Key takeaways
- Google AI Overviews retrieve passages, not full pages -- so paragraph-level structure matters more than overall page rank.
- Query fan-out means Google breaks one search into 5-15 sub-queries. A hub-and-spoke content architecture is now the baseline, not a nice-to-have.
- Citations almost always come from the top 20 organic results, so traditional ranking signals still matter -- they're just the floor, not the ceiling.
- Answer-first writing (a direct 40-80 word answer in the opening paragraph) is the single highest-leverage on-page change most sites can make today.
- Tracking which pages get cited, and by which AI models, is now as important as tracking keyword rankings.
Getting cited in a Google AI Overview is the new #1. That's not hyperbole -- for a growing share of informational and commercial queries, the AI answer box is the entire above-the-fold experience. If you're not in it, you're effectively invisible, regardless of where your blue link sits.
The frustrating part is that most SEO advice hasn't caught up. A lot of what gets shared is either too vague ("write helpful content!") or still anchored in a pre-AI mental model. This guide is an attempt to be specific: here's what actually changes at the on-page level, and here's what to do about it.
How AI Overviews actually work (the short version)
Before touching a single heading tag, it helps to understand the mechanism. When a query triggers AI Mode, Google doesn't just pull the top 10 results. It performs what's called query fan-out: it generates 5-15 related sub-queries, retrieves relevant passages from across its index, then synthesizes an answer with inline citations.
The key word there is passages. Google's retrieval unit is a well-structured paragraph or section, not a full page. This is why a perfectly-written paragraph buried on a page that ranks 15th can get cited while your homepage doesn't. It also explains why a single long page that covers everything in one dense block tends to underperform compared to a tightly structured page with clear sub-sections.
This passage-retrieval model has two practical implications that shape everything else in this guide:
- Every major sub-question on your page needs its own clearly labeled section with a direct answer immediately below the heading.
- Your content architecture needs to match the fan-out. One pillar page won't cover all the sub-queries Google generates -- you need supporting pages that answer the predictable follow-ons.
The foundation: you still need to rank organically first
This is worth stating plainly because some newer GEO content glosses over it. Citations in AI Overviews almost always come from pages already in the top 20 organic results for the head query. If your page isn't ranking, it's not getting cited.
So the starting point is solid traditional on-page SEO:
- Strong topical relevance and keyword targeting
- Internal links from related cluster pages with descriptive anchor text
- External citations to authoritative sources (.gov, .edu, peer-reviewed research, Wikipedia)
- Fast Core Web Vitals -- LCP under 2.5s on mobile is the threshold that matters
- No aggressive interstitials or pop-ups above the fold (Google's quality model penalizes these)
- Mobile-first layout, since AI Overviews appear first on mobile
None of this is new. But it's the prerequisite. Get this right, then layer in the AI-specific optimizations below.
On-page structure for passage retrieval
Write the answer first
The single highest-leverage change most pages can make is moving the direct answer to the top. Google's passage retrieval favors the opening paragraph when it contains a clear, declarative answer to the query.
The target is 40-80 words that directly answer the question in the first paragraph, before any context-setting or preamble. This is the opposite of how most content is written, where the answer gets buried after background information.
A practical test: paste your opening paragraph into a blank doc and ask yourself whether someone could stop reading there and have their question answered. If the answer is no, rewrite it.
Match headings to sub-questions
Each H2 should map to a predictable sub-question someone might ask about the topic. The heading should read like the question (or a close paraphrase), and the first 2-3 sentences below it should answer it directly.
This structure serves two purposes. It makes your page easier for Google's passage retrieval to parse. It also means your page naturally covers the fan-out sub-queries that AI Mode generates -- which is how you get cited across multiple parts of an AI Overview, not just one.
Use descriptive, declarative H1s
Your H1 should match the head query as closely as possible. "How to rank in Google AI Overviews" beats "The ultimate guide to AI search" every time. Declarative and specific wins over clever and vague.
Keep paragraphs tight
Long, meandering paragraphs are hard for passage retrieval to work with. Aim for 3-5 sentences per paragraph, each one advancing a single idea. If a paragraph is doing two things, split it.
Schema markup that actually matters
Schema is one of those topics where the advice often overshoots. You don't need every schema type -- you need the right ones, implemented correctly.
For most content targeting AI Overviews, the relevant types are:
- Article schema with
publishedDateandupdatedDatein JSON-LD (and visible in the HTML -- Google cross-references both) - FAQPage schema for pages with a Q&A section -- this directly maps to how AI Overviews synthesize answers
- HowTo schema for step-by-step content
- Author schema with a credible bio that supports E-E-A-T signals
The updatedDate point is worth emphasizing. AI Overviews strongly favor recently updated content for time-sensitive queries. If you have pages with stale updatedDate values that you've actually refreshed, fix the schema. Google uses it.
E-E-A-T signals at the page level
Experience, Expertise, Authoritativeness, and Trustworthiness aren't just site-level signals -- they operate at the page level too. For AI Overview citation, a few specific signals seem to carry disproportionate weight:
Original data and named entities. Pages that cite original research, proprietary data, or specific named sources get cited more often than pages that summarize what others have said. If you have first-party data -- survey results, internal benchmarks, customer statistics -- surface it explicitly on the page.
Author bylines with real credentials. A byline linked to an author bio page with verifiable credentials (publications, professional history, relevant expertise) helps. Thin or anonymous authorship is a signal in the wrong direction.
External citations to authoritative sources. Linking out to peer-reviewed research, government data, or well-known reference sources signals that your content is grounded in something real. This is counterintuitive for SEOs trained to keep users on-page, but it's the right call for AI citation.
Query fan-out and content architecture
This is where most sites have the biggest gap. A single well-optimized page can get you cited for the head query. But AI Overviews often synthesize answers from multiple sources across multiple sub-queries. To win more of that surface area, you need content that covers the fan-out.
The practical approach: before writing or optimizing a piece, map the likely sub-queries. Google's "People Also Ask" boxes and the related searches in AI Mode itself are the fastest way to do this. You're looking for the 8-15 questions that someone asking the head query would predictably follow up with.
Each of those sub-questions should either have its own section on the pillar page (if it's a brief answer) or its own dedicated page (if it warrants depth). The pillar page links to the sub-pages; the sub-pages link back. This hub-and-spoke architecture is what lets a single topic cluster dominate multiple parts of an AI Overview.
Tools like Promptwatch can help map this systematically -- the Answer Gap Analysis feature shows exactly which prompts competitors are getting cited for that you're not, which is a fast way to find the sub-queries your content architecture is missing.
What kills AI Overview citation
A few things reliably prevent citation even when the content is good:
Aggressive interstitials. Pop-ups, cookie consent overlays that block content, and newsletter modals above the fold all hurt. Google's quality model treats these as signals of a poor user experience, and they reduce citation eligibility.
Thin or duplicate content. Pages that largely restate what other pages already say don't get cited. AI Overviews favor pages that add something -- a unique angle, original data, a more complete answer.
Stale content without updated dates. For queries where recency matters, an outdated updatedDate is a disqualifier. If you've refreshed a page, make sure the schema and visible date reflect it.
No passage-level structure. A page that's one long block of text with no sub-headings gives passage retrieval nothing to work with. Even if the content is excellent, it's hard to extract a clean citation from it.
Tracking your AI Overview visibility
Optimizing for AI Overviews without tracking them is like running a PPC campaign without conversion data. You need to know which pages are being cited, how often, and for which queries -- otherwise you're guessing at what's working.
A few tools worth knowing about:
Promptwatch tracks AI Overview citations alongside 9 other AI models, with page-level visibility data and traffic attribution. The crawler log feature is particularly useful here -- it shows when Google's AI crawler hits your pages, which is a leading indicator of citation activity.
Semrush has added AI Overview tracking to its core platform, which is convenient if you're already using it for traditional rank tracking.
Surfer SEO has built out AI search optimization features alongside its content scoring, making it useful for both writing and tracking in one workflow.
SE Ranking's AI visibility toolkit covers Google AI Overviews alongside other AI search engines, with keyword-level citation data.
Here's a quick comparison of what to look for in an AI Overview tracking tool:
| Tool | AI Overview tracking | Page-level citations | Content gap analysis | Traffic attribution |
|---|---|---|---|---|
| Promptwatch | Yes (+ 9 other AI models) | Yes | Yes | Yes |
| Semrush | Yes | Limited | No | No |
| Surfer SEO | Yes | Yes | Partial | No |
| SE Ranking | Yes | Yes | No | No |
The traffic attribution column matters more than it might seem. Knowing you're being cited is useful; knowing that citations are driving actual visits and conversions is what justifies the optimization investment.
A practical prioritization framework
If you're looking at a site with dozens of pages and limited time, here's how to prioritize:
-
Find pages already in the top 20 for head queries. These are your citation candidates. Pages outside the top 20 need traditional SEO work first.
-
Check for answer-first structure. Pull the opening paragraph of each candidate page. Does it directly answer the query in 40-80 words? If not, that's your first edit.
-
Audit heading structure. Do the H2s map to predictable sub-questions? Is there a direct answer in the first 2-3 sentences below each one?
-
Check schema. Is Article schema present with accurate
publishedDateandupdatedDate? Is FAQPage schema implemented where there's a Q&A section? -
Look for original data or named entities. Does the page cite anything specific -- a study, a statistic, a named expert? If not, find something to add.
-
Map the fan-out. What are the 5-10 sub-questions someone asking this query would follow up with? Are they covered on this page or on linked sub-pages?
This sequence lets you make meaningful improvements to existing content without starting from scratch. In most cases, the answer-first rewrite and heading restructure alone move the needle.
The content types that get cited most
Not all content formats are equally likely to get cited. Based on how AI Overviews synthesize answers, a few formats consistently outperform:
Definition and explanation pages. "What is X" and "How does X work" queries are heavily represented in AI Overviews. Pages that answer these clearly and completely, with good sub-structure, get cited often.
Comparison pages. "X vs Y" and "best X for Y" queries generate AI Overviews that cite multiple sources. Being one of three cited sources is more achievable than being the single top organic result.
Step-by-step guides. HowTo schema maps directly to how AI Overviews present procedural content. A well-structured guide with numbered steps and HowTo schema is a strong citation candidate.
Pages with original statistics. If your page is the primary source for a specific data point, you'll get cited every time that data point is relevant to a query. Publishing original research -- even small-scale surveys -- pays disproportionate dividends here.
The image signal people overlook
AI Overviews frequently include image carousels, and the images come from cited pages. An image with descriptive, keyword-relevant alt text on your page increases the chance that both the text and the image get surfaced.
This isn't a major ranking factor, but it's a free win that most pages miss. Every image on a page targeting AI Overview citation should have alt text that describes what the image shows in the context of the query -- not just a filename or a generic label.
Putting it together
The shift from traditional SEO to AI Overview optimization isn't a complete rebuild -- it's a layer on top of what already works. Pages that rank organically, have clear structure, demonstrate real expertise, and answer questions directly are the ones getting cited.
What's new is the precision required. Passage retrieval rewards paragraph-level clarity. Query fan-out rewards content architecture. Citation eligibility rewards original data and E-E-A-T signals at the page level, not just the domain level.
The sites winning in AI Overviews right now aren't necessarily the ones with the highest domain authority. They're the ones that made it easy for Google to extract a clean, credible answer from a specific passage. That's an on-page problem, and it's one you can solve.