AI SEO in 2026 — What Actually Gets You Cited (and Where llms.txt Really Stands)

I published a version of this post in March 2026 recommending llms.txt as one of three foundations of AI SEO. The evidence since then has not been kind to that advice, so this is a rewrite rather than a touch-up. If you implemented llms.txt on my recommendation, nothing you did causes harm — but you should know what it is and is not doing.

What changed

In June 2026 Google updated its documentation to state plainly that llms.txt has no effect, positive or negative, on Search rankings or AI Overviews. Search simply ignores the file. This followed earlier comments from Google's Gary Illyes that the company does not support llms.txt and has no plans to, and from John Mueller comparing it to the long-discredited keywords meta tag.

The usage data tells the same story. A SE Ranking study across 300,000 domains put adoption at roughly 10%. More tellingly, an analysis of over 500 million AI bot visits across a 90-day window found only 408 requests that targeted llms.txt at all. As of 2026 no major AI provider — OpenAI, Google, Anthropic, Meta or Mistral — has publicly committed to reading or acting on the file in production.

So should you delete it?

No. It costs nothing to serve, it does no harm, and there is a plausible future for it in agentic systems and developer-facing documentation platforms, where a machine-readable index of a site genuinely helps. Keep it if you have it.

What you should stop doing is treating it as an SEO deliverable, paying for it as a line item, or expecting it to produce citations. It is a bet on a standard that has not been adopted, not a ranking factor.

What actually earns citations

The mechanisms that work are less novel and more demanding.

1. Structured data that matches the page

Schema.org markup in JSON-LD has become the primary way AI assistants understand what a page is about — who wrote it, what it sells, what it costs, what question it answers. It is no longer a finishing touch applied after launch; it is part of the foundation. The critical detail most sites get wrong is consistency: your JSON-LD must agree with what a human sees on the page. I wrote about that failure mode separately in Schema Drift.

2. One page, one intent, answer first

An answer engine extracts a passage, not a page. Content that buries its answer under 600 words of preamble does not get quoted, because there is no clean span to lift. Put the direct answer near the top, then expand. Give each page one job rather than three.

3. Real author identity

Experience, expertise, authoritativeness and trust are doing more work than they used to, precisely because generating plausible text became free. A named author with a verifiable track record, consistent across the site and off it, is a signal that synthetic content cannot cheaply fake.

4. Entity consistency

Describe yourself the same way everywhere — same business name, same service names, same locations, in your schema, your copy and your profiles elsewhere. Models build an internal picture of who you are from many sources; contradictions dilute it.

5. Being mentioned, not just linked

Brand mentions across platforms increasingly matter alongside traditional backlinks. Being discussed is itself evidence.

Do not abandon ordinary SEO

The most common overcorrection in 2026 is treating AI search as a replacement for organic search. It is not, at least not yet. Measured organic traffic from Google to websites is down only around 2.5% year on year — a real decline, but nowhere near the collapse the louder commentary implies. Classic technical SEO, fast pages, clean crawling and useful content still carry most of the load, and they happen to be the same foundations answer engines rely on.

The honest summary

AI SEO in 2026 is mostly technical SEO done properly, plus clarity of identity and structure. There is no file you can upload to shortcut it. I would rather correct my own published advice than leave a page up that sends people down a dead end.