Shopify took over
llms.txt. Almost nobody
lost anything real.

Shopify claimed the
path, and every custom
file stopped resolving.

In late May 2026 merchants who had put a custom llms.txt at the root of their Shopify store found it no longer resolved. Shopify had begun serving the path platform-wide and redirecting it to agents.md. A thread titled, with some feeling, that Shopify had killed llms.txt for all stores ran from 24 May into mid-July.

The practical fix arrived quickly. By 27 May the community had established that a templates/llms.txt.liquid template restores control over what gets served at that path, which is a normal Shopify pattern and takes a few minutes.

So the operational story is short: the path was claimed, a template workaround exists, and if you had a custom file you can have one again. What makes this worth more than a changelog note is what the threads surfaced while arguing about it.

One thing to flag for anyone auditing a store: the redirect means a request to /llms.txt now returns Shopify's response rather than a 404, so a checklist that tests only for the presence of the path will report a pass on every Shopify store whether or not the merchant has done anything. If you are assessing this, fetch the content and read it rather than checking the status code.

A sensible proposal
that got adopted faster
than it got validated.

llms.txt was proposed as a convention rather than a standard: a markdown file at the root of a site giving a language model a curated map of what the site is and which pages matter. The analogy everyone reached for was robots.txt, and the analogy is where the trouble starts.

robots.txt works because the crawlers it addresses agreed to honour it, and that agreement predates almost everyone reading this. A convention only functions when the consuming party has committed to reading it. llms.txt was adopted broadly on the reasonable expectation that model providers would follow, and that commitment has not arrived in any form you can point at.

That gap produced a familiar sequence. The proposal was covered as a best practice, agencies added it to audit templates, tools shipped generators for it, and it appeared on a large number of sites within months. None of those steps required evidence that anything consumed the file, because each step was reacting to the previous one rather than to data.

It is worth saying that the file may still turn out to be useful. Conventions do get adopted, sometimes years later, and the cost of serving one is close to zero. The error is not having the file. It is the confidence with which its effect has been asserted, and the budget that has followed the confidence.

Everyone argued about
access. Almost nobody
asked if it mattered.

The most useful contribution in either thread came from a merchant who had been measuring rather than theorising. His observation was that hits on the llms.txt file were real but very small, and that he had seen no evidence content in that file showed up in citations, as opposed to what assistants pull from rendered HTML and structured data.

That is the whole question, and it went largely unanswered while the thread filled with workarounds. A file being fetched is not the same as a file influencing an answer. Crawlers fetch a great deal they do not use, and the gap between fetch and citation is exactly where an unproven tactic hides.

It is worth being fair about why the idea is appealing. llms.txt proposes something genuinely reasonable: a curated, machine-readable summary of what a site is and what matters on it. If assistants used it as intended, it would be a clean solution. The problem is that adoption by the model providers is not established, and the file has been widely deployed on the assumption that it will be.

The pages that get cited
have a shape, and it is
not a text file.

This does not have to be argued from theory, because AI citation became measurable during 2026. Bing Webmaster Tools reports citations from Microsoft Copilot and partners, including the grounding queries that produced them, and Microsoft Clarity reports the share of sessions arriving from AI referrers.

On this site, over the three months to early September 2026, that instrumentation recorded roughly 6,800 citations across 117 distinct grounding queries, with AI-referred sessions running at about 11% of total sessions. None of that is attributable to an llms.txt file, because this site does not serve a curated one. The pages being cited are ordinary HTML articles.

The shape of the cited pages is consistent enough to describe. They carry a dated comparison or benchmark table with named entities in it, a direct answer near the top rather than buried, structured data that matches the visible content, and specific figures rather than ranges of adjectives. The single most-cited page on this site is a platform comparison built exactly that way.

The last row is not an argument against having the file. It costs almost nothing and may matter later if adoption changes. It is an argument against treating it as the AI visibility project, which is how it has been sold.

For a store, the real
constraint is that your
product pages say nothing.

There is a merchant-specific version of this that matters more than any file at the root. When an assistant decides which particular product to name in an answer, it works from the product page. A two-line description gives it nothing to work with, and a large share of Shopify catalogs are exactly that.

One contributor to the AI search discussions put the number on his own catalog: two thirds of product descriptions under eighty words. That is not unusual, it is close to the median for a store that grew quickly, and it is the actual bottleneck. No amount of curation at the root compensates for a catalog that does not describe itself.

The work is unglamorous and it is the work: materials, dimensions, fit, use case, what it is not for, and the specific comparison a buyer is actually making. Written for a person, which is also what makes it legible to a model. The analysis of which products surface in ChatGPT covers what separates the ones that get named from the ones that do not.

There is a scale objection to this, and it is fair: a catalog of four thousand SKUs cannot be rewritten by hand. It does not need to be. Citation follows the products people actually ask about, which is a small fraction of any catalog. Pull the twenty products that carry most of your revenue and the twenty that get the most search impressions, accept the overlap, and write those properly. That is a week of work rather than a quarter, and it covers the products an assistant is realistically choosing between.

The mismatch between
your schema and your
page is the real bug.

If the citation evidence points at rendered content and structured data, it is worth being specific about what usually goes wrong there, because it is rarely absence and nearly always disagreement.

The common failure is structured data that no longer matches the page it describes. A price in schema that the theme has since overridden. An availability value that reflects the product template rather than current inventory. A review count from an app that was uninstalled two years ago. A product name in schema carrying a variant suffix the visible page does not show. Each of these is invisible on the page and legible to every parser that reads it.

A parser that finds a page contradicting itself has a reason to prefer a different source, and it does not tell you it did. This is dull maintenance work with no launch attached to it, which is precisely why it accumulates, and it is worth more than any file at the root. The search approach for brands covers how this sits alongside the rest of the technical surface.

Stop guessing. Three
reports make this
observable this week.

The reason this topic is full of confident claims is that most people discussing it have no measurement. That changed during 2026 and the tooling is free.

1. Bing Webmaster Tools, AI Performance. Citation counts, the grounding queries behind them, your citation share against other domains, and which of your pages are being cited.
2. Microsoft Clarity, AI Visibility. Share of authority, and the proportion of sessions arriving from AI referrers, which is the number that tells you whether any of this is traffic yet.
3. Google Search Console, generative AI features. Impressions in AI surfaces, broken down by page, which is the closest thing to a control group against the Bing data.

Run all three for a month before adopting any tactic, including the ones in this article. Across the sites I have looked at, the grounding queries alone reframe the strategy, because the questions assistants are answering with your content are rarely the keywords you targeted.

For the analytics-side version of this, measuring AI visibility in GA4 covers referrer configuration, and the answer engine optimisation breakdown covers the structural work that follows once you can see the data.

Keep the file. Do not
mistake it for the
strategy.

A reasonable position on llms.txt, given what is currently known, is roughly this. Serve one through the template, keep it accurate and small, spend an hour on it, and never think about it again until there is evidence it does something.

Then spend the time you were going to spend on it upstream. Product descriptions that describe the product. An answer near the top of every page that would otherwise bury it. Structured data that agrees with the visible content rather than contradicting it, which is a surprisingly common failure. Dated figures instead of adjectives.

For agencies and consultants reading this, there is a commercial note attached. AI visibility is being sold right now at prices that assume the tactics are proven, and most of them are not. The defensible version of the offer is instrumentation first, then the structural work the citation data actually supports, with the file included because it is cheap rather than because it is effective. It is a less exciting proposal, and the one that survives a client asking what changed.

That list is unexciting and it is what the citation data supports. The comparison of generative engine optimisation against traditional SEO covers how much of this is genuinely new work rather than good SEO with a new label, and the honest answer is less than the category would like.

The broader lesson from this episode is about how the tactic spread. A plausible file format was proposed, widely adopted, sold as a service, and argued about at length, without anyone establishing that the intended consumers read it. The instrumentation to check arrived afterwards. When a tactic and its measurement arrive in that order, treat the tactic as a hypothesis, and for anything touching how agents actually reach a catalog, check what is being read before optimising what you serve.

Questions merchants ask
about llms.txt and
AI visibility.