How AI agents decide what to recommend — and what creators need to know before it's too late to matter

Try this. Open Perplexity or ChatGPT and search for something you know well — a topic you've covered in depth, written about, answered for clients dozens of times. Look at what it gives back. Then look at the sources cited at the bottom.

Those citations are what's replacing the first page of Google results for a growing slice of the people who used to find you through search.

The mechanism behind that citation layer has a name: WebMCP. Most content creators have never heard of it. Most SEO professionals are just starting to pay attention. The rules are being written right now, and that's exactly when it's worth understanding what you're looking at.

Here's what WebMCP actually is, in plain terms.

An AI agent — the kind that browses the web to answer a question on a user's behalf — needs a systematic way to find and fetch content. WebMCP is a specification for how that works: structured signals that tell an AI agent what's available on a domain, how the content is organized, and how it can be used. Think robots.txt, but designed for machines that are trying to answer questions rather than index pages.

SEMrush covered this in their 2025 piece "WebMCP: What It Is, Why It Matters, and What to Do Now." Their primary audience is technical SEOs managing enterprise content portfolios. The underlying point applies at every scale: if you publish content that could be cited by an AI agent, this is the infrastructure layer where that citation happens or doesn't.

The structural difference between being cited by an AI agent and ranking in a search result is worth getting specific about.

When you rank in Google, the user sees a list of results. Even when they don't click — and Piece 1 of this series covers how often they don't anymore — they at least saw your name. The brand impression exists even without the click.

When an AI agent answers a question, the user sees an answer. The sources are footnotes, and depending on which interface they're using, they may never scroll to them. If your content informed the answer but isn't cited, you contributed to someone's knowledge and got nothing. If you are cited, you get a name-check that a percentage of users will follow.

That's a different game than ranking. Less about discoverability, more about whether the AI considers your content a reliable source worth surfacing. The criteria aren't fully public, but the signals that appear to matter: clear authorship, factual grounding, specific sourced evidence, consistency over time. Long-form, well-sourced, opinion-with-evidence is the format most likely to earn citation. Short FAQ content designed to rank for high-volume informational queries is the format most likely to be replaced by the answer rather than cited as the source of it.

This is where independent publishers have a structural advantage that's easy to miss.

If someone can summarize your piece in four sentences without losing anything important, an AI agent will. The information gets used; you don't get credited. But if your piece contains something that genuinely can't be compressed — a specific data point, a documented test result, a first-person account, a named example no one else covered — the agent has to either cite you or leave it out entirely.

The HouseFresh case from Piece 1 is the clearest example of this. They didn't just recommend air purifiers. They tested them in a real home, photographed the setups, measured filtration efficiency across units. That's content an AI agent has to cite or ignore. It can't reproduce it. The Forbes roundup written by an editor who never touched the products can be synthesized into a paragraph. The HouseFresh documentation of actual test results can't.

The independent creators who are well-positioned for the citation layer are not necessarily the ones with the biggest audiences or the most traffic. They're the ones whose content is specific enough to be irreplaceable.

One honest note the technical coverage tends to step around.

WebMCP is a spec, not a deployed feature. Whether ChatGPT, Perplexity, Claude, and Gemini's web features implement it the same way — or at all, on any particular timeline — is not settled. Publishing a WebMCP manifest on your domain today doesn't guarantee citation tomorrow. The spec may evolve. Implementations will vary.

The value in understanding this layer isn't a checklist you complete and forget. It's positioning. The creators paying attention now are making the same bet the newsletter-first wave made in 2020: that understanding the infrastructure shift before it calcifies is worth something. That bet paid off for the people who built email lists while organic search still looked easy. Not a guarantee the same thing happens here. But the cost of positioning correctly is low, and the cost of being invisible in the next layer when it does settle is high.

This is still early enough that knowing what WebMCP is counts as competitive advantage.

If you want to go deeper than an article can take you on what this means for your specific content — what you've got that's actually citation-worthy versus what's at risk — grab my inbox. It's specific enough that the general version only gets you so far.

Sources: SEMrush, "WebMCP: What It Is, Why It Matters, and What to Do Now"; Moz technical coverage of AI agent citation infrastructure; SparkToro 2024 zero-click research.