llms.txt for AI Search: Does It Actually Move the Needle?

A single plain-text file sitting quietly at the root of a website has become one of the more argued-over ideas in SEO circles this year. Anthropic’s Claude reads it. Perplexity uses it to decide which pages deserve priority. Google’s own search team insists it makes no difference to AI Overviews, while a separate Google team just shipped a browser tool that flags sites for not having one. That file is llms.txt, and Singapore businesses reviewing their AI search strategy for 2027 are increasingly asking whether it belongs on their to-do list.

The honest answer sits somewhere between the enthusiastic case studies and Google’s shrug. llms.txt is neither a switch that unlocks AI visibility nor a waste of an afternoon, and treating it as either tends to lead Singapore SMEs and their marketing teams astray. Before deciding whether to build one, it helps to understand what the file does, who is reading it, and what has happened to sites that have already tried it.

What Is llms.txt, Exactly?

llms.txt is a simple Markdown file, published at yourdomain.com/llms.txt, that gives AI systems and browser agents a curated summary of a site: what the business does, its most important pages, and where to start reading. The idea, proposed by developer Jeremy Howard in late 2024, was meant to save large language models the trouble of crawling an entire site to work out its structure, much like a sitemap helps traditional search engines find pages faster. Two years on, the format has moved from a niche developer convention to something adopted widely across technology, SaaS and publishing, even as adoption in more conservative sectors such as finance and legal stays under 10 per cent among top domains.

Who’s Actually Reading It in 2026?

Support for the format is real, but uneven, and that unevenness is the first thing to understand before investing any effort. As things stand: Anthropic’s Claude and Claude.ai have publicly confirmed they read this kind of file when navigating a site; Perplexity has confirmed the same, using it to help prioritise which pages it pulls into an answer; OpenAI and Mistral show correlated shifts in citation patterns after one is published, though neither has confirmed direct use of it; and Google Search has stated plainly that it plays no part in ranking or AI Overviews.

That last point deserves more attention than it usually gets, because conflicting signals from Google’s own teams run deeper than a single quote suggests. Google’s Search Advocate John Mueller has compared the file to the old keywords meta tag, a signal search engines quietly stopped reading years ago. Yet Chrome’s Lighthouse tool added an Agentic Browsing category in May 2026 that specifically checks for one at the domain root, on the basis that browser-based AI agents may otherwise waste time working out a site’s structure. One Google product treats the file as irrelevant to search visibility; another flags its absence as a technical gap. For a Singapore business deciding where to spend its SEO budget, that split deserves to be named plainly instead of assuming Google speaks with one voice.

What the Evidence Actually Shows

The clearest answer so far comes from a proper before-and-after study, not a survey of opinions. A controlled look at ten live websites across finance, SaaS, ecommerce, insurance and pet care tracked each for 90 days after publishing the file, measuring AI crawl frequency and referral traffic from ChatGPT, Claude, Perplexity and Gemini. Eight saw no measurable change, one declined, and the two that grew did so for reasons unrelated to the file itself: a neobank’s 25 per cent lift tracked back to a banking licence announcement and rebuilt comparison pages, while a SaaS platform’s smaller gain followed the release of 27 downloadable planning templates. It behaved, in other words, much like a sitemap: useful plumbing that documents what already exists, not something that manufactures visibility on its own.

When Building One Still Makes Sense

None of this means the exercise is pointless, only that it is not a growth lever. It tends to earn its keep on sites where a machine genuinely benefits from a shortcut: developer-facing SaaS platforms with deep documentation, marketplaces with thousands of product variants, or multilingual sites where an AI agent might otherwise struggle to find the right regional version of a page. A business expecting meaningful referral traffic from Claude or Perplexity specifically, or one that wants to be ready for the browser-based agents Chrome is now auditing for, can treat this as a low-cost way to cover that base.

For most Singapore businesses, though, the more useful question is whether the fundamentals underneath are solid first. A curated summary cannot compensate for a site that AI crawlers and search engines alike struggle to read. If your crawlability and indexing setup still has gaps, closing those comes first, since pointing AI systems towards broken navigation or thin pages will not change how those pages actually perform.

Where This Leaves Your Site

Treat this as routine technical hygiene rather than a priority project. A basic version takes an afternoon to draft, costs nothing to maintain, and will not hurt anything if the major AI platforms never end up using yours. What actually shifts outcomes for Singapore businesses remains the same as it always has: clear, well-structured content that answers real questions, sitting on a technically sound site. If you would like a second opinion on where your search and AI visibility priorities should sit, feel free to get in touch with our team.