How to rank in Google AI Mode: the playbook for getting cited

"How do I rank in Google AI Mode" is the wrong question, and getting the framing wrong is why most teams waste the first month. AI Mode doesn't produce a ranked list you climb. It produces a written answer, and your job is to be one of the sources it cites inside that answer. Those are two different games.

Here's the part that catches people out. Strong organic rankings don't carry over the way you'd expect. Google's own data shows only 14% of the URLs cited in AI Mode rank in the traditional top 10 for the same query. So a page sitting at position one can be absent from the answer, and a page nowhere near the top 10 can be cited repeatedly. This playbook is how you get on the right side of that.

Ranking and getting cited are different games. AI Mode plays the second one.

What AI Mode actually does.

You don't need the full architecture to act on it, but you do need the shape of it. When you submit a query to AI Mode, it doesn't run one search. It uses a technique Google calls query fan-out: an LLM decomposes your single query into multiple sub-queries, each targeting a different facet of the intent. Each sub-query retrieves independently, and the system synthesises the results into one answer with citations.

The retrieval unit is the passage, not the page. AI Mode lifts and cites a specific claim from your content, not your URL as a whole. That single fact changes most of what follows. If you want the mechanism in full (the patent, Deep Search, how Perplexity and Copilot differ), I covered it in how Google AI Overviews and AI Mode actually work. This post assumes that and goes straight to the doing.

What fan-out looks like in practice

Take a buyer typing "best project management tool for a remote agency". Traditional Search returns ten blue links for that exact phrase. AI Mode does something different: it expands the query into a set of sub-queries, each chasing a facet of what the buyer actually needs to know, then retrieves for each one separately. The answer is assembled from whatever sources best satisfy each facet.

So a single query fans out into something like this:

query: "best project management tool for a remote agency"

fans out to:
  - "project management tools for agencies"
  - "best PM software for remote teams"
  - "project management tools with time tracking and invoicing"
  - "Asana vs Monday vs ClickUp for agencies"
  - "project management pricing for small teams"

each sub-query retrieves its own passages
answer = synthesis of the best passage per facet

Notice what this does to your odds. You don't have to be the best page for the head term to get cited. You have to be the best passage for one of the sub-queries. A page with a genuinely strong section on agency invoicing can get pulled into that answer even if it would never rank top-three for "best project management tool". That's the opportunity. It's also why a thin page that only targets the head term gets nothing: it has no passage to offer for any of the facets the buyer actually cares about.

What changes versus traditional SEO.

Page-level ranking becomes passage-level retrieval. The unit that gets cited is a claim under a heading, not the whole URL.
One SERP to rank within becomes many retrieval entry points, one per fanned-out sub-query.
Link authority stops being the primary gate. Semantic relevance and source-worthiness do more of the work.

The playbook.

Six steps, in the order I'd run them. The first three are about the content on the page. The last three are about the signals around it. None of this is mechanical: deciding what counts as a non-commodity claim takes judgment, and I'll flag where.

1. Cover the fan-out, not just the head term

Take your target query and write down the sub-questions it decomposes into: the narrower specs, the adjacent questions, the "compared to what" and "for whom" facets. The fastest way to find them is to run the query in AI Mode yourself and read which facets the answer actually addresses, then look at the "people also ask" cluster and the related searches. Those are a rough map of how Google is already fanning the query out.

Then make sure your content answers each facet. A page on "best approach to X" should cover "what is X", "X vs Y", "X for B2B", and "what does X cost", either on the same page or across a tightly linked cluster. Each facet you genuinely answer is another retrieval entry point. Miss a facet and you simply can't be cited for it.

2. Write passage-first

Front-load a clear, self-contained claim under each heading. A passage that makes sense lifted out of the page is a passage that can be cited. A claim buried in the third sentence of a meandering paragraph is not. Put the answer in the first line of the section, then support it underneath.

A simple test: take any section of your page, copy the first two sentences, and paste them somewhere with no surrounding context. Do they still state something true and complete on their own? If yes, that's a citable passage. If they only make sense after the three paragraphs above them, the model has nothing clean to lift, and it'll reach for a competitor who wrote it tighter.

3. Make the claims non-commodity

This is the one that matters most and the one most teams skip. If the information on your page could be generated by the model from its training data without consulting you, it provides no grounding value and won't be cited. The model already knows the generic version. It only needs your page when your page has something the training data doesn't: original data, first-hand experience, a specific number from real work, a named methodology.

This is a judgment call, so be honest with yourself. "Internal linking helps SEO" is commodity: every model can generate that. "Across the 40 audits I ran last year, fixing orphaned pages moved crawl frequency before it moved rankings, usually within two weeks" is not. The second one is grounded in work only you did. Audit your key pages for at least one genuinely non-commodity claim each, and if a page hasn't got one, that's the gap to fill before anything else.

4. Structure for extraction

Clean H2 and H3 nesting, one idea per passage, attributable data points with their source named in the text. The model does its own chunking, so you're not chunking for it. You're making each passage unambiguous enough to be lifted without losing its meaning. Tables and clear lists help here too: a well-formed comparison table is about as extractable as content gets, because each row is a self-contained fact.

Name your sources inline rather than hiding them in a footnote. "Google's documentation states X" inside the passage travels with the claim when it's lifted; a superscript footnote marker does not. If the number is yours, say so in the sentence: "in our testing of 612 prompts" is more citable than "in testing".

5. Do the entity and schema groundwork

Organization, Person, and Article schema with consistent sameAs links so the model can resolve who you are and who wrote the claim. This is the entity layer, not the rich-result layer: it has no visible search feature attached, but it's what lets a model attribute a claim to a named, credentialed source rather than an anonymous page. It's the same work I describe in the entity coverage map template.

At minimum: an Organization node on your site with sameAs links to your real profiles, and an Article node on each post with a Person author whose sameAs points to a consistent LinkedIn or equivalent. Consistency is the whole game. If your name, role and description differ across your site, LinkedIn and Wikidata, you're fragmenting the entity rather than building it.

6. Earn source-worthiness

Be the primary source a model reaches for. That means original research, named expertise, and the kind of content other people cite. It's the slowest lever and the most durable one, because it's the hardest for a competitor to copy. The five source categories I use in every audit are the framework I run here.

You don't earn this in a sprint. But you can start: publish one piece of genuinely original data this quarter, attribute every post to a named expert, and get cited by one credible third party in your category. Each of those is a signal that you're a source, not just a page that exists. The compounding is slow and then sudden.

14%

of AI Mode citations come from pages ranking in the traditional top 10

passage

the unit AI Mode retrieves and cites, not the page

many

sub-queries one query fans out into, each a retrieval entry point

What doesn't move the needle.

There's a cottage industry of AI search tactics that Google's own documentation says do nothing. I've tested several of them and seen no effect, which matches what the docs say. Save your time:

llms.txt files. There's no evidence they affect AI search visibility. No major platform has confirmed reading them for ranking.
Manual content chunking for AI. The model chunks your content itself. Pre-chopping it into "AI-friendly" blocks isn't a signal.
AI-phrasing rewrites. Rewriting copy to sound more like an AI wrote it isn't a documented signal and tends to make the content worse for humans.
Inauthentic mention campaigns. Coordinated brand mentions seeded across low-quality or AI-generated content. This is the 2014 link-scheme playbook in a new hat, and it carries the same risk.

If a tactic only makes sense in a world where you're trying to trick the retrieval system rather than be genuinely useful to it, it belongs on this list.

The AI Mode citation checklist. Run this against any page you want cited.

Does the page answer the sub-questions the target query fans out into, not just the head term?
Does each section open with a clear, self-contained claim that survives being lifted out?
Is at least one claim genuinely non-commodity: your data, your experience, a number a model can't invent?
Is the structure clean enough (H2/H3, one idea per passage) for unambiguous extraction?
Is the entity layer in place: Organization, Person and Article schema with consistent sameAs?
Is the page a source other people would cite, or just a page that exists?

Where teams go wrong.

The playbook is simple to state and easy to get wrong in practice. These are the failure patterns I see most, across very different kinds of site. None of them are exotic. They're just the default way most pages get written.

Writing for the head term and stopping

The most common mistake. A page targets "best running shoes for flat feet" or "how to remortgage" or "small business accounting software", nails that one phrase, and ignores every facet around it. In a ranked list that can still work. In AI Mode it leaves most of the retrieval surface uncovered, so the page wins one sub-query at best and is absent from the rest of the answer. Breadth of genuine coverage beats a single perfectly optimised passage.

Confusing fluent with non-commodity

A page can be beautifully written and still have nothing a model needs. Fluency isn't grounding value. A recipe site that says "preheat your oven and use fresh ingredients" is fluent and worthless to a RAG system; a recipe site with tested timings, gram weights, and a note on what goes wrong at altitude has something only that kitchen produced. The same split applies to a law firm, a SaaS company, or a travel blog. Specific, tested, first-hand beats fluent and generic every time.

Treating schema as the whole job

Schema is necessary and nowhere near sufficient. I've seen teams add perfect Organization and Article markup to thin, commodity content and expect citations to follow. They don't. Schema helps a model attribute and resolve a claim; it does not make a weak claim worth citing. Do the entity groundwork, then spend the rest of your effort on the content the schema points at.

Chasing tricks instead of usefulness

Every few weeks a new "AI ranking hack" does the rounds. Most of them are the head-term-stuffing instinct in a new costume. If a tactic only makes sense in a world where you're trying to game the retrieval system rather than be genuinely the best source for a facet, it'll either do nothing now or get neutralised later. Build for usefulness and you never have to unwind it.

One honest caveat to close on. AI Mode is moving fast, and the retrieval behaviour you optimise for this quarter will shift. The reason this playbook holds up is that none of it is a trick. Cover the real intent, make genuinely useful claims, structure them clearly, and prove who you are. That's been the durable core of getting cited since long before AI Mode existed, and it's what survives the next change too. Pick one page that matters, run it through the checklist above, and you'll have done more than most of your competitors will this year.

NOTES

The 14% figure (AI Mode citations coming from pages in the traditional top 10) is from Google's public statements at Google I/O 2025 and subsequent query fan-out documentation. It's a directional signal about decoupling, not a fixed constant.
Query fan-out is documented in Google patent application US20240289407A1, which describes using large language models to generate multiple alternate queries from a single search.
This is a playbook built on Google's first-party AI optimisation documentation and my own testing, not a peer-reviewed study. It's one operator's working approach, made public so others can poke holes in it.

Frequently asked

Is AI Mode the same as AI Overviews?

No. AI Overviews is a single retrieval pass shown above the results. AI Mode is a full AI search experience built on multi-stage query fan-out. They share Google's index and quality baseline but optimise differently. I break the difference down in AI Overviews vs AI Mode.

Do my existing rankings help at all?

They help as a foundation, not as a guarantee. Being indexed, technically healthy, and snippet-eligible is the baseline AI Mode draws from. But ranking position itself is a weak predictor of citation: only 14% of cited URLs sit in the traditional top 10. Treat rankings as table stakes, not the goal.

How do I measure if it's working?

Citation share, not rankings: the proportion of relevant AI answers that cite you, measured across a query set. Run a structured prompt test across platforms, and track the number over time. I cover the method in citation share is the new keyword ranking.

Does schema matter for AI Mode?

The entity layer does: Organization, Person and Article schema with sameAs links help the model resolve who you are and attribute claims to you. The rich-result layer (FAQ dropdowns, review stars) doesn't drive AI Mode citation. Prioritise the entity work.

How to rank in Google AI Mode.