llms.txt will not get you cited by AI

There is an llms.txt file on notamize.com. I wrote it in March 2026.

It is a straightforward thing to write. You describe your business in plain markdown. What you do, who for, the languages you work in, the link to book a call. No headlines, no persuasion, no hero section. Just facts an AI assistant can read without having to guess at your navigation.

What I am not going to tell you is that it made AI assistants recommend me. The evidence says it almost certainly did not.

The claim, stated at full strength

The argument for llms.txt is a good one, and I want to put it fairly before I take it apart.

Large language models read the web through crawlers that have to fight your site to understand it. Navigation, cookie banners, JavaScript, a hero headline that says "Grow Faster" and means nothing. Every token spent parsing that is a token not spent on what you actually sell. So you publish one markdown file at /llms.txt that hands the model a clean, token-efficient summary of your business, and the model understands you better. Better understanding, the argument goes, means more accurate descriptions and more citations.

It is a tidy idea. It costs almost nothing. It has a specification, a growing list of adopters, and it sits at the top of most AEO packages being sold right now, usually as the first deliverable and often as the headline one.

I wrote one.

What the studies actually found

Two independent studies ran at meaningful scale in 2026, and both point the same way.

SE Ranking looked across 300,000 domains and found no citation lift. Sites with an llms.txt file were not cited more often by AI assistants than sites without one.

Ahrefs looked at 137,000 sites and found the mechanism behind that result. In 97% of cases, the llms.txt file was never read by an AI crawler at all. Not misread. Not deprioritised. Never requested.

A file that is never requested cannot influence anything. That is the entire argument, and it fits in one sentence.

Then in May 2026 Google published its own generative-AI guidance and said directly that machine-readable files like llms.txt are not needed for AI search. Not "less important than you think". Not needed.

Three things worth being precise about here, because this is the part that gets overstated in both directions.

First, none of this says llms.txt is harmful. It is a static text file. It does not slow your site down, it does not compete with anything, and no study found a penalty.

Second, no citation lift measured across 300,000 domains is not the same as proof that no effect exists anywhere. It means that if there is an effect, it is too small to find in a sample that size. That is a meaningful result, and it is also not a law of physics.

Third, and this is the part that matters most if you are deciding where to spend an hour: the same body of research is clear about what does help. Structured data does. Schema markup is read, documented and acted on. Google tells you what it consumes and gives you a validator to check your work. The contrast is not subtle. One artifact has a specification nobody enforces and a 97% non-read rate. The other has documentation, a testing tool, and visible consequences when you get it wrong.

And if you sell products, neither of those is your main lever. Your product feed is. That is the structured description of your catalogue that shopping surfaces and AI shopping agents actually parse, and it is the one place where wrong data has an immediate, measurable cost.

The part that lands on your ad budget

This is where I think the AEO conversation goes quiet, and it is the part I care about most, because the Google Ads work I do lives inside accounts rather than on content sites.

Avinash Kaushik has been modelling the downstream numbers through his AEO series, and his figures are worth sitting with. He puts the coming traffic reduction at around 30% for organic and 15% for paid. He also reports that ChatGPT referral traffic converts at 16%, against 1.8% for Google organic.

Read those two together, because separately they mislead.

A 9x conversion rate on a channel that is still a rounding error in your sessions does not save your quarter. What it tells you is that people arriving from an answer engine have already done their comparison. The research step happened somewhere you cannot see and cannot bid on. They arrive at the end of the funnel, not the start.

Now think about what that does to a Google Ads account.

Your brand terms get more expensive to defend. If a third of the organic traffic that used to arrive for free stops arriving, that demand does not vanish, it reappears as paid clicks on your own brand. Kaushik puts the brand traffic mitigation cost at a 15 to 20% budget increase. That is not a strategy change, it is the same revenue costing more.

Your search terms get longer. Answer engines are trained on questions, not keywords, and the query behaviour bleeds back into search. Where that happens, phrase and exact match coverage built for two-word keywords quietly stops matching, and the traffic lands in whatever campaign is loosest. Your own search term report will tell you whether it is happening in your account, which is a better source than anyone's impression of the trend.

Answer surfaces have fewer ad slots than ten blue links. Fewer slots at the same demand is a price increase, and it arrives without an announcement.

None of that gets fixed by a markdown file. All of it shows up in your account this quarter.

Where that hour actually goes

In order, and none of them need a budget. The first one matters most, because it is the only item here that can leave a site genuinely invisible to AI assistants rather than merely unhelped. If you want to see how I sequence work like this, it is set out in the process I follow on an account.

1. Check robots.txt before anything else. This is the one place where a site can be actively invisible to AI assistants, and it happens by accident more often than you would expect. Confirm that GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and Google-Extended are not blocked. One note on that last one: Google-Extended controls Gemini grounding and training use, not normal Search indexing, so blocking it does not hurt your rankings. And if you blocked these crawlers deliberately because you do not want your content used for training, keep the block. That is a trade-off you chose, not a bug. Skip this step only if you know when you last read the file.
2. Add Organization and FAQPage schema, then validate it. This is the part with evidence behind it. Organization for who you are, FAQPage mirroring questions clients genuinely ask you. Then run every block through Google's Rich Results Test, because schema that contradicts the visible page can be treated as spam, which is worse than having none. If you use a tag manager you can deploy this today without touching your theme.
3. Run the discovery test yourself. Open ChatGPT, Perplexity and Gemini and ask four things: what does my business do, is it a good fit for my ideal client, how does it compare to a named competitor, and how do I get in touch. Screenshot all twelve answers. This takes fifteen minutes and it is the most useful fifteen minutes in the list, because it stops being theoretical. You find out which facts are wrong, which service you are not associated with at all, and whether you appear in comparisons. Those are fixable problems with known fixes.
4. If you sell products, fix the feed and stop reading about text files. Titles that describe the product rather than pack it with keywords. Complete attributes, real GTINs, accurate availability, correct prices. Your feed is the version of your catalogue that machines read. A perfect llms.txt next to a feed with 40% of attributes missing is a strange way to spend an afternoon.
5. Then write the llms.txt anyway. Twenty minutes, last, expecting nothing from it. Two reasons. It costs nothing and the cost of being wrong about it is one file. And writing it forces you to state plainly what you do, who for, and what you sell, in language with no marketing in it. That is a useful exercise whether or not anything ever reads the file.

What I do not know

One question

If you have an llms.txt live, go and grep your server logs for it. I want to know whether anything ever requested it, and under which user agent.

That single line in your logs would tell us more than either study did. Has anyone actually looked?