Does the Google Ads search terms report still show the exact query a user typed?

Not always. Google has updated its Ads Help documentation to say that for searches arriving through Google Lens, AI Mode, AI Overviews and autocomplete, the term shown in the search terms report is the best approximation of the user's intent rather than the literal words typed. For a plain text search on google.com the displayed term is still generally the query, but you can no longer assume every row is a verbatim transcript — a growing share are Google's interpretation of what the searcher meant. The practical consequence is that an exact-match negative keyed to a specific string may never fire, because that exact string was never what the system recorded.

Why does this break literal-term negative keyword auditing?

Literal-term auditing assumes the term in the report is the thing you are blocking. If the report shows an approximation of intent, a negative keyword added to match that displayed term can miss the real underlying queries that produced it, and conversely a query you think you have blocked can keep spending under a differently-worded interpretation. Google has also said that identical keywords may not automatically receive priority — AI-based ad group prioritisation picks the group that best fits the inferred intent. So both halves of the old model weaken at once: the term you read may not be real, and the keyword you match may not win the auction the way the match-type label implies.

Which searches are affected by the approximate-intent change?

Google scopes the change to searches that come through its AI and assisted surfaces: Google Lens, AI Mode, AI Overviews and autocomplete. For those journeys the report shows an interpreted intent rather than the raw text. Conventional typed searches on the results page are less affected, but there is no column in the report that tells you which rows are literal and which are interpreted, so in practice you have to treat the whole report as partly approximate and audit accordingly. The share of interpreted rows will grow as more search volume moves through AI surfaces, so the safe assumption going forward is that any given term might be an approximation.

How should I change my negative-keyword workflow in response?

Shift weight away from single exact-match negatives tied to a displayed string and toward patterns and performance. Build negatives from recurring modifier words and n-grams that appear across many rows, because a pattern survives the report's imprecision in a way a single literal term does not. Lean harder on the metrics you can still trust — cost, clicks and conversions per term — to decide what to cut, rather than on the exact wording. And route genuinely ambiguous single tokens to a review queue instead of negating on a guess, since a misread of an approximate term is now easier to make.

Can I still trust conversion data in the search terms report?

Yes. The approximation applies to the search term text, not to the performance columns. Cost, clicks, impressions and conversions attributed to a row are still real spend and real outcomes, even when the term label is Google's interpretation of intent. That is precisely why the response to this change is to anchor decisions on the metrics: a row that spent money and produced nothing is a problem whether or not its term is literal, and an n-gram that accumulates cost with no conversions across dozens of approximate rows is a reliable signal even when no single row's wording can be taken at face value.

Search Terms Report Now Shows Intent, Not Exact Queries

For as long as the search terms report has existed, practitioners have read it as a transcript: the list of actual strings people typed that triggered your ads, which you scan for waste, mine for negatives, and harvest for new keywords. Google has quietly changed what the report is. Its Ads Help documentation now states that for searches arriving through Lens, AI Mode, AI Overviews and autocomplete, the term shown is the "best approximation" of the user's intent rather than the literal query. The report is no longer a record of what was typed; for a growing share of rows it is Google's interpretation of what the searcher meant.

That is a change to the epistemology of the single most important report in search-term hygiene, not a one-off matching quirk. If the term you read might be an inference, then the oldest reflex in the craft — copy the offending string, paste it as an exact-match negative, move on — can quietly stop working. This post is about what actually shifted, why it undermines literal-term auditing, and how to re-base your negative-keyword process on the parts of the report that are still solid ground.

What Google actually changed

The change is documentation, not a dashboard button, which is why it is easy to miss. Google updated its help text to say the search terms report may show an approximation of intent instead of the exact words typed, and scoped it to its AI and assisted surfaces: Lens, AI Mode, AI Overviews and autocomplete. As Search Engine Land and others reported around May 2026, for those journeys the system records what it believes the person wanted, not the characters they entered. There is no flag in the report separating interpreted rows from literal ones, so you cannot audit around the change by filtering it out.

The second half of the change is about matching, and it is the one that stings more. Google has said that for these searches, identical keywords may not automatically receive priority; instead an AI-based prioritisation chooses the ad group that best fits the inferred intent. One practitioner analysis cited in the coverage found an exact-match keyword matching queries that did not contain its words at all, which tells you the match-type label in the report is Google's own classification of intent rather than a promise about the literal string. The report has become descriptive of intent on both axes: the term it shows, and the keyword it credits.

Why this breaks literal-term auditing

Literal-term auditing rests on a single assumption: the term in the report is the thing you block when you negate it. Remove that assumption and the method develops two blind spots at once. First, a negative you add to match a displayed term may never fire, because the real queries behind that interpreted row are worded differently — the same failure mode as a negative that misses a paraphrased search term, now generalised across every AI-surface row rather than confined to the odd close variant. Second, a query you believe you have blocked can keep spending under a different interpretation, so your list looks complete while the waste continues.

The matching change compounds this. Because identical keywords no longer reliably win their own queries, the tidy mental model — exact keyword catches exact query, negative blocks exact string — stops predicting what happens in the auction. This is the same drift that has made exact match show unrelated search terms, taken a step further: now even the terms you are reading to diagnose the problem are themselves approximations. Auditing one literal string at a time was always slow; the new problem is that it is also unreliable, because the string is not guaranteed to be real.

What in the report is still reliable

The approximation applies to the term text, not to the numbers attached to it. Cost, clicks, impressions and conversions on a row are still actual spend and actual outcomes, regardless of whether the displayed term is literal or interpreted. That distinction is the whole basis of the response: you stop trusting the label and keep trusting the ledger. A row that took real money and returned nothing is a problem whether or not its wording is verbatim, and that judgement needs no faith in the string at all.

Patterns across rows are the other thing that survives. Any single interpreted term might be wrong, but a modifier or token that recurs across dozens of rows — and accumulates cost without conversions as it goes — is a signal the approximation cannot wash out, because the imprecision is in individual wordings, not in the aggregate shape of where money goes. This is why the report remains genuinely useful even as it becomes less literal: it still tells you, accurately, which clusters of intent are spending and which are converting. You are simply reading it at the level of the cluster now, not the line.

Re-basing negatives on patterns and n-grams

The concrete shift is to move weight away from single exact-match negatives keyed to a displayed string and toward negatives built from recurring words. Run an n-gram analysis over the report: break every term into its component words and short phrases, sum cost and conversions by n-gram, and you surface the intent words that are bleeding budget across many rows at once. An n-gram that spends across forty approximate terms and converts on none is a safe negative even though you would not trust any one of those forty terms on its own — the pattern is real precisely because it is an aggregate.

Negating the modifier rather than the exact term also insulates you from the wording problem directly. If "free" is the thing that never converts, a phrase-match negative on "free" blocks the whole family whether the report rendered it as "free tool," "tool for free" or some interpreted paraphrase in between. That is the same logic as grouping by pattern instead of negating single terms, and it matters more now than it did when the report was literal: a pattern negative does not care which of several approximate surface forms Google chose to display, because it matches the underlying word in all of them.

Letting performance, not wording, drive the cut

When the term is uncertain, the metric is the arbiter. Sort the report by cost, and for each high-spend cluster ask what the performance says rather than what the wording implies: did this spend produce conversions, assisted value, or anything at all over a fair window? A cluster that spends materially and returns nothing gets cut on the strength of its numbers, and you do not need to resolve exactly what each approximate term "really" was to make that call. The report's new imprecision is a reason to trust the columns you can verify over the label you cannot.

Ambiguity that cannot be resolved by the numbers should go to a review queue, not to a negative. A single interpreted token — one word, high spend, unclear intent — is now easier to misread than it was, so negating it on a guess risks throttling good traffic you could not see. Park it, gather more data, and decide when the performance picture is clear. The discipline is the same as it always was; the change is that the bar for "I understand this term well enough to block it" is higher when the term itself is an approximation.

Why this is not the same as hidden search terms

It is worth separating this change from the privacy thresholding that has long kept some terms out of the report entirely. Hidden terms are rows you never see; approximate terms are rows you do see but should not read literally. They are different gaps that call for different responses: the hidden and redacted terms problem is about reconciling spend that has no visible term at all, while the approximation problem is about not over-trusting the terms that are visible. Conflating them leads to the wrong fix — you cannot pattern-mine your way to a term that was never shown, and you cannot threshold-audit your way past a term that is shown but interpreted.

Taken together, the two changes point the same direction: the search terms report was never going to reconcile perfectly with your campaign totals, and it is now also not a verbatim record of what remains. The response is not to abandon it — it is still the best view you have into where search spend goes — but to use it as a map of intent and spend rather than a transcript of queries. Anchor on the numbers, act on patterns, and treat any single displayed term as a hypothesis the performance data either supports or does not. For the broader workflow this sits inside, start from the guide to reading the search terms report and layer this approximation caveat on top of it.

For the foundational walkthrough of the report this change modifies, see the Google Ads search terms report guide. For the pattern-based negative method that holds up when individual terms cannot be trusted, see n-gram search-term analysis, and for choosing patterns over single strings deliberately, see grouping patterns versus negating single terms.