Tracking Brand Visibility in Google AI Overviews and AI Mode

AI Overviews is an enrichment: it appears above classic results for queries Google deems suitable, synthesises a short answer from pages in its index, and links its sources. The user didn't opt into an AI experience — it happened to their search. AI Mode is the opposite: a deliberate conversational surface where the user asks in natural language and follow-ups are expected, much closer to a ChatGPT session that happens to run on Google's index. They differ in trigger, length, tone and — measurably — in who they name. On our own account last month, AI Mode produced roughly twice the brand mentions of AI Overviews (91 against 46, of 2,940 runs). Same brand, same Google, same month. Any tool or article that treats "Google AI" as one surface is averaging two different games.

Neither surface has a public API, so honest tracking reads the rendered answer — the thing a searcher actually sees — from clean, neutral sessions. That capture route has real properties to respect. It must be locale-true: AI Overviews triggered from Germany in German are different answers from the same query in English from the US, so per-market checks matter more here than anywhere. And it must be sampled like everything else in this field: AI Overviews doesn't even trigger on every query every time, so presence is a rate across repeated runs, never a screenshot (the sampling argument we made in full in Scraping vs API applies doubly to surfaces where rendering is the only option). That's how CrunchJunkie measures both surfaces — rendered answers, neutral sessions, per-prompt market and language, repeated runs with margins — alongside the API-based engines, in the same report with the same sample-size discipline.

Engines like ChatGPT and Claude decompose a prompt into their own web searches before answering — fan-out queries, which they expose and we capture verbatim. Google's AI surfaces don't do this, by construction: they already sit on a search engine, so the user's query is the retrieval. There is no hidden intermediate search to optimize for. The strategic consequence is bigger than it sounds. For ChatGPT visibility, you optimize for machine-generated search strings that differ from the prompt. For AI Overviews and AI Mode, the target is the query itself — which is to say, the surface rewards exactly the asset you may already have: pages that rank for the buyer's actual search. Classic organic strength doesn't transfer perfectly to AI Overviews inclusion, but it is the entry ticket; pages absent from the index don't get synthesised from at all.

Since mid-2026, Search Console reports generative-AI performance for Google's surfaces — the first first-party data in this category. Know its edges before leaning on it: impressions-level only (no clicks, no CTR, no query detail for AI surfaces), Google properties only, and nothing about what the answer said. It will tell you that pages surfaced in AI experiences; it cannot tell you whether your brand was named, how you were described, or who was named instead. That's the split to build on: Search Console for Google's own impression telemetry, answer-level tracking for the part clients actually ask about — were we in the answer, ahead of whom, said how. The two disagree in informative ways; a page with AI impressions but no brand mentions is being used as raw material without being recommended, which is its own diagnosis.

Strip the folklore and three levers remain. First, index presence and rank for the buyer queries themselves — both surfaces synthesise from pages Google already ranks for the query, so the old work is the new work. Second, liftability: clear claims, sourced statistics and quotable passages raise the odds your page is the one the synthesis borrows — the same signals shown in the KDD 2024 GEO research to raise citation likelihood by up to ~40%, and they apply here with the least friction because retrieval is already solved. Third, machine readability: server-rendered content and clean structured data, because a page the indexer parses cleanly is a page the synthesiser can quote precisely. What doesn't appear on that list: llms.txt (Google doesn't use it), fan-out optimization (there is none here), and blocking-related fixes that matter for other engines — Google's AI surfaces read the index, not your server, which is why our crawler-log study found Google fetching a site just 24 times while OpenAI fetched it 1,421 times. Different plumbing, different fixes.

Because AI Overviews trigger probabilistically and AI Mode answers drift, both surfaces punish anecdote-driven strategy harder than the chat engines do. Get a baseline first: your buyer prompts, both surfaces, repeated runs, per market — the free visibility check is the two-minute version, and full tracking runs both Google surfaces alongside the other eight engines with every number carrying its n. On the accounts we watch, the Google surfaces are routinely where the biggest gap between assumed and measured visibility lives — in both directions.