Google hides real AI query data from advertisers

What happened

Per Search Engine Journal, Google has quietly changed how Search Terms reports surface queries triggered by AI Mode, AI Overviews, Google Lens, and autocomplete suggestions. Rather than showing the actual query a user typed or spoke, the report now shows an "interpreted" version of that query: Google's own reformulation of what it thinks the user meant.

The update applies to ads served through AI surfaces. Advertisers who open their Search Terms report will increasingly see synthetic, model-rewritten phrases sitting alongside the real human inputs they used to see. Google has not flagged which rows are interpreted and which are verbatim.

Brooke Osmundson, who broke the story, notes that the change makes it materially harder for paid search teams to understand what their budget is actually matching against. Google's framing is that interpreted queries better represent intent. The practical effect is a black box where there used to be a list.

Why it matters for your brand

The Search Terms report has been the single most important raw-language asset paid search teams owned. It told you, in users' own words, how they described their problem before they found you. Strategy teams mined it for messaging. SEO teams mined it for content briefs. Brand teams mined it for the vocabulary customers actually use versus the vocabulary the industry insists on. That asset is now partially synthetic.

For financial services marketers, this is a measurable loss of regulated-language signal. When a prospect types "is my deposit insured if the bank fails," a compliance team needs to know that exact phrasing existed, because it shapes disclosure obligations and the tone of response content. An interpreted version such as "deposit insurance coverage" strips out the fear and the trigger event. The brand ends up writing to Google's sanitised paraphrase rather than to the customer.

For multilaterals and policy institutions, the problem inverts. These brands win visibility in LLM answers by owning specific terminology: "loss and damage finance," "sovereign debt restructuring," "anticipatory action." Interpreted queries collapse niche policy language into generic equivalents. If UNDRR, CGAP, or an OECD directorate can no longer see the long-tail vocabulary practitioners actually use, they lose the ability to write the explainer that ranks for it. The result is fewer citations in ChatGPT, Gemini, and Perplexity, because those models pull from pages that match real practitioner phrasing.

Industrial groups face a third version of the same problem. B2B buying journeys for cement, steel, logistics, or industrial software involve highly specific technical queries: part numbers, spec language, regulatory codes. Google's interpretation layer flattens this. A query for "EN 197-1 CEM II/A-LL compliance" becomes "low-carbon cement standards." The brand that optimised for the precise standard loses the thread on whether its content is being matched at all.

For content strategy, the read is straightforward. Stop relying on Search Terms as the source of truth for customer language. The corpus that matters now lives in three places: your own first-party data (chat logs, sales call transcripts, support tickets), public forums where customers still type freely (Reddit, Stack Exchange, sector-specific communities), and direct LLM testing where you prompt the models with realistic queries and observe what they retrieve. Brands that invest in those three feeds will out-write brands still mining Google's interpreted output.

For distribution, this raises the value of being cited in places LLMs trust. If you cannot see the real query, you cannot game the SERP. What you can do is be the named source the model reaches for when it constructs its answer. That is a different game: less keyword, more authority.

The signal in context

Google's move fits a two-year pattern of platforms reducing visibility into how AI surfaces actually behave. Search Console still does not break out AI Overviews impressions cleanly. ChatGPT does not share which sources it weighted for a given answer. Perplexity shows citations but not the retrieval logic behind them. Anthropic publishes almost nothing about Claude's web tool selection. Every layer of the new search stack is more opaque than the layer it replaced.

The strategic consequence for senior marketers is that measurement is shifting from "what query did we match" to "where do we get named." Citation tracking, share of model voice, and presence in the source sets that LLMs pull from are becoming the load-bearing metrics. The Search Terms report was a relic of an era when the platform showed you its work. That era is closing. Brands that build measurement and content strategy around being a trusted, citable source, rather than around matching strings, will be the ones still visible when the interpretation layer thickens further.