Someone showed me a list of 14 changes they were making so AI assistants would start citing their site. Schema on every page. An FAQ block under every article. Headings rewritten as questions. An llms.txt file at the root.

Every item on that list was about being easier for a machine to read. Not one was about having something worth reading.

That’s the whole problem. An assistant cites a source when it can’t answer the question by itself. If your page says what 10,000 other pages already say, there’s no reason to cite you over any of them.

The short version

  • Assistants cite out of need. Neat markup gives them none.
  • Google itself says extra markup and llms.txt-style files won’t win you a spot.
  • 1% clicked through from Google’s AI summaries. The real prize is being named.
  • Your own costs and failures get quoted. One question tells you if a page will.

A citation is an admission

A language model answers from what it already holds. It reaches for a source when the answer needs something it doesn’t have: a figure that changed, a first-hand account, a claim contestable enough that it needs attribution.

A citation isn’t a reward for good formatting. It’s the system admitting it needed you.

Which makes “how do I get cited” a much less comfortable question than it looks. The answer is: hold something it can’t produce.

Google says the checklist doesn’t work

Google’s own documentation on AI features and your website is direct about this. “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.” And: “There’s also no special schema.org structured data that you need to add.”

Google’s May 2025 guidance on performing well in its AI features comes at it from the other end. The instruction is to provide “unique, non-commodity content.”

Non-commodity. That word is carrying the entire sentence.

Structured data still has real jobs. It gets you a review star, an event date, a price in a product listing. What it has never done is make a generic paragraph worth quoting. Marking up an answer nobody needed doesn’t create a need for it.

The tactics stay popular for an obvious reason. They’re finishable. You can install schema on a Tuesday and tick it off. You can’t tick off “have a number nobody else has.”

The prize is smaller than you think

Pew tracked 68,879 real Google searches in March 2025. When an AI summary appeared, 1% of visits clicked a link inside it. I’ve written about what that does to traffic in Search traffic is falling. Here it matters for a different reason.

The value of a citation was never the click. It’s being the source an answer is built on: named, in front of someone who is being told the answer rather than hunting for it. That’s a credibility outcome, not a traffic one.

It’s worth having. It isn’t worth 14 tasks of markup.

What you own that a model doesn’t

If you run a business, you’re already sitting on the exact category of thing a model can’t generate. You just don’t publish it, because it feels too specific to matter or too internal to share.

Your own numbers. What it actually costs you to win a customer in your category. Your real return rate. How long the job takes against what you quoted. A model can produce an industry average from memory. It can’t produce yours.

Your own failures. The supplier change that cost you a season. The pricing model you rolled back after 3 months. Nobody publishes these, which means nothing has trained on them, which means there’s no alternative source to cite. I’ve published a few of mine. They’re the pages that get quoted back to me.

Your own process. Not the tidy version for the website. The actual sequence, including the awkward step everyone in your trade knows about and nobody writes down.

A position someone could argue with. If nobody could disagree with the page, it isn’t a claim, it’s a summary. And a summary is precisely what the model was going to write on its own.

Publish those and the mechanics take care of themselves. A page with a number in it that exists nowhere else is a page an assistant has to point at, because the only way to use the number is to say where it came from.

The test

One question before anything goes out. Could this page have been written by someone who has never done the work?

If yes, it doesn’t deserve a citation, and it isn’t going to get one. No amount of markup fixes a page whose only problem is that it’s ordinary.

A model doesn’t need another explanation of what you do. It needs the number you’ve never told anyone.

Sources: Google Search Central, “AI features and your website”, last updated 10 December 2025 (https://developers.google.com/search/docs/appearance/ai-features). Google Search Central Blog, “Top ways to ensure your content performs well in Google’s AI experiences on Search”, 21 May 2025 (https://developers.google.com/search/blog/2025/05/succeeding-in-ai-search). Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 July 2025 (https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)