AI assistants have become a real source of visitors, and most analytics tools still cannot tell you how many of them you get. Adobe reported that AI-referred traffic to US retail sites grew 138% year over year in May 2026, and that AI-referred visitors now convert 54% better than visitors from other sources. If a channel is growing that fast and converting that well, you want to see it. The problem is that a large share of it arrives with no source attached, and lands in your reports as “direct”.
This post explains why that happens, what an llms.txt file can and cannot do about it, and how a measured link closes the part of the gap that matters most: knowing when someone acted on what an assistant told them.
Why AI traffic hides in “direct”
Web analytics decides where a visit came from by reading two things: the referrer, which the browser sends to say which page the visitor left, and any tracking tags in the address, such as utm_source. When neither is present, the visit is filed as direct.
AI traffic loses those signals often. Links opened from an assistant’s mobile or desktop app frequently carry no referrer. Links copied out of an answer and pasted into a message or a browser carry none either. Browser privacy settings can strip the page path or the referrer entirely, and a redirect can drop a tag if the destination does not keep it. OpenAI documents that ChatGPT search adds utm_source=chatgpt.com to the links it shows, which helps, but only for that product and only when the tag survives the trip.
The result is a channel that is growing quickly and is partly invisible. You see a rise in direct traffic, and you cannot say how much of it came from an assistant recommending you.
What llms.txt is, and what it is not
llms.txt is a plain Markdown file published at the root of a website, proposed by Jeremy Howard in September 2024. It gives AI agents a short, curated summary of a site: who you are, what you offer, and links to the pages that matter, so an agent does not have to guess from your navigation.
It is worth being clear about its limits. It is a proposed convention, not a standard that AI providers have committed to. Google’s Search team has said repeatedly that it does not use the file, and its May 2026 guide to AI features lists llms.txt among things site owners do not need. At the same time, Chrome’s Lighthouse 13.3 added an experimental audit for it under agentic browsing, the AI labs publish one for their own developer documentation, and thousands of sites now do the same.
So the honest position is this: llms.txt costs little, may help agents and assistants that read it describe you accurately, and is not a ranking factor. Publish it for clarity, not as a growth hack.
Where a measured link fits
Here is the useful part. Whatever an assistant reads about you, the moment that matters commercially is when a person acts on it: downloads your app, books a call, opens a quote request, scans through to a product. That action happens on a link.
A short link that counts its own clicks does not depend on referrers or tags at all. It records the visit at the moment the link is opened, before your website or the store is involved, so the count is the same whether the person came from an assistant’s app, a copied answer, or a private browser window.
That gives you a simple way to measure an AI channel: use a dedicated link for it, and publish that link only there.
In practice:
- Keep your page links canonical. The links in
llms.txtthat point to content pages should be the real addresses of those pages. Agents read them to understand you, and some do not follow redirects. Do not replace your content links with short links. - Use a dedicated measured link for the actions. For the few links that represent a decision, such as “download the app”, “book a demo” or “request a quote”, publish a short link that exists only in your
llms.txtand nowhere else. Any open on that link came from something that read the file. - Tag the destination. Put
utm_sourceandutm_mediumvalues in the destination address, so that when the visit reaches your site, your analytics files it under the right channel instead of direct. - Read the counts together. The link tells you how many times the action link was opened and from which kind of device; your analytics tells you what those visitors did next. Neither alone is the full picture.
Two cautions keep the numbers honest. An agent that fetches a link while researching you is a machine visit, not a person, so treat early counts as a signal of interest rather than a count of customers. And a link appearing in an answer depends on the assistant choosing to show it; publishing a link does not guarantee it will be used.
Why the link layer is a security decision too
Once you publish links for machines to read and repeat, the link itself becomes part of your attack surface. Three questions are worth asking of any link platform.
Who can change where a link goes? If someone other than you can repoint a link, an assistant that faithfully repeats your published link can send people somewhere you never chose. Destinations should be changeable only from your own authenticated account.
Who else shares the domain? Short links live on the vendor’s domain. When a widely shared short domain is abused for spam or phishing, security filters and browsers can start treating every link on it with suspicion, including yours. A platform that is monitored for abuse protects its own domain, and with it your links.
What does it add to your stack? A measurement tool that requires a script on your site or an SDK in your app adds code you have to review, update and defend. A link-layer tool adds nothing to either.
Why 1Link for this
TAPQUAD’s 1Link® was built for this layer. Every link belongs to an authenticated 1Link account, and only that account can change its destination. The platform is monitored for abuse, so a link you publish keeps going where you sent it, and it runs on two official domains only, 1Link.app and 1lk.io. There is no SDK to embed and no script to add, so it adds nothing to your codebase or your attack surface. Every link comes with a dynamic QR code and real-time reporting by device and platform, and for apps and games it routes each visitor to the correct store across iOS, Android, macOS, Windows and Amazon Fire.
Behind the app is the TAPQUAD team that runs the platform, so when something looks wrong in your numbers, there are people to ask. If you want to try it on your own llms.txt, start a free 30-day trial: one activated link is enough to measure one action.
Sources: Adobe Digital Insights figures on AI-referred retail traffic, as reported by Digital Commerce 360, June 2026; llmstxt.org; Search Engine Journal’s summary of Google’s guidance on llms.txt and the Lighthouse 13.3 agentic browsing audit; OpenAI’s documentation of utm_source=chatgpt.com for ChatGPT search referrals.