Top 5 GEO Myths You Should Ignore

If you spend enough time reading “GEO” advice on LinkedIn or X, you might come away believing that AI search runs on a mysterious collection of secret files, special markup, Markdown pages, and optimization tricks that traditional SEO professionals were apparently too unimaginative to discover. Just add an llms.txt file, sprinkle in some FAQ Schema, convert everything to Markdown, stop worrying about rankings, and congratulations: ChatGPT will surely fall hopelessly in love with your website.

Unfortunately, AI systems have proven stubbornly unwilling to follow the rules invented for them on social media. As researchers have started testing many of the most popular GEO claims, a less exciting picture has emerged: some are based on reasonable technical theories that simply lack evidence, while others dramatically overstate how different AI search is from the search ecosystem that came before it. So before you rebuild your website around the latest GEO hack somebody posted between screenshots of their prompt-tracking dashboard, let’s look at five of the biggest myths, and what the evidence actually says.

LLMs.txt is a proposed web standard that gives website owners a simple, machine-readable file, typically placed at /llms.txt, for pointing large language models and AI agents toward the content they consider most important, often using concise descriptions and links to preferred resources. Proponents argue that it can help AI systems understand a site’s structure and subject matter more efficiently, reduce the need to parse complex HTML or navigation, and make high value content easier for AI crawlers, assistants, and agents to discover and retrieve. Some advocates go further, suggesting that LLMs.txt could improve a brand’s chances of being cited or represented accurately in AI-generated answers, function as a kind of AI-focused counterpart to robots.txt or XML sitemaps, and eventually become a useful part of GEO/AEO strategy.

In June, Ahrefs analyzed server logs and live traffic from 137,210 websites to see whether llms.txt files are actually being used by AI systems. About 28% of the sites had an llms.txt file, but 97% of those files received no requests at all during May 2026. Among the small minority that were fetched, 96% of requests came from bots, but most were not the AI search and assistant bots site owners typically hope to influence. Only 19.5% of requests came from identified AI bots, and AI retrieval crawlers such as OAI-SearchBot and PerplexityBot accounted for just 1.1% of total requests. Ahrefs also found that AI bots did not probe for nonexistent llms.txt files, suggesting they generally discover them only when linked, indexed, or explicitly directed to them.

The study concludes that llms.txt currently offers little evidence of improving visibility in ChatGPT, Perplexity, AI Overviews, or similar search experiences. Its clearest present-day use appears to be with AI agents and coding tools, which accounted for more traffic than AI retrieval systems, while a notable share of requests came from SEO, GEO, validation, and research tools studying the standard itself. Ahrefs also cautions that a bot fetching the file does not prove it actually used the contents. Their overall takeaway is that llms.txt may have future value in an increasingly agentic web, but not AI Search, and it is inexpensive to implement, but today it should not be treated as a meaningful GEO/AEO ranking or citation tactic.

It's worth noting that Google uses LLMs.txt files throughout parts of its own website and has also included LLMs.txt evaluation and audit tools as part of its Lighthouse "Agentic browsing audits". It may seem as if because of this, these files have influence within AI search. However, representatives from Google have distinguished the difference between AI Search Visibility and agentic web optimization. The first pertains to increasing visibility in AI systems, the second focuses on AI agents' ability to understand and engage with a site.

Markdown is a lightweight text-formatting language that uses simple characters to define elements such as headings, links, lists, and emphasis without the additional code and structural overhead found in a typical HTML document. Proponents of using Markdown for GEO argue that serving cleaner, stripped-down versions of web pages to AI crawlers can make content easier and more efficient for large language models to retrieve, parse, and understand by removing navigation, scripts, styling, and other boilerplate that may dilute the main content. Some also claim that Markdown can reduce page size, improve token efficiency, and present information in a format that more closely resembles the structured text commonly used in LLM training and retrieval pipelines. However, while these arguments are technically plausible, there is currently limited evidence that serving Markdown instead of well-structured HTML directly improves visibility, citations, or mentions in AI-generated responses.

Profound tested 381 pages across six websites, randomly serving HTML to one group and clean Markdown to another for AI bots. The Markdown group had a small directional increase in bot visits, but the difference was not statistically significant. The median page saw roughly one additional bot visit over three weeks. ChatGPT-User showed a roughly 20% directional Markdown advantage, but the researchers explicitly stopped short of concluding that Markdown caused better AI visibility.

An even larger ongoing experiment from Trakkr is especially relevant to OpenAI. Their randomized test covers more than 8,000 pages and multiple AI crawlers. As of August 2026, they found that OAI-SearchBot reached Markdown and HTML pages at essentially the same rate. ChatGPT-User was also effectively neutral. Interestingly, GPTBot actually showed a substantial preference for HTML, not Markdown.

FAQPage Schema is a type of structured data, usually implemented with JSON-LD, that explicitly identifies a page as containing a series of questions and answers and defines the relationship between each question and its corresponding answer. Proponents of using FAQPage Schema for GEO argue that this machine-readable structure can make content easier for AI systems to interpret, extract, and associate with specific user questions. They claim that clearly labeling question-and-answer relationships may improve the chances that content is selected for retrieval, summarized accurately, or cited in AI-generated responses. Some also argue that FAQ markup can help reduce ambiguity by giving language models and search systems a more explicit representation of a page’s meaning. However, while these claims are plausible from a semantic-structure standpoint, current research has not established that adding FAQPage Schema by itself consistently increases visibility or citations in AI chat systems.

Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026 and compared them with roughly 4,000 matched controls. For ChatGPT, adding schema was associated with only a +2.2% change in citations, which was statistically indistinguishable from zero. Google AI Mode was similarly neutral at +2.4%, while AI Overviews actually declined slightly relative to controls. Their conclusion was essentially that adding schema by itself did not meaningfully increase AI citations.

An experiment from The GEO Lab ran 480 queries across ChatGPT, Gemini, and Perplexity and found FAQ-marked pages had a 6.7% citation rate versus 8.3% for pages without FAQ schema, effectively a null result rather than a benefit.

Google also argues against treating FAQ Schema as a special AI optimization tactic. Its guidance for AI Overviews and AI Mode says there is no special schema.org markup required for AI features and that existing SEO fundamentals remain the main requirement. Google specifically says structured data should accurately match visible page content. Google also fully deprecated its FAQ rich-result feature in May 2026,

Some GEOs argue that traditional search rankings do not matter, or matter far less in AI visibility because generative systems do not simply reproduce Google’s top results when constructing answers. They point to studies showing relatively low URL-level overlap between organic rankings and citations in systems like ChatGPT, Perplexity, and Gemini, as well as the use of query fan-out, retrieval systems, reranking, and passage-level selection that can surface pages ranking well for related subtopics rather than the user’s original query. From this perspective, a page does not need to rank in the top few positions for the exact prompt to be mentioned or cited by an AI system, which has led some practitioners to frame GEO as largely independent of conventional rankings. The problem with that interpretation is that it can confuse imperfect overlap with no relationship at all: emerging research suggests that search visibility can still play an important upstream role in determining which pages and domains enter the retrieval pool, even though ranking alone does not determine which sources ultimately appear in the generated response.

AirOps analyzed 15,000 prompts, 43,233 original + fan-out queries, and 548,534 retrieved pages in ChatGPT. They found that 55.8% of cited pages ranked in Google's top 20 for at least one original or fan-out query. More strikingly, pages ranking #1 in Google were cited by ChatGPT 43.2% of the time, about 3.5× more often than pages outside the top 20.

Some GEOs argue that traditional SEO best practices matter less in an AI-driven search environment because large language models do not evaluate pages in exactly the same way that Google’s ranking systems do. Their argument is usually that AI systems can surface brands and information that do not rank prominently in conventional search results, especially when those brands are frequently mentioned across trusted publications, communities, databases, and other third-party sources. From that perspective, GEO becomes less about optimizing pages for rankings and more about increasing the likelihood that a brand, product, or concept appears in the information environments an AI system uses to construct an answer.

There is some truth behind that argument, but taking it to the point of ignoring SEO is a mistake. Most major AI answer systems are still heavily dependent on search-enabled retrieval-augmented generation, or RAG, when they need current or verifiable information. ChatGPT Search, for example, searches the web and uses retrieved sources when forming answers, while Perplexity describes its own process as conducting multiple web searches, crawling sources, and synthesizing those results into a response. In practical terms, this means that crawlability, indexability, internal linking, canonicalization, page performance, clear site architecture, and other long-established SEO fundamentals can still determine whether a search or retrieval system can reliably discover and understand your content in the first place.

Ignoring SEO can therefore undermine the very GEO strategy a team is trying to build. If important content is blocked from crawling, poorly linked, duplicated across URLs, rendered unreliably, or difficult for search systems to index, an AI system that depends on web retrieval may have fewer opportunities to encounter or cite it. GEO should be treated as an extension of strong SEO and broader digital marketing rather than a replacement for them: technical SEO makes information accessible, content and entity optimization make it understandable, and brand building and digital PR increase the likelihood that it is reinforced across the wider web. As long as AI assistants continue to use search and web retrieval as major inputs for current answers, abandoning SEO fundamentals is less a new optimization strategy than a decision to weaken one of the primary discovery channels those systems still rely on.

Prompt Engineering: Writing prompts in a way that improves AI output quality.

One more thing: AI is only as good as its operator, and if you are reading this newsletter, you’re better than most!

Till next time,

PS: Let me know what you think of this issue, or anything else here: [email protected]