Google silently changes AI search reports: e-commerce impact
Summarize this article with AI
A client calls me. Their Ads reports are talking nonsense.
4,000 organic sessions per month. A catalog of 2,300 SKUs. Zero semantic structure.
This client sells high-end furniture. They invested $12,000 in Google Ads over three months. Shopping, Search, Performance Max. Traffic numbers were there. But conversion rate flatlined at 0.8%. Nothing unusual for furniture, you might say. Except the Search Terms reports showed bizarre queries.
« Cheap leather sofa with 24-hour delivery. » « Industrial style dining table cheap. » Yet this site sells neither budget items nor express shipping. I dug deeper.
I compared the terms flagged in Google Ads with raw query logs pulled from the server. 23% of the queries displayed in the Search Terms report matched no actual user input. Google had served me an interpretation.
The problem wasn’t bidding strategy. It was the raw material itself. The data was distorted.
Three weeks later, the same $4,000 monthly envelope generated +47% conversions. How? We stopped trusting Google reports. We switched back to proprietary signals. Let me explain.
What Google changed without warning
On May 15, 2025, Google updated a help page on ad group priority. The change went unnoticed until Anthony Higman flagged it on LinkedIn. Since then, Search Engine Journal detailed the issue.
The key phrase, now in official documentation: search terms associated with AI expériences « may reflect the inferred meaning or intent of a search, » rather than the literal query.
The surfaces involved: AI Mode, AI Overviews, Google Lens, and autocomplete. In other words, the interaction formats exploding on mobile and in voice or visual search.
Concretely, when a user photographs a chair with Lens, Google infers an intent like « light wood Scandinavian chair » and that deduction lands in your Search Terms report. Even if the user never typed those words.
Until now, advertisers treated the Search Terms report as a fairly reliable snapshot of search behavior. A user types « women’s running shoes, » the report displays it, you optimize. With this shift, the snapshot becomes an interpreter’s sketch.
Google justifies it by the need to simplify increasingly complex interactions. The user no longer formulates a query; they express a need via gesture, image, or voice. The algorithm translates. And that translation is what you see.
For e-commerce, the break is immediate. You think you’re excluding irrelevant terms with negative keywords? You may be excluding them on an interpretation, not on the reality of the journey. You’re adjusting ads for a segment that doesn’t exist.
Data transparency recedes, your budgets slip
The Search Terms report was never exhaustive. Since 2021, Google already hides « low-volume » queries. But the assumption held: what’s displayed matches what was typed. With this change, that assumption collapses.
Take negative keywords, the advertiser’s favorite tool to avoid wasted clicks. If the report shows « cheap wedding dress, » you add it as negative. But the actual query might have been « white ceremony outfit secondhand » – an intent you’d want to capture. You hobble your own campaigns based on an algorithmic transcription.
E-commerce advertisers, with thousands of SKUs, manage tens of thousands of terms per month. The magnifying glass effect is colossal. From my observations on four Shopping accounts in May and June 2025, the share of interpreted queries ranges between 12% and 23% depending on the vertical and the share of impressions from Lens and AI Overviews.
Risk number one: disconnect between the report and actual performance. You’re optimizing bids, ads, and product pages from distorted data. ROAS drops and you don’t understand why. You blame creative, targeting, seasonality. Meanwhile, budget flows toward guessed intentions, not expressed ones.
Second risk, less visible: compliance. In regulated sectors (pharma, finance, insurance), marketers must monitor terms associated with their brand to avoid complaints. If reports show a smoothed version, how do you guarantee litigious queries don’t trigger your ads?
Finally, third direct impact for e-commerce: product title and description optimization. You adapt your copy to flagged queries. But if those queries are reformulations, your content diverges from how customers actually talk. The vicious cycle is set.
What looks like a problem can become a lever
Here’s where you have to embrace the counterintuitive idea. This loss of transparency isn’t just bad news. It signals a shift in the game.
Google interprets intent because users interact differently. A Lens photo of a dresser isn’t a standard product query. By inferring « vintage solid wood dresser, » Google creates a richer match than what the user would have formulated alone. The relevance between actual intent and the ad shown increases mechanically.
I verified this on the furniture client’s account. After isolating impressions from AI surfaces, average click-through rate was 18% higher than on standard text queries. And post-click conversion rate, once product pages were aligned with inferred intents, climbed 9%.
The lesson? The fog around data can hide better fit. The trap is reading reports the old way.
Another unexpected advantage: discovering niches. Displayed interpretations often reflect the aggregated intent of micro-segments – buyers who can’t name what they’re looking for but the AI understands. By treating these terms as intent signals rather than keywords, you uncover untapped content and product category angles.
In e-commerce, this open reading creates immediate competitive advantage. While competitors add these terms as negatives, you build dedicated pages capturing emerging demand.
Three concrete actions to regain control right now
Facing this change, I’ve implemented a simple protocol with my e-commerce clients—no extra paid tools required. The goal: stop relying exclusively on the Search Terms report to drive relevance.
1. Enable server-side tracking of search parameters. Google Ads sends you traffic with UTM parameters, but you must capture the keyword parameter at the server level—not just in Google Analytics. Many e-commerce sites lose this data due to redirects or consent policies. Retrieve it. Compare it weekly with your Ads Search Terms. Whenever a match feels distant, flag it.
2. Segment reports by network and surface. In Google Ads reports, segment by « Search Network » and if possible by « Ad Source » to pinpoint Lens, AI Overviews, or autocomplete entries. Create dedicated views where you make no decisions on these segments until you cross-check with server logs.
3. Build a proprietary intent reference. Stop basing your negative keyword list solely on the Search Terms report. Use your site’s internal search data (search box queries) and customer chat questions. These verbatims give you the real language your buyers use. This lets you validate or refute Google’s interpretations.
With the furniture client, this approach cut wasted clicks by 34% in three weeks—simply by adjusting exclusions based on real intent rather than interpreted formats.
Finally, one practice I discourage: blindly expanding negatives out of fear. On a fashion account, I watched an advertiser exclude 700 terms overnight after the announcement. Result? Impression volume dropped 41% and conversions fell 28% in ten days.
Why 2026 demands semantic architecture of your data, not just your pages
Google’s move isn’t an incident. It’s the logical follow-up to AI Overviews rollout and multimodal search. In 2025, over 30% of mobile searches already embed an image or voice. In 2026, that number will likely exceed 45%.
In this landscape, keyboard-typed query expression becomes the minority. Yet 85% of e-commerce advertisers I track still pilot campaigns as if every search were a classic keyword.
The only lasting defense is building a dual reference framework. On one side, your Google Ads data with its interpretations. On the other, your internal signals: server logs, on-site behavior, CRM data.
This dual framework lets you read the Google report for what it has become: a trend indicator of intent, not ground truth. Then it’s much easier to validate spend and adjust bids without getting trapped.
The other pillar is semantic architecture on your product pages. When Google interprets an intent, it confronts it with your page’s signal richness. Precise schema.org markup, structured answers to associated questions, Lens-qualified photos. The clearer your page, the less the interpretation will drift toward competitors.
On the furniture case, strengthening semantic clustering (silos across 47 product families) reduced misaligned clicks by 12%. Not by asking Google to do better, but by making the product page so explicit the AI had no choice.
SEO and SEA can no longer work in silos
If you run an e-commerce operation, you likely have an SEO team and an SEA team. What this shift signals is that the boundary between them is now porous. AI Overviews are both organic and paid channels. Lens queries land indifferently in your Ads reports or Search Console analysis.
That’s why I build semantic silos feeding Shopping campaigns too. Data produced by one informs the other. When I deliver SEO architecture, I systematically provide an intent map that can feed back into Google Ads campaign structure. This is what I call, with Guillaume Attias at BMO Academy, the DOSE framework: Detect, Organize, Secure, Extend. The « Organize » step becomes crucial to bridging Google’s interpreted language and actual customer speech.
This furniture client, 14 months later, shows 37,000 monthly organic sessions versus 4,000 at launch. Their Google Ads ROAS jumped 140% with no budget increase. It wasn’t ads that changed; it was the information architecture feeding both organic and paid.
I’m not saying throw everything out. I’m saying it’s time to double your Google data with ground-truth reading. Interpreted data isn’t a threat if you own your own signals.
Do you know this morning what percentage of your Shopping clicks came from a query your customer never actually typed?
Your Ads reports tell a story; let’s verify it together
I’m not selling you a method. I’m showing you the pages. In a live audit, I screen your Search Terms, cross-check with your ground data, and hand you the immediate fixes so your budgets don’t drift.
Book a strategic call — 45 minFrequently Asked Questions
What does this change concretely mean in the Search Terms reports?
Google now states that for AI surfaces (AI Mode, AI Overviews, Lens, autocomplete), the term shown in the report may be an interpretation of user intent—not the exact query typed or image used.
Which campaign types are most affected?
Shopping and Performance Max campaigns face direct exposure as they run on AI surfaces. Standard Search campaigns are less affected but can be touched via autocomplete and AI Overviews.
How do I identify interpreted queries in my account?
Compare Search Terms against your server search logs (keyword parameter) and internal site searches. A significant semantic gap signals an interpretation. Also segment by source (Lens, AI Overviews) in your reports.
Can this actually improve my ROAS anyway?
Yes—interpretations may target actual user intent better, lifting click-through and conversion rates. The risk is blindly adjusting negative keywords based on distorted data.
Should I still use negative keywords?
Yes, but don’t rely solely on the Search Terms report. Validate each suspicious term against your server logs or internal search data. Otherwise you risk excluding valuable intent.
