← Back to Blog

Inside the AI Content Engine Built to Get Your Pages Cited by LLMs

By Marc Sirkin, Chief Growth Officer at Walk West and Founder of Marc Sirkin Consulting ·

People do not just Google a question and click over to your website anymore. They ask an LLM, and the LLM decides which pages to cite. So content teams now have a second question to answer alongside "how do I rank": what makes an AI answer engine pull my page into its answer in the first place?

In this session, Marc Sirkin walked through the connected system he uses to answer that question. It is a back-to-basics content stack built mostly inside Claude: a content creator that never starts from a blank page, an evaluator that scores a page for SEO, readability, and AI visibility, an AI slop detector, and a semantic drift analyzer he built himself. This recap covers how each piece works, what actually earns citations in LLMs, and how to decide whether to build a system like this or buy one.

What problem is this content system solving?

Search behavior is splitting. A lot of people now start inside an LLM when they are trying to solve a problem, understand what they are looking for, and only then head back to a traditional search engine. Google still handles the large majority of search volume, but the friction of asking a chatbot is far lower than running a search and sifting results, so the shift is real and moving quickly.

That creates two problems. First, overall clicks from both search engines and LLMs are dropping, and the mechanism for earning a click out of an LLM is not proven yet. Second, you cannot see what people are typing into these tools, so you are optimizing partly in the dark.

Marc's answer is not a clever trick. It is a return to fundamentals: know your ICP, know the problems they are trying to solve, and produce genuinely good content that other people want to talk about. His view is that AEO and GEO (answer engine and generative engine optimization) are layers built on top of SEO, not a replacement for it. If the fundamentals are wrong, no amount of AI-era tactics will save the page.

What are the four parts of the content stack?

The system is four connected tools, most of them running inside Claude so there is no copying and pasting between apps.

Tool What it is What it does
Content Creator A Claude Project of structured markdown files Helps you draft content using your own voice, frameworks, and hooks so you never start from a blank page
Content Evaluator A Claude skill Scores a page for SEO, readability, and GEO, and returns a prescriptive fix list
AI slop detector A prompt-based check Flags AI writing tells and helps rewrite them
SimCheck A Streamlit app he built Measures semantic drift with cosine similarity and maps where a page loses focus

The principles are portable. You can rebuild the same approach in ChatGPT or Copilot, though the implementation differs by platform.

How does the Content Creator work?

The Content Creator is a Claude Project made up of structured markdown files: an ICP file, a profile of the business, a formatting and structures file, a hooks file, a frameworks file, even a video construct. The structures file holds around eight to ten copywriting frameworks, including AIDA and ABT, each with real examples of how the author likes to write.

Two things make it useful. It has a menu system: you type "menu" and it shows every content move available, and each item is anchored to one of the underlying files. And it is programmed not to just draft. It asks a lot of questions first, infers whether you are making a blog post or a short social post, suggests an H1 and H2 skeleton for SEO and GEO, and surfaces flags before it writes a word.

A detail worth stealing: the markdown files were not hand-written. They were generated through an interview process, where the AI asked questions and turned the answers into the files that then fuel the project. That is how you get your actual voice and domain expertise into the system rather than generic output.

Why build it in Claude?

The recurring reason given for anchoring the system in Claude is context retention. Loading a large amount of context (voice examples, frameworks, hooks, ICP) and then asking for the same output across models, Claude held onto the nuance of those input files more reliably than the alternatives tested. Not necessarily a better writer, but better at not forgetting the context you gave it. That matters when the whole point is to steer the model with your own material.

How does the Content Evaluator score a page?

The Content Evaluator is a separate Claude skill you can chain directly from the project with a slash command. It is informed by Google's quality rater guidelines, a Princeton research paper on generative engine optimization, and the work of well-known SEO and GEO practitioners, and it is refreshed roughly monthly as new research lands.

Give it a live URL and it does more than read the words. It fetches the page, pulls the schema, reads the metadata, and inspects the technical scaffolding, because getting cited in LLMs depends on structure (clear H1s and H2s), schema and JSON data, and keyword construction, not just the copy. If there is no live page, the skill has a fallback protocol that scales back to SEO, readability, and GEO dimensions on the content alone.

The output is a scorecard across SEO, readability, and GEO, plus a prescriptive, section-by-section fix list you can hand to a writer or, inside Claude, simply tell it to apply. In one live example it flagged an entity-mapping issue (the brand was named only once on the page), no freshness signal, and no schema plan.

What actually gets you cited by LLMs?

Several practical signals came up across the session, and they line up into a checklist worth keeping:

  • Citations and proof points on the page. If a page has no attributable evidence, it does not get cited. This is one of the biggest lessons of GEO.
  • Passage-level extractability. Machines get confused easily. One point per paragraph beats three points crammed into one. A tightly focused paragraph that answers a specific question is what gets surfaced as the answer.
  • FAQs. Question-and-answer blocks let an AI pull a clean answer for a specific query.
  • Freshness. LLMs reward fresh content, and citations have a short shelf life, coming and going faster than links did in the classic SEO era. Evergreen pages need periodic refreshing: move things around, add new citations. An open question is whether just updating a date counts, or whether you need genuinely new content.
  • First-party data. Original data appears to help because it adds something new to the knowledge graph rather than restating what already exists.
  • Earned mentions and PR. Press releases get picked up quickly because of their format, syndication onto authoritative domains, backlinks, and verifiable facts. The links are nice, but the citation and the mention are what LLMs respond to. The durable version of PR is getting people to talk about what makes your brand genuinely interesting.
  • Discoverable technical documentation. Publicly available docs can be a differentiator. Moving documentation to an LLM-friendly format has been linked to better AI discoverability.

A useful theme underneath all of this: some current wins come from gaming the system, the way early SEO did, but that gap will likely close. What is left when it does is interesting products, a clear ICP, and good content. Which is the same thing search engines have rewarded all along.

What does the AI slop detector catch?

Near the end of the process, Marc runs the draft through an AI slop detector. It flags the familiar tells, the "it's X, not Y" constructions and the rest of the giveaways, and helps decide where to rewrite. Even with a voice and a set of structures loaded, the model tends to ignore some of what it was given, which is exactly why layering multiple evaluation passes matters.

A related tactic that surfaced: deliberately building in variance so the output does not read as generic. That can mean pattern interrupts in the style of classic direct-response copywriting, so a hook or headline lands as "I had not thought of it that way" rather than as recognizable AI slop. Some practitioners go further and preserve 85 to 90 percent of their own words from a recorded transcript, using AI to shape rather than to generate. Others assign numbered copy variations and use a random selection so repeated outputs vary.

What is SimCheck, and why does cosine similarity matter?

SimCheck is a Streamlit app Marc built. It runs a sentence-transformer model (BAAI BGE) and reads a piece of content from a URL or pasted text, then returns a SIM score out of 100 for how ready that content is for AI answer engines, plus a drift map showing where the page loses focus.

Cosine similarity is the concept underneath it, and it is how LLMs work under the hood. The simple version: if someone says "bat" with no context, you cannot tell whether they mean a baseball bat or a vampire bat. Content has the same problem. When we try to make several points in one article, the writing drifts, and the model loses the thread of what the page is actually about.

In the demo, an article about DKIM email authentication was seeded with off-topic lines about favorite conference snacks and a weekend hiking recap. The drift map showed the page getting weak toward the end, with a low score. Removing the noise and re-analyzing pushed the score up and flattened the drift map: one page, one topic, clearly about DKIM. The tool also returns an action plan pointing to the exact off-topic sections to cut or rewrite.

The takeaway connects back to the whole system. Your writing examples, patterns, and hooks are all context you feed the model so it gets closer to what you actually want, instead of hallucinating its way toward the wrong meaning of "bat."

Should you build this or buy it?

There are strong paid tools in this space that scrape your site, recommend topics, and show how you rank against competitors across a set of prompts in each LLM. They are useful, but there is a lot of inference happening, because none of them truly know what people are typing into an LLM. Every company's setup is different, so the recommendations need to be read in context.

The lowest-inference move available today is to build a simple report in GA4 and filter for LLM referral traffic, using a regex like:

.*(chatgpt|openai|perplexity|claude|gemini|copilot).*

Line up the pages that are actually getting traffic from ChatGPT, Perplexity, and the rest, then go read those pages and ask what earned the citation. Is it a proof point, a fresh update, an FAQ? Add an FAQ to a page that lacks one and watch whether traffic moves. That is real evidence rather than inference.

From there, the build-versus-buy call comes back to fundamentals. If you know your ICP and your content pillars, targeted content is straightforward to plan. And because LLMs seem to favor tightly focused pieces, part of the answer to producing at volume is structural: instead of one article covering five things, write five articles each covering one thing.

Key Takeaways for GTM Teams

  • Treat AEO and GEO as layers on SEO, not a replacement. Fix the fundamentals first: ICP clarity, content pillars, clean structure.
  • Write one point per paragraph. Passage-level extractability is what gets a paragraph surfaced as an AI answer. Add FAQs for the same reason.
  • Put proof on the page. No citations or attributable data means no citation from the LLM.
  • Plan for freshness. Citations decay fast. Build a cadence to refresh evergreen pages with new data and sources.
  • Get your voice into the system. Use an interview process to generate voice, hook, and framework files so the AI steers toward your writing, not generic output.
  • Layer your evaluations. A creator, an evaluator, a slop check, and a drift check catch different problems. Any single pass misses things.
  • Measure with GA4 first. LLM referral tracking gives you lower-inference evidence than most paid tools about which pages are actually working.
  • Favor focused content at volume. Five pieces about one thing each tend to outperform one piece about five things.

Want to Attend the Next Session?

The GTM Clawllective is a community of go-to-market operators building and sharing real AI workflows. Sessions run every Friday, and recaps and session content are published to the Collective Hub each week so you can catch anything you missed. Next Friday's session is sales-focused, and everyone is welcome. Join the community to get the calendar and the hub.

About the Speaker

Marc Sirkin is Chief Growth Officer at Walk West and the founder of Marc Sirkin Consulting, with more than 30 years in marketing, media, and copywriting. He was previously CEO of Third Door Media, publisher of Search Engine Land, MarTech, and the SMX conference series, before its acquisition by Semrush. His consulting work centers on rebuilding digital presence across SEO, content, email, and paid strategy with AI-native tools.

He also hosts the "Playbook Broken" podcast and writes on Substack about integrating AI into go-to-market work.

Watch the session replay: Inside the AI Content Engine Built to Get Your Pages Cited by LLMs