What Actually Gets You Cited by AI? LLMs.txt, Schema, and Entity Signals, Ranked

Thoughts, ideas, and perspectives on design, simplicity, and creative process.

What Actually Gets You Cited by AI? LLMs.txt, Schema, and Entity Signals, Ranked

A male digital strategist in round glasses reviewing analytics on a smartphone against a yellow wall, flanked by glassmorphic UI panels showing LLMs.txt vs Schema rankings and GEO entity confidence scores for local businesses, events, and brands
You are ranking. You are publishing. You’re performing all the things your SEO playbook instructed you to do. However, if someone searches for ChatGPT, Perplexity, or Google’s AI Overviews and types in a question you are in, your competitor’s name will be part of the answer, and your name won’t be.

This is not a rank problem. It is a selection problem. Traditional search engines took you to the most relevant page; a new generation of LLM-powered search engines creates a single answer and then selects which sources to trust. The question is no longer “rank highly. It is “chosen as a trusted source by the AI.

There are three levers that are the basis of every talk of how to get that selection, namely: llms.txt, schema markup, and entity signals. They are not equal. Some engines, some accelerants, and one is primarily insurance. In this article, they have ranked them truthfully based on peer-reviewed research and large-scale measurement studies, and they have mapped them onto a 5-layer visibility model that visualizes exactly where each lever plays in the AI answer pipeline.

The Split That Broke SEO

For 20 years, there was one number that could answer almost any question – yours or anyone else’s: Your position on the page. Well, if you ranked, you were seen. This contract has been broken, and the statistics are stark.

New independent measurement reveals that just 17% of AI Overview citations are from pages ranking in the top ten organic results on Google (Straight North 2026), a significant decrease from about 76% in mid-2024 (Straight North 2024). The win-to-get-cited game has taken off. Even as SparkToro found that in early 2026, 68% of all US Google searches were never clicked, if an AI Overview does show up, it typically steals about 58% of the click-through rate (Meev). By mid-2026, Google AI Overview coverage was estimated to be between 40-48% of all searches.

The focus here is on the commercial, not the theoretical – the click is no longer the goal. It’s inside the synthesized answer is. Generative Engine Optimization (GEO) is not just a modification of SEO; it’s a new mindset in producing, distributing, and asserting information during the AI era. They have identified this transition in five aspects: Information retrieval, NLP, the RAG architecture, Vector representations, and machine readability. This will be the AI Visibility Framework (AIVI), and we will be returning to this shortly, as it’s the map that will help you know which lever to pull first.

#1. Entity Signals: The Engine of AI Trust

Entity signals are the best lever as they address the only question the retrieval layer truly cares about – “Is this brand a true, distinct, and recognized thing?
In machine terms, an entity is not your web page. The concept your business represents on the web: your brand, your business, your location, and your reputation in review sites, directories, the news and social media, as well as your knowledge bases, such as Wikidata and Wikipedia. If these descriptions match, the model creates a solid and strong representation of your brand with confidence. If they don’t agree or if there are very few, you’re a ghost.

The evidence for entity dominance is the strongest in the field:

 

  • Ahrefs (May 2025) analyzed 75,000 brands and found that web mentions correlate with AI Overview visibility at 0.664, versus backlinks at just 0.218. Brands in the top mention quartile averaged 169 AI Overview mentions versus 14 for the next quartile down, more than a 10x gap. Mentions, not links, move the needle.
  • CiteLens (June 2026) benchmarked citations across engines and found that Claude’s citations track brand recognition more than ranking, with 58% of its citations going to sites with a Wikipedia presence. ChatGPT followed neither Google ranking nor brand popularity closely, but entity authority still mattered more than position.
That’s the reason entity signals are towards the top rank. They’re running on two AIVI layers at once: information retrieval (when they’re able to retrieve based on semantic similarity, it’s better, and it’s retrieved more often), and vector representations (when the description of your brand is consistent across the web, it creates a stronger embedding of your brand). A model can’t fix no brand.
Strengthening the entity is not a technical matter, but a commercial one, with a canonical brand description in place everywhere, a Wikidata entry and structured Organization data, reviews that are earned, and a website that clearly states who you are and what you do. It is slow. It’s also the only moat in AI search that a competitor can’t clone in an afternoon.

#2. Schema Markup: The Clarifier, Not the Citation Switch

Schema.org structured data is the second most important lever and is a very misunderstood one, so let’s be clear on what it actually is.
Schema is annotation that can be read by machines. It contains unambiguous information to crawlers and models, such as: this is an article by a named author, this is a product with a price, this organisation is the same as the one at that Wikidata URL. It resides at the machine readability level of the AIVI framework.
This is the truth that most agencies gloss over: there’s no direct link between schema and citations. Google clearly says that there is no requirement for special Schema markup to be present on a page to be featured in AI Overviews or AI Mode, and that publishers should not rely on structured data. Any page that contains all schema types does not need to be marked, and any page that contains no marking will be marked as it is. It appears that much of schema’s impact is the correlation, i.e., teams disciplined well enough to use clean schema tend to be disciplined well enough to write well-sourced content; it’s the content that gets the citation.

So why does schema rank second? Because it is the highest-leverage clarifier in the stack:

 

  • Entity disambiguation. sameAs links and a complete Organization graph connect your brand to the knowledge bases models trust, reinforcing layer one.
  • Passage clarity. Well-structured pages chunk more cleanly during retrieval, so the right passage, not a tangled sidebar, gets pulled into the answer.
  • Role signaling. Article, FAQ, Product, and HowTo types tell the model what kind of statement each block is, reducing the chance your pricing page gets quoted as a definition.
The secondary but real role is backed by practitioner evidence: in a few case studies of graph-first, server-side JSON-LD implementations, the indexing times were faster and the behavior of the model with citations improved within 4 to 8 weeks, not because the model was charmed by the markup, but because the ambiguity went away. FAQPage was also completely phased out of Google’s SERPs in May 2026, despite the presence of FAQ rich results on the page, so it is no longer a SERP feature; rather, it is a structuring tool for machines. Publish Q&A in FAQPage, community threads in QAPage. Then deploy it server-side, where it is stable and in sync with the real data, rather than injected by plugins that the crawler might not run.
The second lever is schema, which forces the rest. It cannot save an unrecognized entity or thin content, which is why it’s underneath entity signals and above the next contender.

#3. llms.txt: Cheap Insurance, Not a Strategy

llms.txt is the most talked-about and least proven of the three. It is a plain Markdown file at your site’s root, proposed by Jeremy Howard in September 2024, that summarizes your site’s structure and points AI systems at the pages worth ingesting. Think of it as a curated table of contents written for language models rather than search indexers.

The adoption story has two halves, and agencies usually tell you only the first:

 

  • Publishers adopted it. Anthropic, Cloudflare, Stripe, Vercel, Perplexity, and others publish llms.txt files, and community directories track many hundreds of implementations.
  • Consumers did not. No major AI platform has publicly confirmed that it systematically reads llms.txt in its production retrieval pipeline. Server-log studies have found AI crawlers generally do not even request the file, and Chrome’s Lighthouse audit treats a missing llms.txt as “not applicable,” not a failure.
  • The most rigorous measurement available, a Search Engine Land test tracking ten sites for 180 days (ninety days before and after adding the file), found measurable AI-traffic increases on just two of the ten sites (12.5% and 25%), with the other eight showing no significant change.

So why rank it third instead of dismissing it? Because it costs almost nothing. A well-curated llms.txt takes a few hours: a clear summary of what your business does, links to your money pages, pricing, and key documentation, and an optional full-content companion file. If agentic browsing matures, and Chrome’s new Agentic Browsing audit category suggests the industry is preparing for it, you will already be legible. That is insurance, priced correctly. What it is not is a citation strategy. Any agency selling llms.txt as the centerpiece of an AI-visibility program is selling you a smoke detector as a fire department.

The Order of Operations

If you are budgeting for AI visibility this quarter, the sequence is now obvious:

 

  • Entity audit. Is your brand resolvable as an entity? Consistent descriptions, Wikidata presence, earned mentions, reviews. This is the foundation everything else stands on.
  • Schema graph. Server-side JSON-LD as a connected entity graph, not scattered markup. Disambiguate the brand, type the content correctly, structure the passages
  • . llms.txt. Ship it in an afternoon as insurance, then move on.
  • Measure citation share, not rankings. Track how often each engine mentions and cites you, and who it recommends instead.
  • That sequence, in that order, is what a serious AI-visibility program looks like. It is also exactly what our free 48-hour AI visibility audit measures: your entity resolution across engines, your schema’s machine readability, and your current citation share against competitors, delivered with the gaps ranked by revenue impact. If you want to know whether the AI trusts your brand yet, this is the fastest way to find out.

Frequently Asked Questions

Have Questions About Our Marketing Services? We Have Answers!

While this is not confirmed, it’s also not ruled out that major AI platforms read llms.txt in production retrieval. It is simple to do and could be important as agentic browsing grows as an option, but it will be a plan B, not a plan A.

Not directly. Google says there is no special schema needed for AI Overviews, nor is there any need for pages to be marked up for them. The true power of Schema is eliminating the ambiguity: disambiguate entities, retrieve cleaner passages, and type them correctly for more accurate quoting of your very best content.

The entity signals are the machine-readable facts that appear in a consistent way on your website, on wikis and knowledge bases such as Wikidata, on review sites, directories, and in the news – about your brand. The one thing that determines AI citation best is strong entity signals.

Because there are more and more systems that are separate for retrieval of AI answers and classic ranking. Just 17% of AI Overview citations are from the top ten organic results, compared to about 76% in mid-2024. Sources are chosen on a basis of trust or entity authority, not position.

The research and measurement clearly indicate the order: entity signals (brand is a trusted entity), schema markup (machine-readable clarity), and lastly, llms.txt (unproven, cheap insurance). All three are supported by content that the Princeton GEO study revealed could be quoted by the models- that is, specific statistics, named sources, and direct quotations- which the study found could raise the visibility of AI-answer content by up to 40%.

To Get Started, Simply Fill Out
The Form Below!