup to surface high-cost zero-conversion words. Second, for each candidate theme, confirm it against conversions and landing-page behaviour rather than trusting the row text alone. Third, add phrase or broad thematic negatives on the confirmed words, placed at the right level and checked against your active keywords so you do not block something you are bidding on.
Keep the written log the rest of your negatives use — what you excluded, at what level, and why, with a date — because thematic negatives touch a lot of traffic and interpreted reporting makes their effect harder to eyeball after the fact. If volume drops on a campaign, the log is the only reliable way to find and reverse the broad negative that went too far, since you can no longer reconstruct the exact blocked strings from the report. The account that adapts to interpreted terms is not the one that stops using the search terms report; it is the one that reads it as a map of intent, writes negatives against themes, and proves the waste with data the model cannot reword. For the fundamentals this builds on, start with the search terms report walkthrough.