How to Track Google AI Mode in Search Console

On June 3, 2026, Google rolled out a significant update for SEO professionals: dedicated Search Generative AI performance reports in Search Console. These reports provide long-awaited visibility into how often your pages appear within AI-driven features like AI Overviews and AI Mode. However, the new interface has notable gaps. It lacks query-level data, and as of August 11, 2026, I confirmed on my own property that neither the Search Analytics API nor the BigQuery bulk export provides access to this generative AI data. That means there is currently no official, automated route to track AI Mode traffic by specific search terms.

Yet, the queries are surfacing anyway. Fragments of real AI conversations, including replies like “yes go on,” follow-up prompts, and full pasted requests, have begun showing up as standard queries in the performance report. SEO analyst Anastasia Kourou flagged these unusual strings on LinkedIn, prompting a response from Google’s John Mueller. He confirmed that this data has always been present in the performance report and directed users to Google’s official documentation.

I have documented my own 16 months of data in a separate, detailed write-up, and I have also built a method for extracting these queries using my Google Search Console MCP server. Below, I will walk through the practical approaches for pulling AI-related queries out of Search Console, and I will introduce a free tool powered by a custom machine learning model that automates the entire process.

The Race To Uncover AI Queries

Once the existence of these conversational strings became public knowledge, the SEO community quickly began developing ways to isolate them.

The earliest attempts, dating back to June 2025, relied on simple word-count regex filters. These surfaced long, conversational queries by length alone, typically anything with 32 or more words. Barry Schwartz compiled these initial efforts, crediting Metehan Yeşilyurt, Vijay Chauhan, and David Konitzny. Since then, the techniques have become far more sophisticated.

Method 1: Glenn Gabe’s Full Inventory via Excel

The performance report interface limits any table to 1,000 rows, which means on larger sites, most of your query data remains hidden. Glenn Gabe’s solution bypasses the UI entirely. He retrieves the complete query set through the Search Analytics API using Analytics Edge in Excel, then uses Claude to organize the list and flag potential AI Mode queries. If you are comfortable working in spreadsheets, this approach places the entire inventory directly within your existing workflow.

Method 2: Jean-Christophe Chouinard’s Custom Regex

Jean-Christophe Chouinard tackled the same problem from inside the report itself. On August 14, 2026, he published a regex pattern on LinkedIn designed to flag conversational strings directly in the performance report’s Query filter using the Custom regex option. He developed this pattern by comparing common LLM prompts, drawn from his own data sources, against strings that were rarely seen in Search Console before AI Overviews emerged.

The result is a lengthy alternation that covers prompt verbs (write, draft, generate, summarize, explain, act as), greetings, acknowledgements (yes, ok, sounds good, yes go on), refusals, and follow-ups (more, continue, show me more, any other options). Paste the pattern into the filter, and you get an instant, free view inside the report you already have open.

Separately, his January 12, 2026 case study demonstrated that AI Mode clicks are tracked in Search Console, but the underlying queries are almost exclusively anonymized. That caveat applies to every method discussed here.

Method 3: Amin Foroutan’s Advanced GSC Visualizer

Amin Foroutan’s Advanced GSC Visualizer is a free Chrome extension that adds advanced charting, annotations, and an AI assistant to your Search Console data, complete with one-click API access. However, it does not include a filter specific to AI Mode.

Method 4: My MCP Servers

My Search Console MCP includes a built-in conversation detector tool called genaiconversationqueries. My BigQuery MCP runs the same detector against the bulk export, which is where the anonymized pool of queries resides. Both run locally, ensuring your data moves only between you and Google. Their strength lies in automation and scale, with query-level labels applied to every row. The main limitation is the setup requirement; you need an MCP client and configuration, making these tools best suited for technical users.

Comparing The Four Approaches

| Method | Strong At | Limits | | :— | :— | :— | | Glenn Gabe, API into Excel with Claude | Full query inventory in an Excel workflow | Per-run sorting, pattern-based | | Jean-Christophe Chouinard, custom regex | Instant free filter inside the report | English pattern list | | Amin Foroutan, Advanced GSC Visualizer | Charting and exploring without exports | No AI Mode specific filter | | My MCPs | Automation and scale, labels on every query | Needs an MCP client and config |

Looking down the Limits column, the first three methods share a common weakness. Pattern lists and per-run sorting cannot reliably catch edge cases such as rank tracker probes, agent harness prompts, pasted strings, or “my location is” probes. They are also built for English, so a reply in Tamil or a code-mixed Hinglish string passes through unlabelled. Method 4 was designed to solve this problem with a custom ML model.

Why An ML Model Is Necessary

Classifying every query into named buckets, across multiple languages, while including the strange machine-generated strings, is extremely difficult to do accurately with conventional methods. The sheer number of edge cases makes rule-based systems unreliable. The only dependable way to achieve this level of classification is to train a machine learning model. So I decided to build one.

How I Built The Detector

The detector operates in two parts. Deterministic rules handle the exact classes, such as reply artifacts, tracker probes, and agent harness prompts. A trained model manages the fuzzy boundary between conversational, long-tail, and ordinary queries.

The model is based on FacebookAI’s xlm-roberta-base, which is pre-trained on 100 languages. I fine-tuned it on my own labelled Search Console conversation data, plus synthetic rows across eight languages, including code-mixed Tanglish and Hinglish. I validated it against just over 120,000 scored queries and retrained it twice on whatever broke. It runs quantized on a scale-to-zero Cloudflare container, and the same model now powers a free tool that requires no setup.

The Tool

That tool is the AI Mode and AI Overview query classifier, available free on this site. No sign-up, no email required. If the GSC MCP route sounded like too much effort, this is the same detector with all that effort removed.

How It Works

Drop in Search Console or BigQuery query exports as CSV files, and you can stack several files into a single run.

Step 1: The model classifies your queries into one of seven buckets.

Step 2: If you add a Search Console generative AI report, the tool extracts page data and maps it against your queries. In your report, you will see an “AI page” column, and the full URL appears when you export the CSV.

The deterministic rules label the obvious buckets right in your browser, and those queries never leave your device. I do this to minimize the amount of data I process and to speed things up. Everything else goes to the model, gets classified in memory, and nothing is stored. Each query comes back with a bucket and a confidence score. In plain terms, the buckets are: full conversational queries, short replies like “yes go on,” follow-up pivots like “what about the pro plan,” rank tracker probes, agent harness prompts, pasted strings, and ordinary searches. You can filter by bucket, classifier, and confidence, then export every labelled row to CSV with no row cap.

How To Use It

Five steps, start to finish.

1. Get your queries out. Under about 1,000 queries, the performance report’s normal export is already complete. Larger sites can stack several filtered exports, pull the full list through the API with Search Analytics for Sheets or Looker Studio, or dump the BigQuery bulk export table to CSV. If you are an SEO, you already know how to get more than 1,000 queries out of GSC.LimitationsThe tool comfortably handles up to 100,000 unique queries per run in the browser. A large run requires keeping the tab open, which I admit is a very 2010 way to compute. The second limit applies to every method of tracking AI Mode in Google Search Console, including mine. Google anonymizes rare queries; conversation strings are rare, and a portion of the pool never shows up. Treat the output as an undercount. My BigQuery export shows that 57.7% of my impressions sat in the anonymized pool over the last 59 days, measured on August 11, 2026.The model costs me real money to run, and as you know me, I do not sell anything, and none of my tools require sign-ups or paywalls. So I have to put some limitations and caps in place so everyone can use this tool. Please play nice. You do not want to be the person who ruins things. The tool is in beta, and if lots of people start using it, it will slow down, and processing may take longer. In that case, wait and try again.Also, kindly note that I offer my time and expertise free of charge, so please do not demand customer support. Yes, I have to write this because some people do not act like reasonable adults.Over 100,000 QueriesAbove 100,000 unique queries, we need to do a batch run instead. Send yours in, and my agency runs it as a proper batch job, then emails you the complete results. We can also turn the findings into a full, actionable report or assist with your wider AI SEO work.