Anastasia Braitsik stands at the intersection of data science and digital visibility, guiding brands through the seismic shifts of the generative search era. As a global leader in SEO and content marketing, she has spent years dissecting how algorithms interpret human intent across a hundred different languages. Today, she breaks down the latest evolution in Google Search Console—the introduction of generative AI performance reports—and reveals the hidden methods for uncovering the conversational queries that Google hasn’t yet officially exposed through its API.
The transition to generative search has fundamentally altered how performance data is captured and interpreted by digital marketers. Our discussion focuses on the new Search Generative AI performance reports released on June 3, 2026, and the technical hurdles of accessing query-level data that remains hidden within the user interface. We explore the “leak” of conversational fragments in standard reports, compare manual extraction methods like custom regex and Excel workflows, and dive into the development of a custom machine learning model designed to classify millions of unique, multi-lingual search strings. Finally, we address the persistent challenge of anonymized data and the necessity of specialized tools to quantify the true impact of AI Overviews on web traffic.
Google Search Console recently unveiled generative AI performance reports, yet the lack of query-level data in the API or BigQuery export leaves many marketers feeling like they are looking at a puzzle with missing pieces. How are you navigating this discrepancy to understand what is actually driving impressions in AI Overviews?
The rollout of the Search Generative AI performance reports on June 3, 2026, was a landmark moment for our industry, finally providing a dedicated view of impressions within AI Overviews and AI Mode. However, as of my latest re-verification on August 11, 2026, neither the Search Analytics API nor the BigQuery bulk export has been updated to expose this specific generative data. This creates a significant visibility gap because, while we can see the aggregate volume in the UI, we cannot officially map those impressions to specific user queries. To navigate this, I have been analyzing 16 months of historical data to identify the exact moments where conversational strings cross over into standard reporting. It requires a pivot from relying on official documentation to building custom extraction layers that can bridge the gap between what Google shows in the dashboard and what is actually happening on the server side.
It was quite a revelation when conversational fragments like “yes go on” began appearing as ordinary queries in performance reports. Could you describe the impact of this “leak” on our understanding of user behavior within AI Mode?
The appearance of these fragments—ranging from simple acknowledgments like “ok” to entire pasted prompts—confirmed that the data for AI Mode interactions has actually been available in our performance reports all along, just hidden in plain sight. When these strings started surfacing, it felt like pulling back a curtain on a completely different type of search behavior that is much more fluid and iterative than traditional keyword entry. I’ve seen cases where follow-ups and conversational pivots are recorded as standard queries, providing a rare sensory glimpse into the “back-and-forth” nature of modern search. This “leak” is invaluable because it allows us to see the messy, human side of AI interaction, even if Google’s official stance is that this data is categorized under ordinary performance metrics. By capturing these strings, we can begin to model the full lifecycle of an AI search session rather than just seeing a single, isolated click.
With several experts developing different workarounds—from Excel macros to custom regex—which approach do you find most effective for a team trying to scale their AI traffic analysis without getting bogged down in manual sorting?
For teams that need to scale, the choice really depends on their existing technical stack and the volume of data they handle daily. If your team lives in spreadsheets, the method of pulling a full query inventory through the Search Analytics API into Excel and then using an LLM like Claude to flag patterns is highly effective because it avoids the 1,000-row limit of the standard UI. On the other hand, the custom regex approach published on August 14, 2026, is a brilliant “quick-fix” for smaller sets, as it allows you to filter for prompt verbs like “summarize,” “explain,” or “act as” directly within the Search Console interface. For high-volume environments, I prefer using MCP servers that run locally, as they can automate the detection of conversation queries across entire BigQuery exports where the anonymized data lives. Each method has its own friction point, but the goal is always to move beyond the pattern-matching of English strings and toward a more robust, automated classification system.
You’ve moved beyond simple pattern matching to develop a classifier based on FacebookAI’s xlm-roberta-base. What specific complexities in conversational search necessitated a machine learning model over traditional deterministic rules?
Deterministic rules and regex are excellent for catching specific verbs like “write” or “draft,” but they fail miserably when faced with the “fuzzy” boundaries of conversational intent, especially in a multilingual context. I chose to fine-tune the xlm-roberta-base model because it was pre-trained on 100 different languages, allowing it to understand the nuance of code-mixed strings like “Tanglish” or “Hinglish” that standard English-centric filters miss. After validating the model against over 120,000 unique queries, it became clear that a machine learning approach was the only way to accurately distinguish between a long-tail search and a genuine AI conversation. The model handles the complex edge cases, such as agent harness prompts or rank tracker probes, that would otherwise clutter our data with false positives. By running this model in a quantized, scale-to-zero container, we can process massive datasets with a level of accuracy that a list of keywords simply cannot match.
Your new tool promises to categorize queries into seven distinct buckets, such as “follow-up pivots” or “agent harness prompts.” How does the system balance the heavy lifting of ML classification with the need for data privacy and speed?
The architecture of the tool is designed to be as “browser-first” as possible to ensure that sensitive search data stays under the user’s control. We use deterministic rules to label obvious buckets—like simple reply artifacts—directly in the browser, which means those queries never even have to leave your device. For the more complex strings that require the model, we send them to a memory-only environment where they are classified and then immediately cleared, ensuring nothing is ever stored on our servers. This hybrid approach allows us to maintain high speeds even when a user drops in a CSV containing 100,000 unique queries, providing a confidence score for each classification. It’s a very 2026 way of handling big datmaximize the local processing power of the user’s machine and only use the cloud for the “intelligence” heavy lifting.
Even with advanced detection, the shadow of anonymized data looms large, with some reports showing over 50% of impressions hidden in the “anonymized pool.” How should SEOs interpret their findings when a significant portion of the conversation remains invisible?
We have to accept that any data we extract is essentially a conservative undercount of the total AI activity on a site. My own BigQuery analysis from August 11, 2026, showed that approximately 57.7% of impressions over a 59-day period were tucked away in that anonymized pool, which is a staggering amount of invisible traffic. Because conversational strings are often highly unique and rare, they are the first to be anonymized by Google’s privacy thresholds, meaning the “leaked” queries we do see are just the tip of the iceberg. I advise marketers to treat the visible data as a representative sample—a directional compass rather than an absolute ledger. If you see a specific category of conversational queries growing in your visible reports, you can safely assume that a proportional growth is happening in the anonymized shadows as well.
What is your forecast for the evolution of generative search reporting?
I anticipate that by the end of this year, the pressure from the search marketing community will force a more transparent integration of AI Mode data into the official Search Analytics API. We are moving toward a reality where “search” is no longer a destination but a continuous dialogue, and our reporting tools must evolve to track the sentiment and “flow” of a conversation rather than just the final click. As machine learning models become more efficient, I expect to see these classification capabilities baked directly into the platforms we use, making it easier to see how AI Overviews are influencing the entire marketing funnel. For now, the most successful SEOs will be those who aren’t afraid to build their own tools and use models like xlm-roberta to find the signals that Google isn’t yet ready to hand over on a silver platter.
