Published: 2026-04-16

Forget Google SEO: How to Rank in ChatGPT, Claude and Perplexity

Chapters / key moments (click to jump — plays here on the page)

Craig Hewitt runs Castos.com — 400+ pages of content, strong Google rankings — but he's invisible in ChatGPT, Perplexity, and Claude. The reason: Google and LLMs evaluate and relay information differently. In this video he shares the seven concrete changes his team made to fix that, anchored by a free open-source Claude Code project called SEO Machine.

Source video

"Forget Google SEO: Rank in ChatGPT, Claude & Perplexity" by Craig Hewitt — Watch on YouTube →

Key Takeaways

LLMs and Google read pages differently. Google rewards relevance signals spread across the whole document. LLM scrapers pull from the top of the page down and weight the first clear answer to a query most heavily. A great Google article can fail completely in LLM retrieval if the answer is buried under a narrative introduction.
The #1 fix: answer first. For any best/top/how query, the first one to two sentences must directly answer the question. "AI scrapers pull from the top to the bottom of the page. Don't bury the answer behind the narrative." A long ramp-up intro is a liability in LLM SEO even if it signals quality to human readers.
Craig shows a counterexample from his own site — an article that opens with "YouTube users watch 700 million hours of podcasts" — interesting context, but not an answer. An LLM evaluating that page would move on to a competitor that leads with the direct answer.
SEO Machine is the open-source Claude Code project Craig uses to execute this strategy at scale. It's a skeleton — input your business, product, and customer info, and it generates targeted content. The biggest improvement is the write command, which instructs the model to always answer the question before adding context.
This is practical SEO content strategy that applies to any site that wants LLM visibility: direct answers, structured information, question-first framing. The same principles that make content good for LLMs also make it better for humans who want fast answers.

The Core Shift: Answer First, Context Second

Traditional SEO content is often structured as: hook → context → narrative → answer. This works for Google, which can evaluate the whole document for relevance signals, and for humans who scroll to find information.

LLM retrieval is different: the scraper reads top-to-bottom and surfaces the first clear response to the query. If your answer is in paragraph five, after two paragraphs of background and one of statistics, you lose to any competitor whose answer is in paragraph one.

The rewrite rule Craig uses: "For any best, top, or how query — the first one to two sentences must answer the question." This is what the SEO Machine's write command enforces: it flags the answer-first rule as critical and instructs the model to never open with narrative context when an answer is possible.

SEO Machine — Free Claude Code Project

Craig's open-source SEO Machine project is a Claude Code skeleton designed for content at scale. You provide: information about your business, your product, and your target customers. The project's commands handle the rest — article writing, structure, internal linking patterns, and LLM-optimized formatting.

The project has been in development for 6–8 months with multiple major iterations. It's publicly available (link in the source video description). This is an example of Claude Code being used as the content production engine itself — not just for software development.

Related on OpenClawDatabase

The RALPH Loop: Craig Hewitt's Claude Code Production Workflow — his coding methodology with the same tool
Skills Guide — extending Claude Code for content workflows
Claude Cowork Hub — another Anthropic interface for content workflows

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Answer the question in the first one or two sentences
<strong>The single biggest change described.</strong> The instruction added to the writing command, marked critical: for any "best", "top" or "how" query, the first one to two sentences must answer the question. The reasoning given is mechanical — <strong>AI scrapers pull from the top of the page downward, so do not bury the answer behind a narrative</strong>. The counter-example is his own article opening with an interesting statistic that answers nothing.
Require a TLDR block immediately after the introduction
Before the first H2 body section, a required key-takeaways block. For a listicle that means naming the answer and then the steps: do A, then B, then C. The pattern is answer-in-the-opening plus structured summary, so the first screen carries the whole answer.
Reinforce the rule in the context files, not just the command
The same "give the answer quickly" guidance is repeated in an SEO guidelines document in the project's context directory — so the rule holds whether the model entered through the command or not.
Put author authority above the fold
An author block establishing why this person can speak on the topic, <strong>as high on the page as possible</strong>, rather than buried at the bottom. Named as part of Google's E-E-A-T thinking but applying equally to LLMs. Use a person's name, not the company's.
Research which prompts you should be appearing in
<strong>The most distinctive technique in the video.</strong> A <code>/research-ai-citations</code> command takes a topic and generates around <strong>100 prompts</strong> a real user might type into ChatGPT or Perplexity where your product could plausibly appear — clustered into general recommendations, comparisons, feature-specific, use-case, pricing and migration. You then check manually whether your brand shows up for each. Output is a markdown file with the audit, prompt clusters, results analysis, gap analysis and an action plan.

Gotchas

The whole workflow is packaged as a free open-source Claude Code project skeleton the presenter maintains; he is also promoting his own business throughout. The techniques transfer without the repository.
The citation-research step is explicitly the most manual part — the tool generates the prompts, you check the results yourself.
It tells you where to focus, not how to write better. The presenter is clear it is a targeting tool, not a quality one.
The author-authority point is structural advice about your site, not something the content generator can fix.

← Back to News digest · See also: OpenClaw guide