How to Get Cited by AI Search: What the Evidence Actually Says in 2026

Every marketing blog in 2026 is selling a trick for getting cited by ChatGPT, Perplexity, and Google's AI Overviews. Most of the tricks share a feature: nobody tested them. This article is the opposite. It covers what controlled studies actually found, which popular claims fall apart under scrutiny, and what remains standing when the hype is stripped away.

The short version: there is no schema hack, no magic file, no prompt trick. What gets cited is what has always gotten cited, just measured differently: original information, clearly structured, from sources the web already talks about.

Key takeaways

• Adding JSON-LD schema produced no measurable citation uplift in the best-controlled study available (Ahrefs, 1,885 pages, 2026). Schema is hygiene, not a growth lever.
• The famous "FAQ schema gets 3x more citations" claim is a correlation misread as causation. Cited pages were more likely to have schema because well-run sites do both.
• Google explicitly says no special markup or AI files are needed for AI Overviews or AI Mode. A 54-study meta-analysis scored llms.txt 2.0/10, lowest of all factors.
• What correlates with citations: original data, quotable stats, clear Q&A structure, and brand mentions across the web (0.664 correlation vs 0.218 for backlinks).
• E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) remains the right quality framework, with Trust as the load-bearing element.

Start with how AI search actually works

Before tactics, the mechanism. Google has stated that its generative AI features are rooted in core Search systems, and that normal technical and quality requirements for Search still apply, with no additional technical requirements for AI Overviews or AI Mode. In practice: a page must be crawled, indexed, and eligible for a normal search snippet to be eligible as a supporting citation in an AI answer. Discovery, retrieval, and citation are three separate outcomes. Being crawled doesn't mean being retrieved; being retrieved doesn't mean being cited.

Each AI provider runs its own crawlers and fetchers with different purposes, and blocking one bot doesn't uniformly remove you from every AI product. There is no single "AI crawler" and no universal GEO score, despite what vendor dashboards imply. Anyone selling you a single number for "AI visibility" is selling an analytical framework, not a platform metric.

The FAQ schema myth, debunked properly

This deserves a full section because it's the most repeated false claim in the industry, and this publication will not repeat it.

The claim: "pages with FAQ schema get 3x more LLM citations." The origin: an Ahrefs observational analysis of millions of URLs found that pages cited by AI tools were almost three times more likely to carry JSON-LD structured data than pages that weren't. That statistic is real. The conclusion drawn from it is not.

Here's the problem, stated plainly: correlation is not causation, and this is a textbook confounded correlation. Websites that bother implementing structured data are overwhelmingly the same websites investing in technical SEO, better content, faster pages, and digital PR. The schema is along for the ride. Concluding "add FAQ schema, get cited" from this data is like concluding that wearing running shoes makes you fast because fast runners wear them.

Somebody actually ran the experiment. The same Ahrefs research team took 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matched each against comparable pages that didn't, and tracked citations across Google AI Overviews, Google AI Mode, and ChatGPT. The results: AI Overviews citations changed by −4.6% (the only statistically significant figure, and it went down), AI Mode by +2.4%, ChatGPT by +2.2%, with the positive numbers indistinguishable from zero. As the authors put it, adding schema produced no major uplift in citations on any platform.

One honest caveat the researchers note: the study tested pages that were already visible, each with 100-plus AI Overview citations before adding schema. Schema may still matter further upstream, gating rich-result eligibility and supporting entity recognition. The takeaway stands: schema is a floor, not an accelerator.

You'll also see vendor studies claiming big FAQ schema wins, such as a 2025 Relixir analysis of 50 sites reporting 41% versus 15% citation rates. Treat vendor marketing as what it is: the primary document is inaccessible, the methodology isn't published, and the vendor sells the tactic it's measuring. The controlled data says otherwise.

What the platforms themselves say

There are only two first-party statements worth citing, and they point in opposite directions, which is itself informative.

Google's Search Central documentation states directly: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." That is as clear as a platform gets.

Microsoft's Fabrice Canel, a Principal Product Manager on Bing, confirmed at SMX Munich in March 2025 that schema markup helps Microsoft's language models understand content. That's a genuine on-record statement from a company running an AI search product, and it deserves weight. But note what it says: schema helps models understand content. It does not say schema gets you cited. OpenAI, Anthropic, and Perplexity have said nothing on the record about schema at all.

The honest synthesis: structured data is good engineering that aids machine comprehension. It is not a citation lever. Do it because it's correct, not because it's a hack.

llms.txt and the lottery-ticket tactics

The llms.txt file, a proposed plain-text standard telling AI crawlers what a site contains, is the other heavily marketed shortcut. The evidence verdict is blunt: a 54-study meta-analysis published by Zyppy in May 2026 swept experiments, patents, and case studies across the field and scored llms.txt 2.0 out of 10, the lowest of all claimed citation factors, finding no credible evidence of impact.

That doesn't make the file harmful. It's cheap to add and may help as documentation. But the industry sells llms.txt generators as investments while its own evidence base grades the tactic 2 out of 10. File it under lottery ticket, not strategy.

E-E-A-T: the framework AI search actually rewards

If schema isn't the lever, what is? Start with the quality framework Google actually uses to evaluate content: E-E-A-T, from its Search Quality Rater Guidelines. That's Experience, Expertise, Authoritativeness, and Trustworthiness, with Trust as the most important element and the thing the other three build toward. "Experience" was added in December 2022 to credit first-hand, lived knowledge over generic rewrites.

Here's how E-E-A-T translates into AI-citation practice:

Experience: Show your work. First-hand testing, original measurements, dated observations, and specific details ("we tested 14 VPNs over 30 days") beat generic summaries. AI systems are trained to prefer content with verifiable, specific claims because those are citable.

Expertise: Put real credentials on the page. Named authors with relevant backgrounds, cited sources, and correct technical depth. Anonymous, thin content is the first thing both raters and retrieval systems discount.

Authoritativeness: Be the source other sources cite. This is where the data gets interesting: an Ahrefs analysis of 75,000 brands found web mentions correlated with AI visibility at 0.664, compared to just 0.218 for backlinks. Being talked about, quoted, and referenced across the web matters roughly three times more than link graphs for AI citation. Digital PR beats link building in the AI-search era.

Trustworthiness: Be transparent about who you are, when you published, and what you don't know. Clear authorship, contact information, correction policies, and as-of dates. For YMYL topics (money, health, legal), this isn't optional; it's the price of admission.

What actually earns citations: the honest playbook

Synthesizing the controlled studies and platform statements, here's what the evidence supports, roughly in order of impact:

1. Publish original data. Proprietary research, surveys, benchmarks, and measurements are the most-cited content type because they're the only source for the numbers. If every AI answer needs a statistic, the site that produced the statistic gets cited.
2. Write quotable, self-contained answers. Structure content so a single paragraph or list directly answers a specific question. AI systems extract and cite passages; make your passages extraction-ready with clear headings, definition sentences, and tables.
3. Earn brand mentions across the web. At a 0.664 correlation, this is the strongest measured factor. Get quoted in journalism, cited in research, discussed in communities. This is slow, unsexy work, and it's the actual game.
4. Keep the technical floor solid. Fast pages, clean semantic HTML, accurate schema, canonical URLs, accessible content. None of this gets you cited; all of it keeps you eligible. The floor matters because AI Overviews draw from pages eligible for normal search snippets.
5. Be specific and current. Dated, precise claims ("as of October 2026") outperform vague evergreen text because AI systems prefer recency and verifiability for time-sensitive queries.
6. Measure honestly. Track a panel of real prompts weekly, capture what the AI actually says, and check your referral logs: ChatGPT referral URLs include utm_source=chatgpt.com, per OpenAI. Diagnose gaps, fix the source content, and re-measure. Consistency beats intensity.

What doesn't make the list: FAQ schema as a citation tactic (debunked above), llms.txt (2.0/10), keyword-stuffing for "AI readability," and any vendor dashboard score presented as a platform metric.

If you're applying this to a Canadian business context, the same principles hold whether you're writing about Black Friday email marketing or explaining open banking: original numbers, clear answers, real authorship, and mentions from Canadian publications.

Practical next steps

• Audit your top 20 pages for extractability: does each answer one specific question in a quotable passage? If not, restructure before doing anything else.
• Add or fix schema markup as hygiene (Article, FAQ, Product where accurate), with zero expectation of a citation bump. Never add markup that misrepresents the page.
• Start a simple prompt panel: 10–20 real customer questions, run weekly through ChatGPT, Perplexity, and Google AI Mode. Record whether you're cited and what the answer says about you.
• Invest in digital PR over link schemes: original research, expert commentary for journalists, and community presence earn the brand mentions that correlate most strongly with citations.
• Put E-E-A-T on every page: named authors, bios, publish dates, sources cited, and a corrections policy. For money and health topics, add explicit as-of dates and disclaimers.
• Skip the lottery tickets, or buy them cheaply: an llms.txt file costs nothing to add, but don't budget real money or strategy around it.

The bottom line

Getting cited by AI search in 2026 is not a technical trick; it's a publishing strategy. The controlled evidence killed the two most popular shortcuts: FAQ schema doesn't move citations (the "3x" claim was correlation mistaken for causation), and llms.txt scores 2 out of 10 across 54 studies. What works is what the data supports: original research, quotable answers, clean technical foundations, and above all, being mentioned across the web by sources that matter. E-E-A-T isn't a hack either; it's the quality bar. Clear that bar consistently, measure honestly, and the citations follow. Anyone promising faster should be asked for their control group.

Sources

• https://semarkglobal.com/blog/schema-markup-ai-search-what-helps-llms-understand-you
• https://forgeandsmith.com/blog/schema-wont-get-you-cited-by-ai-heres-what-its-good-for/
• https://www.scribblersindia.com/blog/schema-markup-for-ai-search/
• https://dev.to/edo911/how-ai-systems-read-the-web-in-2026-discovery-retrieval-citations-and-agents-4i1k
• https://dev.to/edo911/how-to-measure-ai-search-visibility-in-2026-the-evidence-first-geo-framework-nb3
• https://medium.com/@bsim7116/do-llms-txt-and-schema-markup-get-you-cited-by-ai-what-54-studies-say-9082280f8a38