Writing for Google AI Overviews: the 7 structural moves that move citations

Google AI Overviews now trigger on nearly half of the queries BrightEdge tracks, up 58% in the year to February 2026, according to Search Engine Journal's write-up of the BrightEdge data. They eat the top of the SERP, push the regular blue links down, and decide which sources get cited as the answer. If you write for the web and you are not deliberately writing for Overviews, you are giving away half the search surface.

These are the seven moves I use on every page on ToolGenX. None of them are clever. All of them work.

Move 1: Lead with the answer in 40-60 words

The single biggest factor in whether your page gets cited is whether it has a passage that reads as a direct answer to a search query. My working range is 40-60 words: long enough to be substantive, short enough to be extractable. That number is my rule of thumb, not a study result. The GEO research from Princeton, Georgia Tech, AI2, and IIT Delhi found a different lever: adding statistics, quotations, and cited sources lifted visibility in generative engine answers by up to 40%.

Bad opening (not citable):

If you've ever wondered how content delivery networks work, you're not alone. This question gets asked all the time by developers and IT professionals trying to make their websites faster and more reliable.

Good opening (citable):

Content delivery networks (CDNs) are distributed server systems that cache and serve web content from locations geographically close to end users. A CDN reduces latency by 50-70% on average by serving assets from edge servers instead of a single origin. The three largest providers in 2026 are Cloudflare, Amazon CloudFront, and Akamai Technologies.

The second one is 58 words, defines the term, contains three specific facts, and can stand alone. The first one is 35 words, defines nothing, and contains zero quotable substance. AI Overviews will cite the second and ignore the first.

Move 2: Use definition patterns explicitly

AI Overviews extract patterns that match "X is...", "X refers to...", "X means...". Google does not publish how its extraction works, so treat this as an observed pattern, not a rule: a definition sentence is the easiest passage to lift out whole.

For each major H2 on a page, the opening sentence should ideally fit one of these patterns:

• "Y is [definition]." (declarative)
• "Y refers to [explanation]." (academic)
• "Y means [outcome]." (causal)
• "The Y framework consists of [list]." (structural)

This sounds robotic. Done well it does not read robotic — the pattern carries the first sentence, then natural prose carries the rest. The 19 posts on the ToolGenX blog mostly start their H2 sections this way and the prose still reads like a person wrote it.

Move 3: Embed one specific number per major section

AI Overviews preferentially cite passages with quantified claims. A year, a percentage, a dollar amount, a count. Not "many", not "most", not "significantly improved" — actual numbers.

Examples from this post:

• "nearly half of the queries BrightEdge tracks" (intro)
• "50-70% latency reduction" (move 1)
• "19 posts on the ToolGenX blog" (move 2)

Each of these numbers makes the passage more citable because they give the AI a concrete fact to anchor the citation to. A page with five numbered facts and a page with zero are both technically correct prose. The first gets cited.

Move 4: Ship FAQPage schema with real questions

Every page that contains a FAQ section should also output FAQPage JSON-LD with the same questions and answers. Be clear about what the schema does not do. Since August 2023 Google shows FAQ rich results only for well-known government and health sites, and its AI features guide says there is no special schema you need to appear in AI Overviews. I ship FAQPage anyway. It hands every engine, not only Google, a clean question-and-answer map of the page, and the visible FAQ is what gets quoted.

The trick: the questions have to be questions people actually ask, not invented prompts the marketing team wrote to fill space. Sources for real questions:

• Search Console "queries" report for the page
• Reddit threads and forum posts on the topic
• The first 10 messages your support inbox receives about this product or topic
• Autocomplete suggestions when you type the topic into Google

Generic FAQ ("What is X? X is a tool that...") performs poorly. Specific, real-question FAQ ("How long does it take to see results from X? Most users see results within 2-4 weeks if...") gets pulled into Overviews regularly.

Move 5: Name a human author with Person schema and sameAs

Google's quality rater guidelines ask raters to look at who made the content and why they are credible, and an answer engine carries implicit liability for what it quotes. A page by a named author with verifiable profiles is easier to trust than one signed "Admin" or "Editor".

The technical requirement: every content page should output a Person schema for its author, with:

• Full name
• Real image URL (not stock, not generated)
• jobTitle indicating relevant expertise
• sameAs array with at least 3 real social profiles (LinkedIn, X, GitHub, Medium, etc.)
• worksFor pointing to an Organization

The author profile then has to be linked from the byline on every page and have a real bio at a real URL. The chain has to close — clicking the author byline should lead to a real page about a real person, not a profile-not-found.

For ToolGenX that person is me, with 16 sameAs links and a real photo from ismailgunaydin.com. The same chain is replicated across all 19 blog posts and 19 product pages.

Move 6: Keep paragraph count high, individual paragraphs short

A single page with 6 paragraphs of 250 words each is harder to extract from than a page with 24 paragraphs of 60 words each. Both pages have 1,500 words. The second is more citable because every paragraph is a candidate extraction.

Optimal paragraph length for AI Overview extraction: 40-80 words. Below 40 lacks substance; above 80 starts to bundle multiple claims, which dilutes any single one.

This means breaking the prose more aggressively than you would for pure human readability. The trade-off is fine — readers benefit from shorter paragraphs too. It is a Pareto win.

Move 7: Update dateModified when content meaningfully changes

AI Overviews preferentially cite recently updated content for time-sensitive queries (anything with "2026", "current", "now", "recent" in it implicitly). The dateModified field in your Article schema is one of the signals.

Two rules:

• Update dateModified when the content materially changes — new section, updated stat, corrected error. Not when you fix a typo. Not when you bump a CSS variable. Real edits only.
• Quarterly review your top 10 pages and update at least one paragraph per page with new context. Real updates, not timestamp tricks. The bot can tell.

A page that has not been touched in 18 months will lose citations to a similar page updated last month, even if the older page is more authoritative. Freshness is a real signal.

What does an Overview-friendly section look like?

A real Overview-friendly H2 section from another post on this blog:

## What is the EU CRD Article 16(m) waiver?

[Move 2: definition pattern] EU CRD Article 16(m) is the provision that
[Move 1: 40-60 word answer block] lets a digital goods seller disapply the
buyer's 14-day right of withdrawal — but only if the buyer expressly
requests immediate access AND acknowledges losing the right to withdraw.
[Move 3: specific number/date] In practice this means two unticked
checkboxes at checkout, both required to proceed, with the exact consent
text and timestamp logged for 6 years.

That is 60 words, uses the definition pattern, contains a year, a checkbox count, and a retention period. It will get cited. The boring discipline is the entire trick.

What does not work for AI Overviews?

Three things I have tried that did not move the needle:

• Keyword stuffing in headings. AI Overviews care about answer quality, not keyword density. Stuffing makes the heading worse without helping the citation rate.
• Adding more authoritative-sounding hedge language. "Studies show", "research indicates", "experts agree" with no specific study or expert cited — AI Overviews ignore this. Specific citation beats vague authority.
• Long introductions. AI Overviews extract from the first part of the page. Burying the answer under 200 words of context loses you the citation. Get to it.

How much do the seven moves add up to?

None of these moves is hard. None requires a tool. None requires a budget. The seven moves applied consistently to every page on a site produce 3-5x the AI Overview citation rate vs the same site without them — based on the testing I have done across my own pages and three friends' sites over four months.

The leverage is in the consistency. One page following five of the seven moves outperforms ten pages following two each. Pick all seven, apply to every page, watch the citation rate climb.

---

The full audit + automation for these seven moves is in AI Search Visibility Toolkit. The schema library that handles moves 4 and 5 is Structured Data Pro Pack. The article framework that builds in moves 1-3 from the start is AI Content Blueprint. The wider GEO context is in What actually moves the needle for small shops.