Google AI Overviews now trigger on 80% of branded queries, up from 26%

For most organisations, the company name was the one query you could not lose. As of late September 2026 it returns a Google-generated paragraph first on 80% of branded searches (DemandSphere, published 1 Oct 2026), and 93 of the 100 brands tested got an AI summary. In the UAE and GCC, where Google dominates discovery and journeys start on a phone, that paragraph is now your front page. Three business consequences follow. Branded traffic and last-click attribution quietly degrade, so inbound and pipeline numbers will drift from what your analytics show. The generated answer becomes a reputation surface where stale claims about pricing, coverage or leadership get repeated to prospects. And ownership splits between marketing and engineering, because the fix is entity data, structured facts and citable third-party coverage, not another campaign.

The mechanics matter. A jump from 26% to 80% in about two weeks is a threshold change in Google's system, not an incremental rollout, so it can move again and your rank tracker will not warn you. The Rundown AI's AI Index study (Aug 2026) measured a median 14.7% overlap between the sites in a classic top-10 ranking set and the sites that get AI visibility: Gemini 21.2%, AI Overviews 15.4%, ChatGPT 6.9%, and zero overlap for paid ads. What you rank for and what AI engines say about you are now largely different games. Behaviour compounds the exposure. Gartner puts consumer trust in AI Overviews at 32%, and Bain finds 80% of users rely on zero-click results 40% of the time. If 80% of the safest queries in your funnel get answered by Google, unbranded discovery will be worse.

Most teams will get this wrong in three ways. First, they will try to block. Google's own documentation on AI features and your website (updated 23 Sept 2026) states that Google-Extended controls grounding in the Gemini app and AI Mode, not AI Overviews in Search, and that no platform control keeps content out of AI Overviews while staying eligible for Search. Robots.txt experiments will burn weeks and change nothing. Second, they will keep reporting position and traffic when the metric that matters is citation share of the generated answer: does it say the right thing, and does it name you or a third party. Third, they will publish more English blog posts while the Arabic answers about their brand are wrong, which in this region is where the reputation risk concentrates.

How Azrty would approach it: instrument first, then fix the answer. We would stand up a probe panel of branded and brand-intent queries in English and Arabic (for example, "[brand] pricing", "[brand] vs competitor", "[brand] review"), run it weekly against Google's AI surfaces and the LLM engines, and route every model call through FastLLM Proxy so each model, budget and access rule sits in one OpenAI-compatible gateway instead of scattered API keys. The probe logs the answer text and the cited domains each week, and the dashboard reports answer presence rate and citation share, not rank. The prompts live in a shared, versioned library so the panel is reproducible when people change. Then we fix the entity layer: one canonical page of facts, pricing and policies, Organization structured data in JSON-LD on it, and identical facts across Google Business Profile, LinkedIn and the third-party profiles that grounding models pull from, plus the exact sentences we want an answer engine to lift. That is the work of our digital transformation practice, which maps how the business runs and then builds the brand, site and content layer on top of it.

The view from here: this is an open goal in the GCC. Almost no brand in the region measures citation share of AI answers today. A team that instruments its branded panel and fixes its entity data over the next quarter can own its generated answer while competitors are still debating AI SEO. The risk is the opposite: treating this as a content-volume problem and buying generic AI-visibility output with no measurement attached. The measurement is cheap, the fixes are unglamorous, and the window to be the cited source for your own name is open now.