AI Overviews don't lift the page that ranks. They lift the page that answers. Most ranking-optimized content doesn't answer cleanly, which is why so many sites at position #2 or #5 get cited while the page at #1 gets ignored.
The mechanics of getting cited are different from the mechanics of ranking. Ranking rewards depth, authority, and link equity. Citation rewards a single declarative sentence, in the right place, in the right shape, that an AI extractor can lift verbatim without losing meaning.
This is the second spoke under What Google I/O 2026 Just Changed. After Citation is the new position one shows you why this matters, this is the how. Six mechanics, in priority order.
What "getting cited" actually means
When AI Overviews or AI Mode generates an answer, it does two things: it synthesizes a response from the pages it retrieved, and it shows source links to the pages it pulled from. The links are the citations. Being a citation means your URL appears in that source list and, more importantly, that the answer text references your content.
Per QuickSEO's data (cited in the Launchcodex analysis), the overlap between AI Overview citations and the top 10 organic results is only 17 to 54%. Translation: roughly half of cited pages are not the top-ranking pages. Citation is its own selection layer.
What makes a page selected? Per Google's official optimization guide published May 15, 2026, the answer is "same fundamentals as classic SEO." That's true and incomplete. The fundamentals plus six specific mechanics get you cited. Without the mechanics, you rank but don't get quoted. This is the core of AEO (answer engine optimization), the work of shaping content so an AI answer engine quotes it.
Where does the answer have to live on the page?
The single most important move. State your page's core claim in one declarative sentence, 25 words or fewer, in the first 100 words of the body. AI extractors lift cleanly from the opening. They give up on pages where the answer is buried.
"What is invisible prompting?" → "Invisible prompting is a UX design approach where the product interface itself gathers the context an AI model needs, so users never have to learn prompt engineering."
That sentence is 28 words. It defines the term, names the mechanism, identifies the user, and ends with a stable conclusion. If an AI pulls it verbatim, the source still makes sense. That's what "liftable" means.
> The lift test: read your opening sentence out of context. If it still makes sense and answers the page's question, an AI extractor can use it. If it requires the surrounding paragraphs to make sense, an extractor can't.
The most common failure I see on small-business sites is an opening paragraph that reads like a "welcome to our practice" introduction, with the actual answer buried in section three. Move the answer up and put the welcome below it.
What does a stable factual unit look like?
AI systems gravitate toward content with specific, attributable, citable facts. A page that says "we have decades of experience" gets summarized into nothing. A page that says "we've completed 4,200 cabinetry installations since 1998" gets quoted.
The pattern is specific numbers, specific dates, specific names, specific definitions. Not "we've been around a while" but "founded in 1998." Not "many clients" but "over 200 active clients across Ontario." Not "experienced lawyers" but "Sarah Patel, called to the bar in 2009, partner since 2018."
This is also why proprietary research, original case studies, and first-hand data get cited at disproportionate rates. The AI selects facts rather than inventing them, so give it something distinctive to select.
Why do AI extractors fail on pronouns?
On first reference in a section, use the full name instead of a pronoun: your company name, not "we"; the person's name, not "she"; the procedure's name, not "it"; the state's name, not "this state."
AI extractors do entity resolution, the work of figuring out which real thing each noun points to. Ambiguous references reduce citation eligibility because the extractor can't be sure what the sentence is about. Pronouns are particularly bad in the first 100 words.
This is unnatural prose if you overdo it. The trick is to use full names at the start of each major section, then revert to pronouns inside the section once entity context is established. Read your page top to bottom and ask: if an AI lifted any single paragraph in isolation, would it know exactly what each noun refers to?
Mechanic 4: Question-headed H2s
Pages where the H2s are noun phrases ("Pricing," "Services," "About") get cited less. Pages where the H2s are questions ("How much does an extraction cost?", "What services do we offer?") get cited more. AI systems use H2s to identify the question each section answers.
Listicles and definitional pages are the worst offenders. "Top 5 X" reads like a noun. "What are the top 5 X for Y?" reads like a query. The second version both ranks better in AI Mode and gets summarized better in AI Overviews because it matches the form of the query.
For lawyers and accountants especially, where most queries are phrased as questions ("How do I file a small claims appeal in Ontario?", "What's the tax treatment of a side business under $30,000?"), every section heading should be the exact form of the question your prospect would actually type or speak into AI Mode.
Mechanic 5: Schema markup matches the visible claim
If your page's visible claim is "Sarah Patel is a partner at Patel Halat LLP since 2018," your Person schema should say the same thing. If your prose says "We start at $4,200 for a standard installation," your Offer schema's price field should say 4200. AI systems cross-check structured data against visible content. Mismatches reduce trust signals.
This is the operational meaning of Google's repeated guidance that "structured data must match visible page content." It is also the cleanest reason to keep schema usage focused. Sprawling schema with claims that don't appear on-page does more harm than good. The pages that get cited use schema sparingly and accurately, with each field corresponding to a visible sentence.
We covered the structured-data types that still earn rich results in Every structured data type Google still rewards in 2026. Use that list, skip the deprecated types, and make sure every property mirrors a visible claim.
Mechanic 6: Visible last-modified date
Display the date the page was last updated, in human-readable text, in the byline area. Don't bury it only in dateModified (the machine-readable freshness field inside your JSON-LD, the structured-data block Google reads). Put it where the reader can actually see it.
AI systems prefer fresh sources. "Last updated: November 15, 2025" near the byline signals freshness to both the user and the extractor. Pages without a visible date get treated as older than they are. Pages with a visible date that hasn't been updated in years get treated as stale.
If you update your page when facts change, this is the highest-impact half-hour you can spend on a site. Add the date, keep it accurate, and update it when something material changes.
The unspoken seventh mechanic: brand presence off-page
The six mechanics above are on-page. The seventh is off-page, and it might be the most important: AI systems weight entity mentions and co-occurrence across the web. Sites with strong brand presence in news, podcasts, forums, and reviews get cited more frequently than otherwise-equivalent sites without that presence.
Per Amsive's research, branded-search CTR rose +18% in the AI Overview era while generic CTR collapsed. The mechanism is the same: when an AI system recognizes a brand as a real entity, that brand becomes a higher-priority citation candidate. The off-page work (PR mentions, podcast appearances, review presence, citations in reputable directories) feeds the same signal.
You can't optimize this on the page; you build it over years, which is the reason to start now.
The proof: what this series does
I write every post in this series with the six mechanics built in. Look at the post you just read.
The first sentence is a single declarative answer, 25 words or fewer. The TLDR is a liftable summary. The H2s are mostly question form or noun-phrase answers. The factual claims have specific numbers and named sources. The structured data (which you don't see in the rendered content but is in the page's schema) matches the visible prose. The dateModified is visible. Every external citation is a live URL to the actual source.
This is what citation engineering looks like in practice. The same six moves work for a dentist's homepage, a lawyer's "what to expect" page, an accountant's "tax services" page, or a contractor's project portfolio. The mechanics aren't vertical-specific, they're extraction-specific.
For lawyers and accountants, verticals where most queries are now phrased as natural-language questions and where citation inside an AI answer drives disproportionate commercial value, these mechanics are the most valuable SEO work you can do this quarter.
Common objections
"This is just good writing." Partly. Good writing has clear answers, names entities, uses specific facts, structures with clear headings. Where it differs: most "good writing" doesn't pass the lift test, because it relies on narrative flow that doesn't survive isolation. Citation engineering is good writing plus the assumption that any single paragraph might be extracted from context.
"Google says I don't need to do anything special." The Google optimization guide says the fundamentals still apply, no separate AEO playbook is needed, and to avoid hacks like chunking and llms.txt. That's all true, but "fundamentals" and "citation-ready execution of fundamentals" are different things. The fundamentals get you indexed, and the mechanics get you cited.
"What about FAQ schema?" Per the May 2025 deprecation, FAQ rich results are no longer earned by most domains. The FAQs are still useful for AI extraction (they read like Q&A and the structure matches AI Mode interaction patterns), so we keep them in the markup. They just don't show up as rich results in the blue-links SERP (the traditional search results page) anymore.
What to do this week
Pick your three highest-value pages, the ones that drive booked calls, signed contracts, or completed checkouts. For each one:
- Find the single-sentence answer to the page's main question. If it's not in the first 100 words, move it.
- Audit the H2s. Rewrite any noun-phrase headings as questions the customer would ask.
- Find every "we," "us," "this," and "it" in the first 100 words and replace each with an explicit name.
- Add or update the visible last-modified date.
- Check that any structured data on the page matches the visible prose.
- Identify two or three stable factual units (numbers, dates, named people) and make sure they're stated cleanly.
That's the operational checklist. Three pages, six steps each. A focused afternoon. The pages that come out the other side of that audit are the pages that get cited.
Sources
- Optimizing for generative AI search (Google Search Central, May 2026)
- A new resource for optimizing for generative AI in Search (John Mueller, Google)
- Top ways to ensure your content performs well in Google's AI experiences (Google Search Central, May 2025)
- Using generative AI content on your website (Google Search Central)
- Google I/O 2026 SEO update (Tanner Medina, Launchcodex)
- AI Overviews impact on publishers (Matt G. Southern, Search Engine Journal)