A file about cats should not improve your AI citation rate. Yet when the same correlation-based arguments used to promote llms.txt geo effectiveness were applied to a file named cats.txt — containing nothing relevant — the “evidence” held up just as well. That experiment, reported by Search Engine Journal in 2025, did not just embarrass one tactic. It exposed a pattern of reasoning that is costing marketing teams real budget on interventions with no verified mechanism of action.
What llms.txt Actually Is — and What It Claims to Do
llms.txt is a plain-text file placed at the root of a domain, intended to give large language models a structured summary of a site: its purpose, key pages, and preferred descriptions. The proposal draws a loose analogy with robots.txt, which genuinely influences crawler behaviour because crawlers are programmed to read it. The argument for llms.txt is that models like ChatGPT, Gemini, Perplexity, or Google AI Overviews will read this file during retrieval and weight the site more favourably as a result.
That mechanism has never been confirmed by any of those systems. OpenAI, Google, Anthropic, and Perplexity have not published documentation stating that their retrieval or ranking pipelines read or act on llms.txt. The file exists; the causal chain does not.
The Cats.txt Problem: Why Correlation Is Not a GEO Signal
The cats.txt experiment worked like this: a site added a file named cats.txt — containing descriptions of domestic cats, entirely unrelated to the site’s topic — and then observed citation behaviour in AI systems over the following weeks. Proponents of llms.txt had been pointing to before-and-after citation changes as proof that their file was working. The cats.txt test applied identical logic and found identical “results.”
The implication is straightforward. If any file placed at root correlates with citation changes, the file is not the cause. Citation behaviour in generative systems fluctuates for many reasons: model updates, index refreshes, changes in competing content, shifts in how prompts are interpreted. Attributing those fluctuations to a specific file — without a controlled test and without confirmation from the model provider — is the kind of reasoning the experiment was designed to expose.
Search Engine Journal described this as “GEO astrology”: pattern-matching that feels like signal but carries no predictive or causal weight.
Frequently Asked Questions on llms.txt and GEO Effectiveness
Question: Does llms.txt influence how ChatGPT or Perplexity cite my site?
Answer: There is no published documentation from OpenAI or Perplexity confirming that their retrieval pipelines read or act on llms.txt. Observed citation changes after adding the file cannot be attributed to the file without a controlled test, because citation rates fluctuate independently of any single on-site change.
Question: Is llms.txt harmful to implement?
Answer: Implementing the file is unlikely to cause damage. The risk is not the file itself but the opportunity cost: teams that treat llms.txt as a GEO priority may deprioritise interventions — structured data, extractable content, entity consistency — that have a documented mechanism of action.
Question: What does “extractable content” mean in a GEO context?
Answer: Generative models retrieve content in chunks — short, self-contained passages that answer a specific question. A page structured with clear headings, short paragraphs, Q&A blocks, and bullet lists produces chunks that a retrieval system can lift and include in a generated answer. A page built around long narrative prose or JavaScript-rendered content does not.
Question: Why does schema markup matter for AI citation if Google already knows what my site is about?
Answer: Schema markup in JSON-LD format anchors your brand as an entity in a knowledge graph. Without it, a model has no structured signal to associate your domain with a specific category, product type, or area of expertise. A good organic ranking does not transfer automatically to AI citation; entity recognition is a separate layer.
Question: How do generative systems like Google AI Overviews or Gemini actually select sources?
Answer: These systems use semantic retrieval — matching query embeddings to content embeddings — combined with retrieval-augmented generation (RAG). They select a small number of sources, typically three to five, and synthesise an answer. Selection is influenced by content extractability, entity signals, third-party validation, and freshness indicators such as visible dates and identified authors. A plain-text file at root is not part of that documented process.
Question: What counts as third-party validation for AI citation purposes?
Answer: Mentions of your brand or product on authoritative external sources — editorial coverage, industry directories, platforms with high user-generated content — signal to models that your entity is recognised beyond your own domain. This earned-media layer is distinct from anything you can control on your own site.
Question: How often do I need to revisit GEO optimisation?
Answer: Continuously. Model weights update, retrieval criteria shift, and competing content changes. GEO is not a one-time project. A site that was well-optimised for Perplexity’s retrieval logic six months ago may have drifted out of alignment without any change on its own end.
How to Prioritise GEO Interventions That Have a Verified Mechanism
The following steps focus exclusively on levers with a documented or structurally logical connection to how generative retrieval works. Each step notes where manual effort hits a practical ceiling.
Step 1 — Audit content extractability across your key pages.
For each priority page, check whether the main answer to the page’s topic appears in the first two paragraphs, in plain HTML (not rendered by JavaScript), under a descriptive H2. If the answer is buried in a long introduction or depends on JS execution to appear, a generative crawler may never reach it. Rewrite the opening of each page so the core claim is stated immediately and unambiguously.
Step 2 — Implement JSON-LD schema for your core entity types.
At minimum: Organisation (with name, url, description, sameAs pointing to your LinkedIn and Wikipedia if available), WebSite, and FAQPage on any page that contains Q&A content. For product or service pages, add Product or Service schema with a clear description. These markup types give models structured signals to associate your domain with a specific category — the mechanism that a plain-text file cannot replicate.
Step 3 — Establish entity consistency across the page.
Your brand name should appear in the title tag, the H1, and the meta description in exactly the same form. The first paragraph of your homepage or About page should contain an explicit definition: “[Brand] is a [category] that [does X].” Models use this pattern to anchor the entity. Inconsistency across pages — different name variants, missing definitions — weakens the signal.
Step 4 — Add verifiable trust signals to every key page.
Visible publication dates, named authors with brief credentials, and linked external sources all contribute to the freshness and credibility signals that generative systems use to filter sources. A page with no date and no author attribution looks stale to a model even if the content is accurate and recent.
Step 5 — Build earned-media presence on the platforms models read.
Reddit, LinkedIn, YouTube, and authoritative editorial outlets are consistently represented in AI-generated answers. A brand that exists only on its own domain has a thin off-site footprint. Prioritise one or two channels where your category has genuine discussion, and contribute content that can be retrieved independently of your site.
The arithmetic of manual monitoring alone illustrates the scale problem: if you track four AI systems (ChatGPT, Gemini, Perplexity, Google AI Overviews) across twenty representative prompts per system, that is eighty manual checks per cycle. Run that monthly and you have nearly a thousand checks per year — before you have touched a single fix or re-checked whether a previous change had any effect.
Verified GEO Levers vs. Unverified Tactics: A Comparison
| Tactic | Documented mechanism | Manual effort ceiling | Measurable outcome |
|---|---|---|---|
| llms.txt file | None confirmed by any model provider | Low (one-time file) | No reliable before/after signal |
| Extractable HTML content structure | RAG retrieval selects self-contained chunks | High across large sites | Trackable via prompt monitoring |
| JSON-LD schema (Organisation, FAQPage) | Entity anchoring in knowledge graph | Medium; requires dev queue | Measurable via entity recognition checks |
| Explicit brand definition in copy | Model uses “X is a Y” patterns for entity association | Low per page; high at scale | Trackable via citation accuracy |
| Visible dates and named authors | Freshness and credibility signals in retrieval | Low per page | Indirect; part of trust scoring |
| Earned media on high-authority platforms | Off-site mentions feed model training and RAG | High; ongoing | Share-of-AI-voice over time |
| Continuous prompt monitoring | Direct observation of citation presence | Very high at scale (see arithmetic above) | Direct; the only ground truth |
Case Study: Diagnosing Why a B2B SaaS Brand Was Absent from AI Answers
A mid-sized B2B SaaS company had invested several months in content production and had added an llms.txt file following advice circulating in SEO communities. Their organic rankings were solid. Yet when their team ran systematic prompts across ChatGPT, Gemini, and Perplexity using category-level queries — the kind a buyer would use during vendor shortlisting — the brand did not appear in any generated answer.
The diagnosis phase revealed three structural issues that the llms.txt file could not address. First, the site’s core service pages opened with marketing narrative rather than a direct definition; a retrieval system chunking the page would find no clean, self-contained answer in the first passage. Second, the Organisation schema was missing entirely, and the FAQPage markup on the blog had been implemented incorrectly, rendering it invisible to structured-data parsers. Third, the brand name appeared in three different forms across the site — abbreviated in some titles, spelled out in others — creating inconsistent entity signals.
The team spent the majority of their remediation time on diagnosis: mapping which prompts triggered competitor citations, identifying which pages lacked extractable openings, and auditing schema validity. The actual fixes — rewriting page openings, correcting JSON-LD, standardising the brand name — were completed in a fraction of the total time. Within weeks of implementation, the brand began appearing in Perplexity answers for category queries. ChatGPT citations followed. The llms.txt file remained on the server, unchanged and irrelevant to the outcome.
What This Means for How You Allocate GEO Effort
The cats.txt experiment is not an argument against GEO. It is an argument for GEO done on the basis of documented mechanisms rather than plausible-sounding analogies. The levers that actually influence whether ChatGPT, Gemini, Perplexity, or Google AI Overviews cite your brand — extractable content structure, JSON-LD entity anchoring, earned off-site presence, consistent brand definition — are well-understood and implementable. They are also continuous: models update, retrieval criteria shift, and the competitive landscape changes without warning. The structural limit of any manual approach is not effort per intervention; it is the impossibility of monitoring dozens of prompts across four AI systems on an ongoing basis while simultaneously tracking whether each fix has produced a measurable change in citation behaviour. That is the work aisearchaudit.ai is built to do: running the audit continuously, surfacing the gaps with evidence, and making the before-and-after visible — so that budget goes to interventions that have a mechanism, not to files named after cats.
Run your first AI Search Audit report for free
If you want to see exactly where your site stands across the four major AI systems, aisearchaudit.ai runs a full citation audit and returns a structured report with the specific gaps to fix first. Check our plans or contact us for a walkthrough.
Featured image: photo by Nothing Ahead on Pexels.
