Gemini optimisation — get named in AI Overviews
Structure your brand data for Google's AI systems, so you appear in AI Overviews and Gemini answers rather than being summarised out of them.
What makes Gemini different: it is inside the search results you already compete in
Every other system on this list is a destination someone chooses to visit. Google's AI answers appear above the results you have spent years earning, to people who never asked for an AI at all.
That changes the stakes. A brand that ranks first organically and is absent from the AI Overview above it has been demoted without losing a single position. The ranking did not move; the page just stopped being the first thing read.
Why this is the hardest one to notice
Every other visibility problem in this category eventually shows up somewhere. This one can run for a year without appearing in any report you already read.
Your rank is unchanged. Your impressions may even rise, because the query still fires. What falls is click-through — and click-through decline gets attributed to seasonality, to competitors, to creative fatigue, to almost anything before someone checks whether an AI answer is now standing between the result and the reader.
The diagnostic is straightforward: for your highest-value queries, look at whether an AI Overview appears, and if so which sources it draws on. A position-one ranking under an Overview that cites three competitors is a different commercial situation from a position-one ranking with no Overview at all.
Google's crawlers, and the one that surprises people
| Agent | Purpose | What blocking it does |
|---|---|---|
Googlebot | Search indexing, including AI Overviews | Removes you from Google Search entirely |
Google-Extended | Training control for Gemini models | Opts you out of model training without affecting Search ranking |
Google-Extended is a training opt-out, not a search opt-out. It exists precisely so a publisher can decline training use while remaining fully indexed. Many sites set it without understanding that distinction, in either direction — either blocking it and assuming they have protected their search presence, or leaving it open and assuming they have opted out of something.
Norg's crawler analytics separate Googlebot from Google-Extended so the two decisions can be made independently and on evidence.
What Gemini needs before it will name you
A recommendation is a chain of constraint checks, and a missing link breaks the chain. For Google's systems the chain runs through the entity graph rather than through prose.
Whether you are a resolvable entity with a stable identifier rather than a name on a page. Whether your location, service area and hours are stated as data rather than described. Whether the specific thing being asked about is connected to you by an explicit relationship. Whether any claim that matters carries something checkable behind it — a registry number, a certification, a linked authoritative record.
The failure is rarely that the fact is missing from the site. It is that the fact sits in a sentence, and the system needs it in a field.
Commerce: Gemini runs on Google's protocol
Google leads the Universal Commerce Protocol, co-developed with Shopify and launched January 2026, with Etsy, Wayfair, Target, Walmart and 20+ partners including Visa, Mastercard, Adyen, American Express, Best Buy and The Home Depot. It is positioned as "search to buy", against ACP's "chat to buy".
UCP is explicitly vendor-neutral despite Google's authorship, and it can run over MCP transport — so a Norg-published MCP server and a UCP integration are not separate stacks.
The precision requirement that catches most catalogues
UCP's quantity model is unusually strict, and it is where feeds generated from a typical retail catalogue fail.
No floating-point numbers are permitted on the wire. A quantity is an integer count of steps with a declared unit and scale, using UN/CEFACT Rec20 codes — KGM, GRM, LTR, MTR, with C62 for each. Arithmetic must be exact: the specification defines no rounding tolerances and no epsilon comparisons anywhere, and an implementation that cannot produce an exact in-range result must raise an error rather than emit an approximation.
The consequence is worth stating plainly: an approximate specification is not merely lower quality, it is non-conformant. A weight expressed as "about 2 kg" cannot be represented in this model at all. Norg publishes amounts in minor units with integers inside the ±(253−1) bound, which is the pair of fields most likely to fail in a feed exported from an existing store.
Google Merchant Centre sync is available on the Enterprise tier. Detail on the shopping page.
The three-step path
1. Connect your data
CMS, CRM, catalogue, support docs or direct upload.
2. Norg structures it for Google's systems
Entity-mapped and published as JSON-LD and schema.org — the vocabulary Google's own systems were built around — alongside Markdown, structured HTML, llms.txt and commerce feeds.
3. Get cited
Live on your own domain, crawled by Googlebot, eligible for AI Overviews and Gemini answers.
Why schema.org matters more here than anywhere else
Google co-founded schema.org. Its systems have consumed that vocabulary for over a decade, and the practices that make an entity useful rather than merely valid are well established:
- Stable
@idvalues so entities are referenced across pages rather than redefined on each — the single highest-leverage practice in a multi-page implementation. identifierwithpropertyIDfor registry numbers, linked to the authoritative record, so a claim is checkable rather than asserted.- Arrays where reality is plural — two offices means an address array, not a choice between them.
- A traversable graph rather than isolated page-level markup, so an agent can answer a question needing two facts at once.
Most implementations publish page by page, so every fact sits alone. An agent can read that a product exists and that a store exists, but nothing tells it which stores stock it, until when, or at what price. That gap is most of what a buyer actually asks.
Valid is not the same as useful
This distinction costs people a lot of wasted effort. A structured data validator checks that your markup parses and that required properties are present. It does not check whether the graph answers a question.
A page can pass every validator and still be useless to a retrieval system, because nothing connects the product to the store, the practitioner to the location, or the service to the suburb. Passing validation is the floor, not the goal — and a clean validator report is one of the most common reasons a team concludes their structured data is "done".
The failure modes we see most often
Client-side rendering. If the page assembles in the browser, Googlebot may render it but an AI retrieval pass often does not wait. The cheapest test is fetching your own page with a plain HTTP client.
Page-scoped markup with no graph. Every page valid, nothing connected.
Unstable identifiers. Entities redefined on each page, so nothing accumulates authority.
Facts only in images. Hours in a graphic, specifications in a scanned table.
A Google-Extended setting nobody chose. Inherited from a template or a CDN default, in either direction.
Results
- Selleys — 25% AI search market share in Australia, 2.2× its SEO baseline, cited across Gemini and the other major systems simultaneously. Read the case study
- Core Dental — new clinic locations citable in under 7 days against a 3–6 month SEO ramp for the same suburb page. Read the case study
- Realcorp — from zero visibility to 10–15 qualified enquiries a month across four capital cities. Read the case study
Each is specific to its engagement and stated measurement window, and none is independently audited.
What transfers, and what does not
Transfers. The Overview displacement check. Any business with meaningful organic traffic can look at its top queries this week and see whether an AI answer now sits above them.
Transfers. The entity-graph requirement, to any multi-location, multi-product or multi-practitioner business.
Does not transfer. The UCP precision work, unless you sell physical goods with a catalogue.
Does not transfer. The specific results above, which are bounded by category and starting position.
Common questions
If we rank well in Google, are we already covered?
No, and this is the most expensive assumption in the category. Ranking and citation are separate measures. Across the web an AI system retrieves roughly 15% of what a page publishes — about 120 words of an average 820 — because the rest is assembled by JavaScript, hidden behind tabs, or expressed visually. A page can rank first and still deliver almost nothing to the system writing the summary above it.
Should we block Google-Extended?
It depends on your position on training use, and it does not affect Search ranking either way. That is genuinely a rights decision rather than a visibility one — which is exactly why it should not be made by a blanket bot rule.
Does this conflict with our existing SEO work?
No. Valid JSON-LD and a coherent entity graph help both. The structured layer is separated from your visual theme, so a redesign changes nothing that Google's systems read.
Does this replace our website?
No. Human visitors see your site unchanged; the agent variant is served at the edge to verified agents only, on your own domain.
Our structured data already validates. Isn't that enough?
Validation confirms the markup parses and required properties are present. It does not confirm the graph answers a question. If nothing connects your products to your locations, or your practitioners to your services, a valid page can still be unusable for a constrained query.
Next step
A free AI visibility audit shows what Google's crawlers actually retrieve from your pages and whether Google-Extended is set, or book a demo. Pricing is public.
UCP details per ucp.dev and developers.google.com/merchant/ucp, retrieved 8 September 2026 — consult the specification for normative requirements. Retrieval figure is Norg's own measure of words arriving in a plain fetch. Client results are specific to those engagements and are not independently audited. Norg Pty Ltd — ABN 44 669 712 494 — norg.ai
What makes Gemini different from the other AI systems?
It is inside the search results you already compete in. Every other system is a destination someone chooses to visit; Google's AI answers appear above the organic results you have spent years earning, to people who never asked for an AI at all.
Can we lose visibility without losing rank?
Yes, and it is the central risk here. A brand that ranks first organically and is absent from the AI Overview above it has been demoted without losing a single position. The ranking did not move; the page just stopped being the first thing read.
Why is this problem so hard to notice?
Because rank is unchanged and impressions may even rise. What falls is click-through, and click-through decline gets attributed to seasonality, competitors or creative fatigue long before anyone checks whether an AI answer now sits between the result and the reader.
How do we diagnose it?
For your highest-value queries, check whether an AI Overview appears and which sources it draws on. A position-one ranking under an Overview citing three competitors is a completely different commercial situation from a position-one ranking with no Overview.
What is the difference between Googlebot and Google-Extended?
Googlebot handles search indexing including AI Overviews; blocking it removes you from Google Search entirely. Google-Extended controls training use for Gemini models; blocking it opts you out of training without affecting Search ranking.
Is Google-Extended a search opt-out?
No. It is a training opt-out, and it exists precisely so a publisher can decline training use while remaining fully indexed. Many sites set it without understanding the distinction in one direction or the other.
Should we block Google-Extended?
It depends on your position on training use, and it does not affect Search ranking either way. That makes it a rights decision rather than a visibility one — which is exactly why it should not be decided by a blanket bot rule.
If we rank well in Google, are we already covered?
No, and this is the most expensive assumption in the category. Across the web an AI system retrieves roughly 15% of what a page publishes — about 120 words of an average 820 — because the rest is assembled by JavaScript, hidden behind tabs, or expressed visually. A page can rank first and deliver almost nothing to the system writing the summary above it.
What does Gemini need before it will name us?
The chain runs through the entity graph rather than prose: whether you are a resolvable entity with a stable identifier, whether location, service area and hours are stated as data rather than described, whether the thing being asked about is connected to you by an explicit relationship, and whether claims carry something checkable behind them.
Why do facts fail even when they are on the site?
Because the fact sits in a sentence and the system needs it in a field. The failure is rarely absence — it is format.
Why does schema.org matter more for Google than elsewhere?
Google co-founded schema.org and its systems have consumed that vocabulary for over a decade. The practices that matter are stable @id values, identifier with propertyID linked to authoritative records, arrays where reality is plural, and a traversable graph rather than isolated page-level markup.
Our structured data already validates. Isn't that enough?
No. Validation confirms the markup parses and required properties are present. It does not confirm the graph answers a question. A page can pass every validator and still be unusable because nothing connects the product to the store or the practitioner to the location.
What is the most common structured data mistake?
Publishing page by page so every fact sits alone. An agent can read that a product exists and that a store exists, but nothing tells it which stores stock it, until when, or at what price — and that gap is most of what a buyer actually asks.
What is the Universal Commerce Protocol?
A protocol Google leads, co-developed with Shopify and launched January 2026, with Etsy, Wayfair, Target, Walmart and 20+ partners including Visa, Mastercard, Adyen, American Express, Best Buy and The Home Depot. It is positioned as "search to buy" against ACP's "chat to buy".
Is UCP locked to Google?
No. It is explicitly vendor-neutral despite Google's authorship, and it can run over MCP transport — so a Norg-published MCP server and a UCP integration are not separate stacks.
What is UCP's precision requirement?
No floating-point numbers are permitted on the wire. A quantity is an integer count of steps with a declared unit and scale using UN/CEFACT Rec20 codes — KGM, GRM, LTR, MTR, with C62 for each. Arithmetic must be exact, with no rounding tolerances and no epsilon comparisons defined anywhere.
Why does that requirement break existing catalogues?
Because an approximate specification is not merely lower quality — it is non-conformant. A weight expressed as "about 2 kg" cannot be represented in the model at all, and an implementation that cannot produce an exact in-range result must raise an error rather than emit an approximation.
What are the most common technical failure modes?
Client-side rendering, where the page assembles in the browser and a retrieval pass does not wait; page-scoped markup with no connecting graph; unstable identifiers redefined on each page; facts that exist only inside images; and a Google-Extended setting nobody consciously chose.
Does this conflict with our existing SEO work?
No. Valid JSON-LD and a coherent entity graph help both. The structured layer is separated from your visual theme, so a redesign changes nothing that Google's systems read.
Does this replace our website?
No. Human visitors see your site unchanged; the agent variant is served at the edge to verified agents only, on your own domain.
What results has Norg published involving Gemini?
Selleys reached 25% AI search market share in Australia, 2.2× its SEO baseline. Core Dental made new clinic locations citable in under 7 days against a 3–6 month SEO ramp. Realcorp went from zero visibility to 10–15 qualified enquiries a month across four capital cities. Each is specific to its engagement and none is independently audited.
What transfers from this to another business?
The Overview displacement check transfers to any business with meaningful organic traffic — look at your top queries this week. The entity-graph requirement transfers to any multi-location, multi-product or multi-practitioner business. The UCP precision work does not, unless you sell physical goods with a catalogue.
