Search Console and AI Overviews: how to build the before-and-after of GEO for the client

The question every client asks in 2026 is the same: "did Google traffic drop because of AI?" The official Google Search Central documentation on AI features in Search gives an uncomfortable answer for anyone who needs to prove ROI: clicks coming from AI Overviews (the AI-generated answers at the top of the SERP) and from AI Mode already enter the Performance report, but mixed with traditional search traffic, within the "Web" search type.

In other words: Google confirms these clicks exist and are counted, but it doesn't provide a native filter that says "this one came from an AI answer." Building the before-and-after of GEO (Generative Engine Optimization) for the client, therefore, requires method, not a magic button. This article proposes that method.

What Google Confirms (and What It Doesn't Deliver)

The documentation is direct on a few points worth fixing in mind before any analysis:

• There's no special optimization to appear in AI Overviews or AI Mode. Standard SEO best practices remain fully valid. There's no need to create AI-specific text files, new markup, or specific schema.org structured data.
• The technical requirement is the same as for a regular snippet: the page needs to be indexed and eligible to appear in Search with a snippet that meets the technical requirements. Nothing beyond that.
• Clicks are aggregated into overall traffic in the Performance report, under the "Web" search category. There's no "Appearance = AI Overview" dimension the way there is for results with review stars or sitelinks.
• Google states that these clicks have better quality: according to the documentation itself, users who click from a SERP with an AI Overview "tend to spend more time" on destination pages.

This last point is what gives ammunition to the GEO argument, but it isn't quantified. It's up to whoever measures it to turn "spend more time" into a defensible number.

Why You Can't Isolate 100% (and How to Get Close)

Since there's no native filter, the strategy is to triangulate signals. None of them isolate the AI Overview with surgical precision, but added together they build an honest case.

1. The pattern: impressions rise, CTR falls. When an AI answer starts occupying the top spot for a set of queries, the typical behavior is: impressions for those queries hold steady or grow, but the click-through rate plummets, because the user reads the answer without scrolling down. In the Performance report, this shows up as a divergence between the impressions curve and the clicks curve. It's the classic symptom of AI exposure without a click.

2. Segment by query and by page. Filter long-tail queries phrased as questions ("how to," "what's the best," "difference between"), which are exactly the ones the documentation describes as the territory of AI Overviews and AI Mode, built for "complex questions that require research or comparison." Compare average position, impressions, and clicks for these queries before and after the date the AI answer started appearing for them.

3. Cross-check with Analytics. The documentation explicitly recommends combining Search Console data with Analytics to track conversions and time on page. This is where the "better-quality click" argument turns into a metric: if the average time and conversion rate of organic sessions rise while click volume shrinks, you have the picture of fewer clicks, but more qualified ones, which is GEO's central thesis.

A Before-and-After Roadmap

The sequence below is the path that makes sense to follow in a typical project. Adjust the dates to the client's history.

1. Mark the cutoff date
   (when the AI Overview started appearing for the target queries,
    cross-checked with SERP monitoring)

2. In Search Console > Performance > search type "Web":
   - export 90 days before and 90 days after
   - filter queries phrased as questions / comparisons
   - record: impressions, clicks, CTR, average position

3. In Analytics, for the same period and organic channel:
   - average engagement time
   - conversion rate
   - pages per session

4. Build the comparison table and read both axes together:
   volume (SC) x quality (Analytics)

The correct reading is never a single number. Here's an example of how to present the before-and-after to the client, with illustrative values:

| Metric | Before | After | Reading | |---|---|---|---| | Impressions (target queries) | 100,000 | 128,000 | More exposure, including in AI | | Clicks | 8,000 | 6,400 | Volume drop | | CTR | 8.0% | 5.0% | Zero-click rising | | Average engagement time | 1:10 | 1:48 | More qualified click | | Organic conversions | 240 | 268 | Fewer clicks, more results |

If the picture looks like this, the conversation stops being "we lost traffic" and becomes "we traded volume for qualification." If conversions drop along with clicks, GEO isn't paying off, and it's time to revisit content, not celebrate exposure.

The Role of Structured Data in This Story

Technical honesty is needed here, because the market sells smoke. The documentation states plainly that no special structured data is required to appear in AI answers. What it recommends among best practices is something more discreet: aligning structured data with the page's visible text. In other words, schema serves to reinforce and disambiguate what's already written, not to "get into the AI Overview" through a shortcut.

In practice, well-built structured data (product, FAQ, article, review) helps Google understand entities and trust the content, which supports eligibility both in classic Search and in AI features. The mistake is treating schema as a GEO trigger. It's hygiene, not an isolated lever. When building the before-and-after, structured data enters as a control variable: "the page that keeps consistent markup and accessible text content is the one that sustained conversion even as clicks dropped."

When to Control AI Exposure (and When Not To)

The documentation also covers the other side: how to reduce presence in AI features. Since AI is embedded in Search, control runs through Googlebot itself. The tools are the usual ones for snippets:

• nosnippet, data-nosnippet, and max-snippet to limit the displayed snippet;
• noindex to remove the page from Search entirely;
• Google-Extended to restrict AI training and grounding in other Google systems, outside of Search.

The trade-off is brutal and needs to be said to the client: using nosnippet or max-snippet:0 to "escape the AI Overview" also removes the snippet that attracts clicks in traditional search. You don't just leave the AI answer, you leave the entire storefront. For almost any business that lives off organic traffic, it's not worth it. It only makes sense in specific cases (paid content, paywalls, material that can't be summarized), and even then only after measuring the click drop before deciding.

What Remains Open

The sensitive point remains: until Search Console opens up an appearance dimension for AI answers (the way it already exists for other rich results), every GEO measurement will be well-built inference, not direct reading. The documentation was last updated on December 31, 2025, and still doesn't promise that filter. For the Brazilian professional, the realistic path is to master the triangulation of impressions x CTR x engagement time and present the before-and-after with that explicit caveat. Promising exact isolation of the AI click today is selling what the tool doesn't deliver, and the client finds out at the first audit.