The llms.txt question won't go away. Last week I got asked about it twice. One was a founder who'd been quoted $1,200 for an "llms.txt optimization package," and the other was an engineer who'd read that GPT-5.6 changed everything and wanted to know if the file finally mattered now. They were hoping for opposite answers about the same file, so let's settle it properly with the evidence we have, including our own.
If you're skimming, the short version is no. llms.txt isn't worth strategy or money after ChatGPT 5.6, and 5.6 actually made the case for it weaker. Spending twenty minutes to create one as a courtesy file is fine by me. Anything past that means real hours on a file with no measured effect. The long version follows, because the reasoning is more useful than the verdict.
What 5.6 actually changed about source selection
A lot of people have this one backwards. When GPT-5.6 landed in August and citation patterns went sideways, a chunk of the SEO world decided ChatGPT must be reading new signals, which might mean your courtesy file finally counted. The tracking data says the opposite happened.
Retrieval after 5.6 got more conservative. Site-scoped fan-out queries went from under half a percent of searches to 17-23%, so the model now more often picks which trusted domains to search before it decides what to read. "Official" turned into one of its favorite words to append to its own retrieval queries. And while retrieved URLs doubled, the set of cited domains consolidated. Trace the mechanism shift yourself and you'll see every part of it pointing the same way, toward indexed pages on verifiable domains, weighted by trust.
So where does llms.txt fit into that machinery? The engines index your pages, and courtesy files aren't what sits in those retrieval indexes. The file can't work as a trust signal either, precisely because anyone can write one, and the fan-out queries don't hit it. It was designed as a friendly map for AI systems that choose to fetch it, while the systems deciding your citations in 2026 assemble their answers from somewhere else entirely.
What llms.txt was actually designed for
The origin story is worth a short detour, because half the confusion you'll run into comes from people arguing with a file that never claimed to do the thing they're annoyed it doesn't do.
The llms.txt proposal came out of the developer-tools world in 2024 as a courtesy map. It's a markdown index at your site root that lets an AI system with a limited context window find your important pages without crawling everything, a README for robots, more or less. For documentation sites, where an agent really does benefit from "here are the twelve pages that matter, skip the changelog," that was a sensible, modest idea, and it still is.
What happened next is classic SEO folklore. The file was easy to create, your competitors could see whether you had one, and it sat right next to a scary trend, which is how a tactic goes viral whatever its effect. Within months "add llms.txt" was on every AI-SEO checklist as though it were a ranking submission form. Agencies put it in their audits, and the proposal's humble scope got buried under promises its own authors never made.
The file never failed at the job it was built for. The job it fails at is the one the checklist sellers handed it, so the real question was never "does llms.txt work?" It's "does a courtesy index for context-limited agents affect citation selection in retrieval-augmented answer engines?" Once you phrase it that accurately, the answer stops being surprising.
The evidence pile, honestly stacked
Before you decide, it helps to see how much evidence there actually is on each side.
The case against starts with our own test. When we ran llms.txt across 12 sites for 90 days, we measured no citation lift at all, on any engine. Google has said flatly that it doesn't use the file. The adoption studies keep finding the same gap, too: sites add the file, crawler logs show the big engines mostly not fetching it, and citation behavior doesn't move. That's a first-party test and the largest engine's own statement, with log data backing both.
The case for it is that the file costs you almost nothing and a long tail of smaller agents and tools really does fetch it when it's there. Some developer-facing crawlers read it, and someone's homegrown research agent might use yours this afternoon. That's real, and it's the whole case.
The kindest thing I can say about llms.txt is that it's a lottery ticket that costs you twenty minutes. The harshest is that an industry sprang up selling lottery-ticket optimization for it. Any tactic that's easy to deliver and impossible to measure ends up with someone invoicing you for it, and that industry is a preview of GEO spam generally.
200+ SaaS teams already track their AI citations.
They know exactly when ChatGPT mentions their brand, and when it stops. Do you?
So: worth it?
My actual answer comes in three tiers.
Creating one, once, is fine. It takes twenty minutes to write a tidy little index of what your product is and where the important pages live, and it might help some agent someday. It won't move your citations, and you shouldn't expect it to. I file it mentally next to a favicon, correct to have and silly to build a strategy around.
Maintaining one as a priority is where I say no. If updating llms.txt shows up in your team's quarterly goals, something has gone wrong with the quarter. The file describes your site to systems that mostly don't read it, while the systems that decide your visibility read your actual pages, and that gap is the opportunity cost the next section is about.
Then there's paying for llms.txt optimization, and there I'd tell you to refuse outright. A $1,200 package built around a file with no measured effect is a tax on your anxiety, whatever the invoice calls it. I'd go one step further: a vendor who leads with llms.txt has just told you how they choose deliverables, and effect size isn't the criterion.
Spend the afternoon where the engines actually read
The fair comeback is "fine, then what do I do with the twenty minutes and the anxiety?" I'd make three trades with that time, and you can measure the mechanism behind each one.
The first is rewriting your openings. Most ChatGPT usage answers from your title and roughly the first 200 characters of body text, without ever opening the page. An opening that states the answer plainly is worth more than any courtesy file could theoretically be, because it's literally the text the answer gets built from. You can get through twenty pages in one afternoon.
Next, verify access instead of describing it. An llms.txt that says "welcome, AI systems" does nothing for a site whose CDN bot protection is challenging every AI fetcher. Fetch your key pages as the AI user agents and confirm the content comes back, because that one check finds more lost visibility than every courtesy file ever written.
The third trade is fixing what trust-seeking retrieval actually checks, which means consistent entity facts across your site and profiles, named authors and parseable pricing. Post-5.6 machinery verifies brands against corroboration, and that's the layer where an afternoon compounds.
We watch all of this from the measurement side every week, and I have yet to see a citation shift that traced back to a courtesy file. I've seen plenty trace back to an unblocked crawler or a rewritten opening, and some to a review profile that finally agreed with the homepage. The boring fixes keep winning.
Answering the boss who read about it
If you're asking about llms.txt, odds are you aren't really asking about the file. Usually one of three other questions is hiding underneath it.
Sometimes it's cheap-win anxiety: "is there a twenty-minute thing we're missing while competitors get cited?" That's a fair worry aimed at the wrong file, since the twenty-minute thing you're missing is almost certainly the opening-paragraph rewrite or the CDN check above. Other times it's audit theater, where an agency deliverable listed the file as missing, in red, and now you need an answer. And sometimes your boss has just forwarded a LinkedIn post with "should we be doing this?"
For that last one, here's a reply that works, and you're welcome to copy it: "We evaluated it. The largest engine says it doesn't use the file, a 12-site 90-day test measured no effect, and the retrieval systems that decide our citations read indexed pages, not courtesy files. We added one anyway since it costs nothing, and we spent the real effort on openings, crawler access, and entity consistency, which are the levers with measured effects." It's short, every word of it is true, and it turns the anxiety into the work that pays.
How we'd know if this verdict expires
A "no" in this space has a shelf life, so I'd rather state the expiry test than pretend this verdict is permanent. It flips if any major engine documents the file in its retrieval pipeline. It also flips if fetch logs at scale start showing the big crawlers reading it and citation changes correlating with its contents. Either would show up within weeks in crawler logs and tracking data, and neither has happened through two model generations and a retrieval rewrite.
So we keep re-checking, the way everything in AI search gets re-checked: on a schedule and against data, with the weekly per-engine tracking doing the watching. If the file ever starts to matter, you'll read the update here. I just wouldn't hold your afternoon open for it.
How often does ChatGPT mention your brand?
Most founders have no idea. The answer might surprise you.
The one-line answers
Is llms.txt worth it after ChatGPT 5.6? As a twenty-minute courtesy it's harmless, and as a strategy it's a no. As a paid service, run from it. And if what's really on your mind is "how do I know what's working," that's a tracking question, and no static file was ever going to answer it.
15 hours a week manually. Or 15 minutes with RankControl.
Track citations, monitor competitors, and fix content gaps across every AI search engine. Automatically.
