Key takeaways
- AI Overviews now appear in roughly a quarter of US searches and about half of informational queries, and being cited inside one earns measurably more clicks than a traditional ranking alone.
- The citation mix shifted in 2026: product pages overtook listicles as the most cited format in AI Overviews in late July, so commercial pages now need the same extractability rigor as editorial content.
- Roughly 55% of AI Overview citations pull from the top 30% of a page, so front-loading direct, standalone answers is a structural requirement, not a style preference.
- Technical prerequisites matter more than most checklists admit: AI crawler access in robots.txt, server-rendered HTML, and Core Web Vitals all correlate with citation rates.
- Google killed FAQ rich results in May 2026, so the old "add FAQ schema" advice is dead, even though FAQ-style content structure still helps extractability.
Google's AI Overviews have quietly become one of the most contested surfaces in search. Conductor's benchmarks put AI Overviews on 25.11% of searches in Q1 2026, nearly double the 13.14% from March 2025, and post-update analyses suggest the May 6, 2026 update pushed that higher still. The stakes are asymmetric: Seer Interactive found AI Overviews cut organic CTR by 61% on affected queries where a page is not cited, while getting cited earns 35% more clicks than a traditional ranking alone.
That asymmetry is why "how do I get cited in AI Overviews" has become the most common question I hear from SEO teams. The honest answer is that a lot of it happens off-page (authority, third-party mentions, Reddit and forum presence). But there is a meaningful set of on-page fixes that determine whether your content is even eligible for citation, and most sites I look at are failing at least five of the items below.
This checklist is built on current citation behavior, not 2025-era advice. Two things changed recently that invalidate older playbooks: Google removed FAQ rich results entirely on May 7, 2026, and product pages overtook listicles as the most cited content format in AI Overviews in late July 2026.
What the citation data says (and why older advice is stale)
Before the fixes, it helps to understand what AI Overviews actually cite, because the mix is not static.
Promptwatch's July 2026 citation-type data shows the full picture for AI Overviews: listicles averaged 18.0% of citations for the month but lost the top spot in the final days, with product pages ending July at 17.9% versus 16.2% for listicles. How-tos sat at 15.1%, news articles at 13.5%, and video climbed to roughly 6.3% by late July, up from about 2.7% in January, one of the fastest-growing formats.
Two implications. First, commercial pages now compete for citations directly, which means product pages need structured specs, transparent pricing, and current availability information, not just marketing copy. Second, video is a genuinely underexploited surface; most brands have no video assets competing for that ~6% share.
The sources-per-response data matters too. Google AI Overviews cite roughly 10 sources per answer, about double ChatGPT's ~5, which makes AI Overviews the more forgiving engine for mid-authority domains. There are more slots to win.
One more structural point: AI Overviews run fan-out sub-queries behind the scenes, meaning a single user prompt can trigger citations from multiple pages on the same site. Topical depth creates more citation surface area. A site with one thin page per subtopic is leaving slots on the table.
The 15 on-page fixes
1. Front-load the direct answer in the top 30% of the page
AirOps' analysis of 2,400 AI Overview citations found 55% of citations pull from the top 30% of a page's content. That is a structural finding, not a writing-style preference. If your page opens with a 300-word narrative intro that "sets context," the extractable passage the AI needs is buried below the fold of its retrieval window.
The fix: answer the query in the first two paragraphs, in a form that could be lifted verbatim into an answer. Then support, nuance, and expand below. Definition-first openings, direct numeric answers, and short standalone paragraphs all work.
2. Write for claim density, not narrative flow
The core selection pattern, per Silktide's analysis of what survives the AI Overview filter: AI systems favor passages with high claim density, meaning sentences or short paragraphs containing a complete, standalone, verifiable assertion (a number, a behavior, a timeframe) that makes sense without surrounding context. Long argument-building paragraphs have low claim density and get passed over.
Practically: break long paragraphs into shorter ones. Make each one carry a complete thought. "Our API returns results in under 200ms for 99.9% of requests" is a citable claim. "We believe speed matters, which is why our team has always focused on building fast infrastructure" is not.
3. Audit robots.txt for AI crawler access (the most common silent failure)
This is the fix I'd put first if ranking by frequency of the problem. Sites configured years ago with a wildcard User-agent: * Disallow rule, or that only ever explicitly allowlisted Googlebot and Bingbot, are often unknowingly blocking the AI crawlers that feed retrieval.
Check for these user-agents specifically:
OAI-SearchBot(OpenAI's search index, drives ChatGPT Search citations)Google-Extended(Gemini training and grounding; note this does not affect Google Search ranking)PerplexityBotandPerplexity-UserClaudeBotandanthropic-aiCCBot(Common Crawl, feeds many models' training data)
A nuance most guides get wrong: blocking GPTBot (training) does not remove a page from ChatGPT Search citations, which are governed by OAI-SearchBot and ChatGPT-User. If you want AI search visibility but not to feed model training, you need granular per-bot rules, not a blanket allow or disallow. And remember that robots.txt rules override anything you put in llms.txt; llms.txt is a navigation aid, not an access control.
4. Make sure key content exists in server-rendered HTML, not client-side JavaScript
Many AI crawlers, including OpenAI's, do not reliably render JavaScript. If your product specs, pricing, or FAQ answers are injected client-side by a JS framework, they may be invisible to the systems deciding what to cite. This is a distinct fix from traditional Core Web Vitals work: the question is not "is the page fast" but "is the citable content present in the raw HTML response."
Server-side rendering, static generation, or at minimum ensuring critical text content is in the initial HTML response closes this gap. Check by fetching the page with JavaScript disabled, or by viewing the page source (not the rendered DOM) and confirming the content is there.
5. Add structured data, but stop chasing a magic schema type
Google's Search team has confirmed structured data helps AI Overviews understand, verify, and cite content, and Microsoft has said the same about Copilot. But there is no single schema type that triggers citation. What helps is marking up what is actually on the page: Product, HowTo, FAQPage (for structure, not rich results), Article, Organization, Person.
One dated tactic to drop: Google removed FAQ rich results from Search entirely on May 7, 2026. Adding FAQPage schema no longer earns a visual enhancement. FAQ-style content structure still helps extractability and claim density, so keep the format, lose the expectation of a rich result.
6. Put visible publication and last-updated dates on the page
This one is small and consistently skipped. AI systems weigh content currency, and a visible, honest last-updated date paired with a credentialed author signals both. Practitioner analyses suggest content freshness correlates strongly with citation retention, with one agency reporting pages not updated at least quarterly are 3x more likely to lose citations.
The operative word is honest. Changing the displayed date without a real content delta ("lastmod gaming") is detected and discounted over time. If the date says September 2026, the content should reflect something that happened or was revised in September 2026.
7. Add named, credentialed authors with visible expertise
The December 2025 core update reportedly extended E-E-A-T expectations beyond YMYL topics to all content categories, and a March 2026 update amplified the "Experience" pillar specifically. Generic, ghost-written, fact-aggregator content is the loser in this environment. Lily Ray's read on the May 2026 "Expert Advice" block is that first-hand experience stopped being just a trust factor and became a direct click driver.
The fix: named authors, author pages with real credentials and relevant experience, and content that includes first-hand observation ("we tested this across 40 client sites and found...") rather than aggregated secondhand claims.
8. Restructure product and commercial pages for extractability
This is the 2026-specific fix most checklists miss, because until this year product pages were an afterthought in citation strategies. With product pages now the top-cited format in AI Overviews as of late July 2026, commercial pages need the same rigor as editorial: structured specification tables, transparent pricing, current availability, and short standalone answers to the questions buyers actually ask ("how much does X cost," "what's the difference between X and Y").
If your product page is a hero image, three benefit bullets, and a gated demo form, it has almost nothing for an AI system to extract.
9. Build topical depth to exploit fan-out sub-queries
AI Overviews run fan-out sub-queries behind the scenes, pulling from subtopic sources to assemble one answer. A single prompt can trigger citations from multiple pages on the same site, but only if those pages exist and each one answers a specific sub-question well.
The fix: map the subtopics of your core queries and make sure you have a focused page for each, rather than one broad page that covers everything shallowly. Focused, single-question pages beat broad pages in retrieval, especially in engines with fewer citation slots.
10. Fix Core Web Vitals, with honest expectations
A May 2026 study of 74 domains found 56.8% of pages cited by AI Overviews pass Core Web Vitals on mobile, versus roughly 40% for the web overall. Cited pages beat the web average on LCP (74.3% good versus ~66% web-wide) and roughly match it on INP and CLS. The study's own authors are careful to note this is correlated, not proven causal: a fast page doesn't earn the citation on its own.
Treat CWV as a gate, not a signal. A slow page probably doesn't lose the citation directly, but it correlates with a bundle of other quality factors, and a March 2026 core update reportedly increased CWV weight in ranking generally, with the INP "good" threshold tightening to under 150ms.
11. Add FAQ-format content structure (without expecting a rich result)
Even though FAQ rich results are dead, FAQ-style structure remains one of the highest claim-density formats available. Question-and-answer blocks naturally produce standalone, verifiable assertions that lift cleanly out of context.
The fix: identify the actual questions your audience asks (from support tickets, sales calls, People Also Ask, and AI referral prompts in your analytics), then answer each in one to three sentences directly under the question. Keep answers self-contained.
12. Create video assets for queries where video is cited
Video citation share in AI Overviews climbed from ~2.7% in January 2026 to ~6.3% by late July, one of the fastest-growing citation formats, and most brands have nothing competing for it. YouTube alone reached 4.16% of all AI Overview citations in June 2026, and YouTube plus google.com together account for roughly 6.5% of citations, a self-referential concentration Google is compounding month over month.
The fix: for how-to and demonstration queries in your niche, publish short, well-titled videos with transcripts on the page. The transcript matters as much as the video, since it gives the AI system text to extract.
13. Update content substantively, on a schedule
Freshness matters, but cosmetic updates don't. The pattern to avoid: republishing with a new date and no meaningful content change. AI engines discount this over time. The pattern that works: quarterly reviews that add new data, update numbers, remove outdated claims, and reflect changes in the field.
For pages competing for citations on fast-moving topics, build a review calendar. For evergreen pages, a visible "last reviewed" date with a real review behind it is enough.
14. Match content format to what the query type actually cites
Promptwatch's citation-type data shows different query types pull different formats. How-tos dominate instructional queries, news articles dominate time-sensitive ones, comparisons have their own share, and product pages dominate commercial ones. If you're writing a listicle for a query where AI Overviews cite how-tos and product pages, you're mismatched against the retrieval pattern.
Before creating or reworking a page targeting a citation, check what the current AI Overview for that query cites. If it cites how-tos, structure yours as a numbered process. If it cites product pages, get the specs and pricing structured. Format fit is a quiet ranking factor.
15. Monitor which pages earn and lose citations, then iterate
The last fix is a process fix. Citation patterns shift month to month, as the listicle-to-product-page flip in July 2026 demonstrated. A page that earned citations in March may have silently lost them by September without anyone noticing, because traditional rank trackers don't show AI Overview citations.
You need visibility tracking that shows which of your pages are cited, for which prompts, and whether that's trending up or down. Promptwatch tracks exactly this, with citation trends classified by content type and per-page citation rates, so you can see whether the fixes above actually moved the needle.
Quick-reference table
| # | Fix | Effort | Impact timeframe | What it addresses |
|---|---|---|---|---|
| 1 | Front-load answers in top 30% | Low | Weeks | Extractability |
| 2 | Increase claim density | Medium | Weeks | Passage selection |
| 3 | Audit robots.txt for AI bots | Low | Days | Crawl access |
| 4 | Server-render key content | Medium | Weeks | Crawler visibility |
| 5 | Add structured data | Medium | Weeks | Understanding/verification |
| 6 | Visible dates | Low | Days | Freshness signal |
| 7 | Credentialed authors | Medium | Months | E-E-A-T |
| 8 | Restructure product pages | High | Months | Commercial citation share |
| 9 | Build topical depth | High | Months | Fan-out coverage |
| 10 | Fix Core Web Vitals | Medium | Months | Quality gate |
| 11 | FAQ-format structure | Low | Weeks | Claim density |
| 12 | Video assets with transcripts | High | Months | Growing video share |
| 13 | Substantive update schedule | Medium | Ongoing | Freshness retention |
| 14 | Match format to query type | Low | Weeks | Retrieval fit |
| 15 | Citation monitoring | Low | Ongoing | Measurement |
What not to bother with
A few things I'd explicitly deprioritize, because they consume effort without evidence:
- Chasing a specific schema type as a citation trigger. There isn't one. Mark up what's real on the page and move on. -- llms.txt as an access or ranking lever. It's a navigation aid. It doesn't control access (robots.txt does) and there's no evidence it influences AI Overview citations. Promptwatch's data on markdown in AI search shows markdown files make up just 0.05% of all AI search citations.
- Rewriting everything into listicles. That was the 2025 advice. Listicles still get cited, but they lost the top spot, and over-formating into listicles at the expense of substance lowers claim density.
- Lastmod gaming. Changing dates without changing content is detected and discounted. It's one of the few tactics that can actively hurt you.
How to know it worked
Set a baseline before you start. Track three things: which of your pages are currently cited in AI Overviews, for which prompts, and what your share of citations looks like versus competitors. Then re-check after each batch of fixes, on the same prompt set, so you're comparing like to like.
Expect movement in weeks for the extractability fixes (1, 2, 11, 14) and months for the authority and structural ones (7, 8, 9, 12). And keep in mind that citation share is partly a function of things outside your direct control, third-party mentions, forum presence, and competitor activity, so a flat line after fixing on-page issues usually means the bottleneck has moved off-page.
If you want to compare tracking options, the GEO software directory at bestgeosoftware.com covers the landscape, and ai-rank-tools.com lists rank trackers that include AI Overviews specifically.
The uncomfortable summary: most of these fixes are unglamorous. Front-loading answers, honest dates, real authors, server-rendered HTML, and a robots.txt audit won't generate conference talks. But they determine whether your content is even in the candidate pool, and in a system where roughly 10 citation slots per answer are contested by everything on the web, eligibility is most of the battle.