Search is changing venue: more and more people ask an AI instead of scrolling a results page. That created an entire market selling “shortcuts to show up in AI” - magic files, secret schema, miracle GEO. The truth, backed by recent primary sources (2024-2026), is less sexy and far more useful: AI mostly cites what already has authority. AI visibility is not a new trick - it is reputation made machine-readable.
Key takeaways
- AI search is still a small slice of traffic (less than 1% compared to Google Search, in the top-site sample from Similarweb, 2025), but it changes behavior: when an AI Overview appears, clicks on results drop by about 47% (Pew Research, 2025).
- There is no magic file or markup. Google itself states that it does not use
llms.txt, “AI text files” or special schema to appear in AI features - “it’s still SEO” (Google Search Central, 2026). - Ranking well helps a lot, but does not guarantee anything: part of AI Overviews citations comes from the organic top 10, part comes from deeper results, SERP features, YouTube and sources pulled in by query fan-out (Ahrefs, 2026 - vendor study).
- What drives citation (peer-reviewed study): citing sources, adding statistics, clarity and an authoritative tone increase visibility in generative engines; keyword stuffing reduces it (GEO, KDD 2024).
- Being mentioned outside your site matters: brand mentions and earned media correlate with AI visibility more than backlinks (Ahrefs; Muck Rack, 2026) - correlation, not a guarantee.
The reality of AI search
Start with the real size, because the hype distorts it. In June 2025, in the top-site sample analyzed by Similarweb, AI platforms generated about 1.13 billion referral visits; Google Search, in the same period, 191 billion. In other words, there AI is less than 1%of referral traffic compared to Google. Whoever says “X% of search is already AI” is usually conflating visits to AI platforms with search volume.
What is changing is not (yet) the volume, it is the behavior. Research by Pew Research (2025), with real browsing tracking, showed that 18% of Google searches displayed an AI Overview; and that, when it appears, the user clicks on any result only 8% of the time (versus 15% without an AI Overview) - and only 1% click a link inside the summary itself. Search is heading toward zero-click: you want to be the source cited in the summary, not just the blue page below it.
And how does AI choose whom to cite? Google’s AI features (AI Overviews and AI Mode) use RAG over the index and ranking systems of traditional Search - not a separate index. In Google’s own words: “AI Overviews and AI Mode rely on our core Search ranking and quality systems” (Google Search Central). ChatGPT Search retrieves from Bing’s index; Perplexity runs its own pipeline. In all of them, the pattern is the same: AI retrieves pages and cites a subset.
There is no magic file
This is the part the market does not want you to read. Google’s official documentation has an explicit “mythbusting” section: you do not need to create llms.txt, “AI text files”, special markup or schema to appear in AI features, because “Google Search doesn’t use them” (Google Search Central). What you need is the fundamentals: being indexed, snippet-eligible, with useful, unique content and sound technical structure. AEO and GEO, in practice, are still SEO - plus a clear entity and authority.
The llms.txt case illustrates the hype well. It is a convention proposed in 2024 (llmstxt.org) to give AI agents a summary of the site. Useful in development workflows (Claude Code, Cursor), but, as a search signal, the engines do not consume it in practice: Google declared it does not use it (John Mueller went as far as comparing it to the old keywords meta tag), and log analyses show AI bots ignoring the file and crawling the HTML directly (Search Engine Roundtable).
An honest nuance is worth recording, because it is real: in May 2026, Google Lighthouse (and, by inheritance, PageSpeed Insights) gained an audit for llms.txt- but inside a new, experimental category called “Agentic Browsing”, aimed at AI agents and WebMCP, not at Search ranking. And the test itself makes it clear: if the file does not exist, the result is “Not applicable” - it does not fail you (Chrome for Developers - Lighthouse). In other words: the tool is preparing for the agentic web, but that did not turn llms.txt into an SEO factor. Conflating the two is exactly the mistake this article avoids.
What actually makes AI cite you
If it is not a magic file, what is it? The evidence points to four levers - and all of them are more demanding versions of things you should already be doing.
- 1. Rank well. It remains one of the biggest advantages - but it is no guarantee. Data from Ahrefs (vendor study) shows that a relevant share of AI Overviews citations comes from the organic top 10, while another share comes from deeper results, SERP features, YouTube and sources retrieved via query fan-out. Without ranking, though, you rarely make the cut. And not only on Google: since ChatGPT Search relies on Bing’s index to discover pages (OpenAI) and Copilot is Bing directly, ranking well on Bing - which almost nobody optimizes for - is part of showing up in AI.
- 2. Structure content to be citable. The peer-reviewed study GEO (Aggarwal et al., KDD 2024) tested content tactics on a benchmark of 10,000 queries and found that citing sources, adding statistics, using quotations and clear, authoritative language increase visibility in generative engines (up to +40% in some domains), while keyword stuffing reduces it. Extractable content, with facts and sources, is what AI picks up.
- 3. Be a clear entity. AI needs to understand who you are before citing you with confidence. It is the direct bridge to entity resolution: organization/person schema,
sameAsto official sources, consistent identity. It is not a citation trick - it is what keeps AI from confusing you or ignoring you. - 4. Be mentioned outside your site. Here is the most counterintuitive finding: in vendor studies, brand mentions across the web correlate with AI visibility more strongly than backlinks (Ahrefs, 75,000 brands). And earned media shows up as a very strong source of citations: Muck Rack (2026) analyzed answers from ChatGPT, Claude and Gemini and found 84% of citations coming from earned media - a strong strategic signal, not a universal rule. AI also leans heavily on sources like Wikipedia, Reddit and YouTube (Search Engine Land). Honest caveat: all of this is correlation, not guaranteed causation - heavily cited brands are already big. But the strategic direction is clear: presence and reputation outside your own domain feed AI visibility.
AEO and GEO: two terms, no settled definition
An honest note is due here, because you have probably already run into these two terms - and into definitions that do not match. AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization) are young, market-made acronyms with no consolidated standard. Anyone who reads a lot on the subject tends to reach a logical conclusion - and an almost always partial one, because each source slices the cake differently.
The two most common slices:
- By mechanism type. AEO would be optimizing for answer engines - featured snippets, voice search and the boxes that deliver one direct answer (like the AI Overview). GEO would be optimizing for generative engines - ChatGPT, Gemini, Perplexity - which synthesize several sources and cite a few.
- By work layer. AEO would be the semantics: how you deliver the information - visible, objective, resolutive - the content ready to be cited. GEO would be what makes you deserve the citation: authority, originality, proprietary data, reputation and third-party mentions.
Both are defensible and circulate among serious professionals - but notice they are not even talking about the same thing: one splits by platform, the other by type of effort. And there is a detail that dismantles any rigid taxonomy: the paper that coinedthe term “GEO” (Aggarwal et al., KDD 2024) tested precisely contenttactics - citing sources, adding statistics, writing clearly - and called that GEO. In other words: what many people label “AEO” (the semantics of the content), the academic origin of the term calls “GEO”. The boundary is not a wall - it is a fog.
That is why our position is practical: do not fight over the acronyms. They overlap, change owners depending on who is writing and none of them is a button. What does not change - under any definition - is the work that makes AI cite you: ranking well, structuring extractable content (with facts and sources), resolving your entity and building reputation outside your site. Call it AEO, GEO or SEO on steroids: the name is debatable; the work is not.
The synthesis
AI visibility is not a file you install or a schema you hide in the code. AI tends to cite what has authority, structure and external corroboration. There is no shortcut - there is reputation made machine-readable.
Whoever promises you “showing up on ChatGPT” with a file or a plugin is selling the shortcut Google itself says does not exist. The real path is more laborious and more durable: rank, structure extractable content, resolve your entity and build reputation outside your site. It is the opposite of magic - it is engineering and authority.
At Inodus, AI optimization is a baseline requirement, but an honest one: structure, entity and the fundamentals that make AI understand and cite you. Want to see where your site stands? Run the free online audit.
Frequently asked questions
Do I need an llms.txt file to show up on ChatGPT or Google?+
No. Google states that Search does not use llms.txt or special markup (Google Search Central), and logs show AI bots ignoring the file in practice. It has its uses in development workflows, not as a search factor.
But didn't PageSpeed/Lighthouse start scoring llms.txt?+
In 2026 Lighthouse introduced an experimental llms.txt audit, inside a category aimed at AI agents (not Search ranking). And if the file does not exist, the result is 'Not applicable' - it does not fail you (Chrome for Developers). It is not an SEO factor.
What increases the chance of being cited by an AI the most?+
Ranking well, having extractable content (with facts, statistics and sources), being a clear entity and being mentioned on third-party sites (GEO, KDD 2024; Ahrefs).
Does schema markup make AI cite me more?+
Not directly - Google denies there is any special markup for that. Schema helps with comprehension and disambiguation of your entity, which is an important foundation, but it is not a citation button.
Can I pay to appear in AI answers?+
You can pay for an ad, not for the organic citation - and the difference is everything. ChatGPT has been testing ads since February 2026 (Free and Go plans), with a pilot announced for Brazil (OpenAI); Google shows ads in AI Overviews and is testing formats in AI Mode; and Microsoft Copilot displays ads inside the conversation. In all of them, the slot is labeled as sponsored and kept separate from the answer - OpenAI is explicit that ads 'do not influence the answers' (OpenAI). Perplexity went as far as testing ads and pulled back, arguing they erode user trust (Financial Times, 2026). In short: you can buy a slot identified as an ad; being the source AI cites as trustworthy is still earned - with authority, not media spend.
What is the difference between AEO and GEO?+
They are two market terms, young and without a settled definition - which is why sources disagree. Some split by mechanism: AEO for answer engines (snippets, AI Overviews, which give one direct answer) and GEO for generative engines (ChatGPT, Perplexity, which synthesize and cite sources). Others split by layer: AEO as the semantics of the content (making it citable) and GEO as what makes it deserve the citation (authority, originality, mentions). To complicate things, the paper that coined 'GEO' (KDD 2024) used GEO for exactly the content tactics many attribute to AEO. In practice they overlap, and both reduce to the same foundation: solid SEO, extractable content, a clear entity and authority. Don't fight over the acronym - do the work.
How we interpret the sources in this article
This content distinguishes four types of evidence: official documentation, case studies published by recognized sources, proprietary market studies and emerging research or analyses. Official data is treated as normative reference. Proprietary studies and benchmarks are used as directional signals, not universal rules. Academic research and log analyses about AI are presented as evolving technical evidence, especially where vendors have not yet defined public thresholds.
Methodology and sources
Official positions from primary sources: Google Search Central (generative AI guide and mythbusting) and Chrome for Developers (llms.txt audit in Lighthouse)official. Usage and behavior: Pew Research (independent)research/analysis and Similarwebproprietary. Evidence on what drives citation: the peer-reviewed paper GEO (KDD 2024)research/analysis. Correlation studies (Ahrefs, Muck Rackproprietary) are vendor studies, cited as directional signals, not causation - and flagged as such. Monetization: official announcements from OpenAI about the ad test in ChatGPT (Feb-May 2026)official; Perplexity’s retreat is attributed to Financial Times reporting (2026)research/analysis. This article is informational.