Let’s get one thing out of the way first. Nobody outside a small group of engineers at Google knows exactly how Gemini decides what to say and who to cite. We have official documentation, a few patents, public statements from Google’s Search team, and a pile of our own testing. That gives us a working picture, not a blueprint. So when you read this, treat it as “this is how we think it works, and here is what we do about it,” not “here are the secret levers.”
With that said, there is a lot we can say with confidence, because most of it isn’t new and isn’t magic. It’s SEO, looked at from a slightly different angle.
The biggest myth we have to clear up with clients is that Gemini is a separate search engine you can game with a separate set of tricks. It isn’t. Inside Google’s products, Gemini reads off Google’s live search index. AI Overviews and AI Mode are powered by the same crawling and ranking systems that have decided organic rankings for two decades. Google has said as much directly: its AI features are built on the core Search ranking and quality systems. When the May 2026 core update finished rolling out on the 2nd of June, Google reminded everyone that the same update can change which pages get retrieved and cited inside AI answers, not just classic rankings. That tells you everything about where AI visibility actually comes from.
So if your pages don’t rank, no “AI optimisation” tactic is going to rescue them. AI visibility isn’t a thing you bolt on. It’s what strong, well-structured organic performance looks like when a model is doing the work of answering a question.
To make the rest of this make sense, you need one mental model. There are two completely different ways an AI can “know” something about you, and people mix them up constantly.
The two ways an AI knows about you
Think of it like the difference between what a person remembers and what they look up.
The first way is training data. This is the knowledge baked into the model when it was built. It’s frozen at a point in time (the training cutoff) and it never updates until the model is retrained. When Gemini answers a question about something general and well established, like what a heat pump is or who founded a famous company, it’s often pulling from memory, no live search needed.
The second way is grounding, also called retrieval augmented generation, or RAG. This is the model looking things up in real time while it writes the answer. For Gemini inside Google’s products, grounding means querying Google’s live search index, pulling relevant pages, and using them to build a current, sourced response with clickable links. Google’s own definition is plain enough: grounding retrieves relevant, up to date pages from the Search index so the answer is more accurate and fresh.
Here’s why this distinction matters so much. Training data is slow, frozen, and almost impossible to control. Grounding is live, current, and tracks your organic strength almost directly. For 95% of businesses, grounding is the lever that actually moves, and grounding rewards exactly the SEO fundamentals you should already be working on. We’ll come back to that.
But let’s give training data its proper section, because it’s genuinely useful to understand.
Getting into the training data (and why you can’t really “optimise” for it)
When a model like Gemini is first built, it’s trained on an enormous amount of text. A big chunk of that comes from open web crawls. The most famous one is Common Crawl, a non profit that has been crawling the web since 2008 using a bot called CCBot. Its monthly snapshots run to hundreds of terabytes of text across billions of pages, and most major language models have used it, directly or through cleaned versions of it like C4, FineWeb or RefinedWeb. On top of that, model builders lean heavily on high trust anchors like Wikipedia, books, academic papers and code.
So how does your content end up in there? Honestly, mostly by being on the open web and being good enough to survive the filtering. Here’s the part people don’t like to hear: model builders aggressively clean and deduplicate these crawls. They throw out near duplicate text, downweight what looks machine generated, and lean on sources that are widely referenced and clearly authoritative. A thin page that restates what a thousand other pages already say has very little chance of surviving that filter, let alone shaping how the model talks about your topic. Individual business pages rarely make a meaningful dent on their own.
That gives you four realistic things to influence, none of which are hacks:
Be reachable by the bots that build these corpora. If you’ve blocked CCBot or the Google-Extended directive, you can’t be picked up. More on Google-Extended below, because it’s the single most common own goal we find.
Earn references from sources the model already trusts. This is the big one. Models learn who you are partly from how often and how consistently other respected sources describe you. Getting mentioned and described accurately in industry publications, reputable directories, news, and ideally a Wikipedia or Wikidata entry does far more for training data presence than anything on your own domain. You’re not feeding the model directly. You’re making sure the sources it does eat from describe you correctly.
Publish early. Training data has a cutoff. Content that exists before a model is trained has a chance to be included. Content published after it won’t touch that version of the model, only future ones. So if you care about long-term presence, get authoritative content out there now rather than waiting.
Keep your identity consistent. If your name, category and core facts are described the same way everywhere, the model is far more likely to associate the right information with you instead of getting you confused with something else.
Now the honest caveat. Even if you do all of that, you can’t guarantee inclusion, you can’t control how the model represents you, and the data is frozen the moment training ends, so it ages. This is precisely why we don’t tell clients to chase training data presence as a primary goal. It’s a slow, long-term byproduct of being a genuinely authoritative source. The faster, more controllable game is grounding.
Grounding: where the day-to-day battle is won
When Gemini grounds an answer, it’s running live searches against Google’s index and assembling a response from what it finds. That means your eligibility to be grounded is overwhelmingly a function of how well you rank and how cleanly you’re indexed. If Google can’t crawl, render and index a page, and serve it with a snippet, that page can’t be grounded. Full stop.
This is the bit that connects everything back to ordinary SEO. The work that makes you eligible for grounding is the work you should be doing anyway: clean technical SEO, a tight indexing profile, content that actually answers the question, and a credible link and brand profile. There’s no separate “grounding switch.”
There is, however, one switch you can accidentally flip the wrong way, and we find it broken constantly.
Google-Extended: the setting everyone forgets
Inside Google, there are two separate access layers, and conflating them is the most common client mistake we see.
AI Overviews and AI Mode run on standard Google Search access. You can’t opt out of them without leaving Google Search entirely, and you wouldn’t want to. There’s nothing to gain by trying to hide from them.
The Gemini app and Gemini’s grounded answers via the API are governed by a separate directive called Google-Extended. The catch is that Google-Extended is a directive, not a crawler, so it never shows up in your server logs. That makes it easy to miss. We regularly find sites that blocked it during the AI opt-out wave of 2024 and 2025, often by a CMS default or a plugin, and are now quietly invisible to Gemini grounding without anyone realising. We also find it silently flipped back on or off after CMS updates. So we audit the live file, not anyone’s memory of it. If a client wants Gemini visibility, this has to be allowed and documented.
Query fan-out: one question becomes a dozen searches
This is the part that genuinely separates Gemini-era SEO from the old playbook.
In classic search, you type a query and get a list of results. One query in, one list out. Query fan-out throws that out. When you ask AI Mode or Gemini something, the model breaks your single question into a set of smaller, related sub-questions and runs them all at once, then synthesises one answer from the passages it pulls back. Google describes it in its own documentation as issuing a set of concurrent, related queries to gather more relevant results before answering.
A quick example. Say someone asks: “is underfloor heating worth it for a small bathroom renovation?” The model doesn’t just search that exact phrase. Behind the scenes it’s likely firing off a spread of related searches, something like:
- what is underfloor heating
- electric vs water underfloor heating
- underfloor heating running costs
- underfloor heating for small bathrooms
- is underfloor heating worth it
- underfloor heating installation problems
Then it stitches the best answers to each of those into one response, with links. The user never typed those six searches. The model predicted they’d be useful and went and got them.
If you want to see this for real, run a query through Dan Petrovic’s fan-out tool or Mike King’s Qforia. You feed in a seed query and it shows you the spread of sub-queries a model might generate. It’s a genuine eye opener for the first time.
Now, why does this change anything? Because of what it does to citation. In the old world, ranking number one for your head term was the whole game. In a fan-out world, the answer is assembled from whoever best satisfies each sub-question, and that might be a different source for each branch. Practitioners have catalogued roughly eight recurring types of these sub-queries, things like definitions, comparisons (“X vs Y”), alternatives, costs, “is it worth it” style judgement queries, how-to steps, reviews, and “latest” or current-year variants. The exact taxonomy matters less than the lesson: a single page that nails the head term but ignores the surrounding questions is thin inside that network, and it leaves most of the citation opportunities on the table.
The data backs this up. Surfer analysed around 173,000 URLs and found that pages which also ranked for the fan-out queries were roughly 160% more likely to show up in Google’s AI Overviews than pages that only ranked for the main term. Separate studies have suggested that a large share of pages cited in AI answers, in some analyses a clear majority, aren’t even in the traditional top ten for the original query. Methodologies vary and you should take any single number with a pinch of salt, but the direction is consistent: AI citation is a function of topical coverage and synthesis, not a straight copy of the blue links.
Follow-up queries are a related, separate thing
Worth flagging, because it often gets lumped in with fan-out. Fan-out happens inside a single answer, behind the scenes. Follow-up queries are the conversation that happens after, when the user asks “okay, what about a wet room instead?” and the model carries the context forward into the next answer. Both reward the same thing from your side: content that covers the natural next questions a real person asks, not just the first one. If you map the whole research journey, you’re covered for both.
How we actually research fan-out
We don’t try to guess every possible sub-query by hand. We triangulate from a few sources:
- Fan-out simulators (tools like DEJAN, Qforia) to see how a model might expand a topic.
- AlsoAsked for the People Also Ask tree, which shows how real questions branch.
- Google Search Console data for the actual question-style queries your site already gets impressions for. Tools that surface and filter GSC data make this much faster.
Where those three overlap, you’ve usually found the high value coverage gaps. That overlap becomes the content brief.
Optimise the passage, not just the page
Once you accept that answers get assembled from self-contained chunks pulled from across the web, the way you structure content has to change. The unit that gets cited is the passage, not the whole page. So we write so that the most valuable blocks can stand on their own and be lifted straight into an answer.
In practice that means:
Each section answers one discrete question and makes sense without the paragraphs around it. If you lifted any single H2 block out and showed it to someone cold, it should still read as a complete answer.
Lead with the answer, then support it. Put the citable claim at the top of the block and the detail underneath. Models, and impatient humans, both reward this.
Write with real information density. Specific figures, named entities, dates, concrete steps. “Exercise is good for you” gets filtered out. “Roughly 150 minutes of moderate cardio a week is the standard public-health recommendation” is the kind of grounded, attributable claim that gets picked up. Where a fact comes from a named body or study, say so inline.
Use the format the question implies. If the query is a comparison of several things, a clean table beats a wall of prose, and it’s far more extractable. If order matters, use real numbered steps.
Keep the facts in text. The single most common technical own goal is burying the important numbers inside an image, a PDF, or an interactive widget the model can’t read. If it isn’t in the rendered text, assume it doesn’t exist as far as Gemini is concerned.
Make sure Google knows who you are: entity signals
Gemini leans hard on Google’s Knowledge Graph, which makes entity clarity unusually valuable. Entity clarity just means Google understanding, without ambiguity, who you are, what you do, and what you should be associated with. Ambiguity is what causes an AI to describe you wrongly or hand your expertise to someone else.
Google doesn’t strictly require structured data for its AI features, but used properly it removes that ambiguity. So we ship clean Organization schema (and the right type-specific schema for each client, whether that’s LocalBusiness, Product, Article or others), with sameAs links pointing to verified profiles that anchor your identity. For local clients we line up the LocalBusiness schema with the Google Business Profile so location answers are accurate. And we push for third-party consensus, because an AI’s confidence in recommending you grows when independent, trusted sources already describe and endorse you, not just your own site.
If you want to see what a model currently believes about your brand before you start, tools like WAIKAY (What AI Knows About You) are a useful starting point for spotting gaps and mis-descriptions.
The honest bits people skip
A few things we say out loud to every client, because overclaiming is its own kind of amateur hour.
You can’t cleanly measure “AI traffic.” Google folds AI Overview and AI Mode performance into your total Search performance, so Search Console won’t hand you a tidy “AI” line. The best you can do is approximate. We segment GA4 referrals from gemini.google.com and similar sources, track how often and how accurately your brand gets named in AI answers as a metric in its own right, and watch how the AI describes you. A brand being mis-described is a fixable entity-signal problem, not a ranking one, and catching it early matters because it shapes how a lot of answers represent you.
Freshness is real, but spoofing it isn’t. Google weights recency for time-sensitive topics, so genuine updates earn a re-crawl and can help citation on “best,” “latest” and year-stamped queries. But changing a dateModified without changing the content does more harm than good now. Google can compare an old snapshot to the current page, so faking the date just flags you. Update the date when there’s real new content behind it, and run a deliberate refresh cadence on the pages where freshness actually matters.
Fix the SEO first. This is the one we’ll keep saying until we’re hoarse. If your site has a ranking or indexing problem, AI visibility is the wrong thing to be chasing. Get the fundamentals right (technical health, a clean indexing profile, genuinely helpful content, a credible link and brand profile) and the AI visibility tends to follow, because it’s downstream of all of that. Google is deindexing huge volumes of low-value pages, and ranking losses only make it harder to get cited anywhere. The order of operations is not optional.
One last bit of perspective. For all the noise, AI search is still a small slice of total search volume. It’s growing fast and it’s worth getting ready for, but it isn’t a reason to torch a strategy that’s bringing in real traffic. Build the foundations, cover the questions properly, keep your identity clean, and let the AI visibility come as a result.
The short version
- Gemini inside Google reads Google’s live index. AI visibility is mostly downstream of ordinary SEO.
- There are two ways a model knows you: frozen training data (slow, hard to control, earned through being widely and accurately referenced) and live grounding (fast, controllable, tracks your organic strength). Spend your energy on grounding.
- Check Google-Extended on the live file. It’s the most common accidental block, and it’s invisible in your logs.
- Query fan-out turns one question into many. Cover the whole research journey, not just the head term, and structure content so each passage can be lifted out on its own.
- Build clear entity signals with schema, consistent identity, and third-party consensus.
- Be honest about measurement, update content for real reasons, and fix your SEO before chasing any of this.
Want to know where you stand?
If you’re not sure whether your site is set up to be cited by Gemini, or whether something like Google-Extended is quietly working against you, that’s exactly the kind of thing we check. At SWOT Digital we run AI Search Optimisation audits that look at your technical setup, your indexing, your content coverage and your entity signals, then turn it all into a clear, prioritised plan. Get in touch and we’ll tell you honestly what’s helping, what’s hurting, and what to fix first.
