Google-Extended controls training and grounding for Gemini apps and Vertex AI. It does not govern AI Overviews or AI Mode, which run on ordinary Googlebot rules plus a separate Search Console setting introduced on 3 June 2026. That setting is the real control, and you must not be opted out to appear. South Africa has no access to Google's generative AI performance report yet.
If somebody told you to block Google-Extended to keep your content out of AI Overviews, they gave you the wrong instruction. Google’s own crawler documentation states that Google-Extended “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal”. AI Overviews and AI Mode are Search features. You can block Google-Extended and keep appearing in both.
The control that does govern AI features arrived on 3 June 2026, lives in Search Console, and has nothing to do with robots.txt.
Three Google surfaces get treated as one thing in most guides. They have different controls, different citation behaviour, and different measurability. Separating them is the whole job.
The three surfaces
| Surface | How you control access | Shows sources | Can you measure it |
|---|---|---|---|
| Gemini app | Googlebot crawls; Google-Extended governs training and grounding | Yes, source links in answers | No |
| AI Overviews and AI Mode in Search | Ordinary Googlebot rules, plus the Search Console AI setting | Yes, supporting links | In part, where the report has rolled out |
| Gemini API and Vertex grounding | Googlebot crawls; Google-Extended governs grounding at prompt time | Yes, structured citation metadata | No |
What Google-Extended is, and is not
It is not a crawler. Googlebot does all the fetching. Google-Extended is a robots.txt token that downstream pipelines read when deciding whether your content may train future models or ground answers.
Its documented scope in 2026 covers training models that power Gemini apps and the Vertex AI API, plus grounding through Google Search on Vertex. That is the boundary.
Two newer tokens exist and get missed. Google-GeminiNotebook, renamed from Google-NotebookLM, handles user-triggered fetches for notebook-style research and is scheduled to sunset in August 2026. Google-Agent covers agentic browsing. Both sit outside Google-Extended.
The tension worth knowing about
Google’s position is clear. Blocking Google-Extended does not affect AI Overviews.
An academic study published in April 2026 complicates it. Researchers ran a benchmark of 11,500 queries, 14,212 in total, across Google Search, Gemini and AI Overviews. Among their findings: sites blocking Google’s AI crawler saw lower retrieval by AI Overviews, a difference the researchers report as significant, despite the content remaining accessible.
Those two statements sit in tension and we cannot resolve them. Google documents no causal link. One study measured a correlation. Anybody presenting either as the settled answer is overreaching.
Our practical reading: if you want AI Overviews visibility, leave Google-Extended alone rather than blocking it as a precaution. The upside of blocking is a training opt-out you may not need. The downside sits in that correlation.
The control that governs AI features
Google announced website-owner controls on 3 June 2026 alongside a generative AI performance report. The controls include a Search Console setting that opts your content out of AI Overviews, AI Mode and AI features in Discover.
Google’s own optimisation guide now states that a site must be included in that setting to be eligible for display in AI features. That is a meaningful change to the older line about no additional technical requirements beyond normal Search eligibility. A requirement now exists, and it is a toggle rather than a file.
Check it. A setting you have not opened is a setting you cannot vouch for.
What Google tells you to do, and what it tells you to stop
Google’s guidance stayed consistent through 2026: meet the standard Search technical requirements, stay indexed, stay snippet-eligible.
It names things to avoid, in its own words. llms.txt and similar AI text files, because “Google Search itself doesn’t use them”. Breaking content into tiny chunks, since Google states there is no requirement to do so and no ideal page length. Rewriting content for AI systems or stuffing long-tail keywords, because those systems understand synonyms. Chasing inauthentic mentions across the web. Overfocusing on structured data, which supports rich results in general rather than AI features in particular.
That list is worth keeping. Four of the five appear in South African agency proposals we have read.
What Google recommends instead reads like ordinary competence: crawlability, internal linking, page experience, content available as text, structured data that matches what visitors see, accurate Business Profile and Merchant Center data.
Grounding sends traffic through a redirect
For anyone building on the Gemini API, the citation metadata is thorough. Google returns the search queries it generated, a required Search Suggestions block, source chunks with URI and title, and a mapping from each text segment back to the chunks that support it.
One detail has direct consequences. The URI in each grounding chunk is a Google redirect on vertexaisearch.cloud.google.com, not your domain. Google’s own developer forum carries an open request to return real source URLs, unresolved at the time of writing.
Grounded traffic therefore reaches you through Google’s redirect infrastructure. Your server logs and referrer data undercount it, and the click path stays opaque to you. That is an acknowledged limitation rather than a configuration mistake on your side.
South Africa cannot see the report yet
Google’s generative AI performance report launched to a subset of United Kingdom site owners on 3 June 2026. By late June, sightings appeared in the United States, India and Switzerland. Google’s help documentation still describes access as limited to a subset of owners and publishes no country list. John Mueller has confirmed an incremental rollout with no schedule.
We found no source confirming South African access. Open Search Console and look rather than trusting any list, including this one.
When it arrives, know what it gives you. Impressions, pages by canonical URL, countries, devices, and date breakdowns for Search. It withholds click data. Reporting is said to begin from 18 May 2026 forward, which Google has not confirmed.
Until then, Bing Webmaster Tools remains the only free AI citation measurement a South African site owner can act on today.
What the South African numbers say about Gemini
World Wide Worx ran the South African Generative AI Roadmap 2025 with Dell Technologies and Intel, surveying more than 100 mid-sized and large South African enterprises with over 200 employees, weighted toward manufacturing, logistics, construction and retail. Gemini usage among those respondents fell from 56 percent in 2024 to 41 percent in 2025. Arthur Goldstuck attributed the drop to Gemini not keeping pace with other free tools.
Hold the scope in view. That measures tool adoption inside a hundred or so large companies. It says nothing about consumer search behaviour, and nothing about whether Gemini recommends small businesses.
The Gemini app is available in South Africa on Google’s own supported-countries list, with Afrikaans and Zulu marked as language-level support rather than full regional customisation. English has full support.
Android distribution and Workspace bundling may give Gemini a structural advantage here. We found no South African data quantifying it, so we are not going to assert it.
What the strongest study found
The April 2026 benchmark is the most rigorous work we located: an academic preprint, 11,500 queries, university-affiliated authors, no disclosed commercial funding.
AI Overviews appeared for 51.5 percent of representative real-user queries, positioned above organic results.
The sources differ across surfaces more than most people assume. Average Jaccard similarity between what Google Search, AI Overviews and Gemini retrieved came in below 0.2. Ranking on Google does not carry you into either AI surface.
Traditional Search favoured institutional, government and education sites. The generative surfaces skewed toward Google-owned content. AI Overviews proved less consistent across repeated runs and sensitive to small changes in wording.
The other citation-count figures we found came from commercial vendors, and they contradict each other. Gemini’s average citations per query gets reported as 3.4, 6.2 and 11.0 by three different sellers of visibility services. We quoted a version of the low figure ourselves before checking, and we were wrong to.
What to do this week
Open Search Console and confirm the AI features setting is not opted out. Google now names that as an eligibility requirement.
Leave Google-Extended allowed unless you have a specific reason to refuse training use, and understand you are choosing between a training opt-out and a correlation with reduced AI Overviews retrieval.
Delete any llms.txt work from your scope, because Google says it does not use it.
Then run the queries yourself across Gemini, AI Mode and ordinary Search, and compare who each one names. The sources overlap less than 0.2, so treating one result as a proxy for the others will mislead you.
Where this leaves you
Google gives you more documented control than any other engine and less measurement than South African businesses need. The controls are real, sit in two separate places, and one of them is a toggle most owners have not opened.
Start with the Search Console setting. Then test each of the three surfaces on its own, because Google’s own numbers say they disagree with each other.
If you want that comparison run across Google’s surfaces and five other engines with the raw results handed over, our AI search visibility service sets out the scope, the price, and what we cannot promise.