On 15 May 2026, Google published its optimisation guide for generative search, followed in mid-June by an unusual clarification, reported on 15 June by Search Engine Land: llms.txt files, presented for a year as the compulsory gateway to ChatGPT and AI answer engines, have no effect, positive or negative, on rankings in Google Search. The GEO market (generative engine optimisation, also called GSO) built a good part of its sales pitch on the opposite idea: without these technical tactics, a brand would vanish from AI answers. That premise is false in its most widely sold form, and five independent studies confirm it. GEO remains a real field. The tactics that work are simply different from those most of the market sells.
The file supposed to prevent invisibility in AI answers does nothing
An llms.txt is a text file at the root of a website, modelled on robots.txt, meant to tell language models which content to favour. Since 2025, part of the consultancy market has presented it as a prerequisite for appearing in ChatGPT, Gemini or Perplexity. Google's official guide, in the section that dismantles common misconceptions about generative search, is unambiguous: no machine-readable file, markup or specific Markdown is needed to appear in Google Search, including its generative features, "as Google Search itself doesn't use them". Keeping the file does no harm either, but it serves no purpose: Search ignores it. Google files this as a clarification rather than a reversal: the position has never changed since Gary Illyes confirmed it back in July 2025.
We may as well own it with a smile: aixh.com keeps its own llms.txt, in full knowledge of the facts. It brings nothing in, Google has just confirmed as much, but it costs nothing to maintain, and part of our job is to test ourselves what we advise clients against paying for. It is the doormat of the site: nobody ever wipes their feet on it, but the entrance looks tidy.
Other "essentials" sold by GEO providers rest on the same mechanism, which can be measured in the sector's own figures.
The anatomy of a fear that sells
Kevin Indig (Growth Memo) puts venture capital investment in 80 companies of the AI-search sector at $1.5 billion, including $227 million for AI visibility tracking tools alone: 18 companies selling, broadly, the same mechanism, running scheduled prompts on LLMs and counting brand mentions. Profound, one of the most visible players, has raised $154.5 million in four rounds over eighteen months, up to a $1 billion valuation. Digiday documents a sector in rapid expansion: Edward Cowell (WPP Media) receives a sales pitch on the subject every 24 to 28 hours.
A market growing this fast rarely signals a measured need. It signals, above all, a bubble of marketing spend built on a poorly calibrated worry.
What the data says, vendors aside
When even the tool vendors find nothing
Five independent studies converge on the same verdict on llms.txt, each with a different methodology.
Ahrefs (Louise Linehan and Xibeijia Guan, 15 June 2026) analysed 137,210 domains: 28% publish an llms.txt, but 97% of valid files received no request at all in May 2026. Among the few files that were fetched, less than a fifth of requests came from identified AI tools (GPTBot first among them).
SE Ranking, on around 300,000 domains, measures 10.13% adoption and no significant correlation between the presence of an llms.txt and the frequency of AI citations, to the point that removing this variable from an XGBoost predictive model improves its accuracy: the file added noise rather than information.
Three further analyses point the same way. ALLMO.ai found a single llms.txt URL among 94,614 cited in 11,867 answers from ChatGPT, Claude, Gemini, Grok and Perplexity; of the 50 most cited domains, only one had the file. Search Atlas, which sells AEO tools, measured no advantage across 347 domains. OtterlyAI and Limy.ai find the same near-total absence of AI bot traffic to this file, on volumes ranging from tens of thousands of bot visits to more than 500 million observed events.
Cyrus Shepard (Zyppy) scored 23 factors that influence AI citations from 54 studies, patents and experiments: at the top, URL accessibility and rank in classic search; at the bottom, at 2 out of 10, llms.txt. His conclusion: no credible evidence shows that these files influence AI citations.
"But Google does crawl these files"
The most tempting objection turns the argument against Google: the search engine would have an interest in discouraging tactics that reduce its hold on the web. The figures do not support that reading. Wix AI Search Lab shows that Google does index tens of thousands of llms.txt files, up to around 120,000 in May 2026; being indexed and being used for ranking remain two different things. Google's denial is above all corroborated by players with no stake in the platform war (Ahrefs, SE Ranking, ALLMO.ai, and even Search Atlas, a vendor of AEO tools): when a company paid to prove that a tactic works concludes the opposite, the objection collapses.
GEO itself is real
Dismissing GEO wholesale would be as dishonest as overselling it. The field's founding paper, Aggarwal et al. (Princeton and IIT Delhi, KDD 2024), tested nine strategies on 10,000 queries. The most effective (citing sources, adding verifiable quotations, adding statistics) achieve a relative improvement of 30 to 40% on the position-weighted metric, and of 15 to 28% on subjective impression, with a maximum of 41% for some methods. Keyword stuffing, inherited from classic SEO, does not work. William Spurlock sums up the sector's dividing line well: the sceptics correctly identify the vendors' marketing hype, but miss the real change in how models ingest and chunk content.
The 40% figure is a maximum per method and per domain rather than an average: in real conditions on Perplexity, the gain on the position metric falls to around 22%. These percentages measure visibility within the generated answer. They do not measure revenue.
AI-sourced traffic also converts better than classic search: 4 to 5 times higher for a subscription according to the Washington Post, an order of magnitude confirmed by Microsoft Clarity across more than 1,200 publisher sites (1.34% against 0.55%). A study of 94 e-commerce brands, relayed by Search Engine Land, measures 1.81% via ChatGPT against 1.39% for non-brand organic search. This does not rescue the GEO market's narrative: the volume remains tiny, 1.08% of total traffic on average according to Conductor, and the 94 brands mentioned above generated $474,000 via ChatGPT against $32.1 million via classic organic search. A higher conversion rate justifies taking care of the content fundamentals. It does not justify buying a file nobody uses.
The most readable indicator for checking this effect remains appearance in AI Overviews, the answer Google generates directly within its search results. AIxH tracks this indicator on its own SEO & GSO accounts, before and after the substantive optimisations described above. On several accounts we track, including AIxH's own website, visibility in AI Overviews has been multiplied by 20 since the GEO work began.
The real danger: technical gadgets can cost a lot
The GEO market sometimes exposes companies to a real business risk.
Lily Ray (Amsive) has documented the boomerang effect of aggressive tactics: self-promotional listicles of the "10 best X in 2026" type, designed to rank at the top, suffered sharp drops in organic visibility during a wave of demotions observed in January 2026. A B2B company valued at around $8 billion lost 49% of its organic visibility in twelve days; across a sample of more than 220 sites using AI content tools at scale, 54% lost more than 30% of their peak traffic and 22% lost more than 75%. Since AI answers draw on the same search indexes, losing in classic SEO means losing AI visibility.
Measurement itself is unstable. Rand Fishkin (SparkToro), with Patrick O'Donnell (Gumshoe.ai), ran 2,961 identical prompts with 600 volunteers: fewer than one run in a hundred returns the same list of brands, fewer than one in a thousand in the same order. His factual conclusion is clear: AI rank tracking remains inherently unreliable at the level of an individual query. This instability confirms that a precise "AI ranking", as some tools sell it, is largely statistical noise. It also feeds the worry the market exploits.
GEO also raises genuine governance risks, distinct from the sale of optimisation tactics. A position paper accepted at ICML 2026 documents them: in 2026 Microsoft reported hidden prompts in "Summarize with AI" links, designed to steer assistants towards specific brands, and the OECD AI Incident Monitor records an incident of AI recommendation manipulation in China the same year. These issues belong to AI system security. They have nothing to do with buying an optimisation tactic.
What should you do, concretely?
Demand a primary source before any purchase
Before signing for a GEO tool or service, ask a single question to sort the serious providers from the rest: what is the primary source (Google, OpenAI, or an independent controlled study) proving that this tactic increases visibility? If the person opposite you mentions an llms.txt or "AI-specific" markup, treat it as a warning sign: Google and five independent studies say the opposite. Jeremy Moser (uSERP) offers a simple test: a good provider admits from the outset that 80% of GEO is good classic SEO, and calls the opposite claim "snake oil".
Invest in what has a measured effect
Focus your resources on what scores best in the serious studies (Aggarwal et al., Cyrus Shepard): first-hand expertise, proprietary data, verifiable third-party citations, direct and self-contained answers, comparison tables, editorial freshness, presence on the platforms that models actually read, such as Reddit or Wikipedia. If you run an SME, this work looks very much like well-executed classic SEO. If you manage a large account, add brand authority work and, potentially, content licensing agreements, on the Washington Post model. An AI visibility audit lets you start from the same indicator (appearances in AI Overviews, citations in generative answers) on your own website.
Measure without fooling yourself
Set up tracking of AI referral traffic through UTM parameters or referrer detection, then follow an aggregated frequency of mentions over time on ChatGPT, Perplexity, Gemini and AI Overviews. Never treat an "AI ranking" for an individual query as a reliable indicator: Fishkin has shown that this level of granularity is noise. The most useful decision threshold remains sector-based: if the AI Overviews trigger rate in your sector exceeds 25%, a threshold healthcare clears by a wide margin according to Conductor, with close to 49% of queries affected, GEO deserves strategic priority in your budget. Below 10%, simple monitoring is enough.
AIxH's view
The GEO market sells uncertainty as an immediate danger. The ground it exploits remains young, can be measured rigorously, and rewards fundamentals over gadgets. That is exactly the line AIxH takes as a GSO agency in Luxembourg: optimisations drawn from the studies cited here, aggregated tracking of citations rather than an AI ranking sold query by query, and a GSO method that extends SEO instead of replacing it. Before buying a technical gadget, have someone measure what AI answers already say about your brand: it is the only starting point that does not rest on fear.