To show up in ChatGPT and Perplexity answers, you optimise for retrieval, not ranking: these engines pull passages that directly, cleanly, and self-containedly answer the question a user asked, then cite the source they lifted from. That means answer-first paragraphs, question-shaped headings, schema markup, crawler access, and topical depth. This post is deliberately meta — the blog you are reading is built on exactly this structure, so every technique below is one I am running, not theorising about.
How AI engines actually surface a source
Traditional SEO targets a ranking algorithm: hundreds of signals produce an ordered list of pages. Generative engines work differently. When ChatGPT browses or Perplexity answers a query, the pipeline is retrieval-shaped: fetch candidate documents, extract the passages most relevant to the question, and synthesise an answer with citations to the passages actually used. Google's AI Overviews sit somewhere between the two, but the citation step is still extraction-based. The practical consequence: you are not competing to be the best page on a topic, you are competing to contain the best passage for a specific question. A mediocre domain with a crisp, self-contained answer can get cited over an authority site that buries the answer in narrative.
Step 1 — structure content to be extractable
Put the answer first. If a heading asks a question, the first sentence under it should answer that question completely, in a form that survives being quoted out of context — the paragraph you are reading sits under a how-to heading and opened this post by answering the title directly. Burying the answer under 300 words of throat-clearing intro is the single most common reason decent content never gets cited: the retrieval step scores passages, and a passage that gestures at an answer loses to one that states it. Frame H2s as the actual questions people ask, keep one idea per section, and make each section quotable on its own.
Step 2 — use schema markup deliberately
FAQPage and HowTo structured data are machine-readable shortcuts to exactly the question-and-answer pairs these engines want to lift. A FAQPage block hands the crawler a clean question string paired with a concise answer string — no extraction guesswork required. HowTo schema does the same for step sequences. Article and BreadcrumbList schema round out the picture by clarifying what the page is and where it sits in the site. Schema does not guarantee citation, and anyone claiming a secret schema formula is selling something. What it verifiably does is remove ambiguity about what your content contains, which is the whole game in a retrieval pipeline. Every post on this site ships FAQPage schema generated from the same FAQ content you see rendered on the page.
Step 3 — make sure AI engines can actually crawl you
ONE TACTIC A WEEK
None of this matters if the bots are blocked. Check your robots.txt for the relevant user agents: GPTBot (OpenAI), PerplexityBot, Google-Extended (Google's AI training signal), and ClaudeBot (Anthropic). Many sites blocked these wholesale during the 2023-era backlash and forgot. The tradeoff deserves an honest framing: allowing these crawlers means your content can be used in AI answers, which brings citations and brand visibility but can also answer the user's question without a click. Blocking them protects nothing if your goal is to be found. For a business site whose content exists to attract clients, allowing them is the coherent choice — invisibility in AI answers is a real cost, and it grows every quarter as more search behaviour moves into these interfaces. Also verify your content renders without JavaScript; retrieval crawlers are not guaranteed to execute your client-side rendering.
Step 4 — build topical depth, not isolated pages
One good page is a lucky citation; a linked cluster is a durable one. Engines infer subject-matter depth from the neighbourhood a page lives in: internal links between related posts, consistent terminology, and coverage of adjacent questions all raise the odds that your domain is retrieved for the topic at all. This post links to related pieces on tracking setup and agent evaluation because they are genuinely related — but that linking also builds the cluster structure that lets an engine cross-reference answers across the site. Write the cluster before you expect results from any single page in it.
Step 5 — verify you are actually being cited
There is no Search Console for AI engines yet, so verification is manual. Ask ChatGPT (with browsing) and Perplexity the questions your pages target and see which sources they cite — run the checks in fresh sessions so your history does not bias retrieval. Watch referral traffic for chatgpt.com and perplexity.ai referrers in your analytics; both send measurable clicks when they cite you. Log the queries you win and lose over time; citation patterns shift as these engines update, and losing a citation you previously held is the signal to look at what changed. It is crude, but it is honest data, which beats a vendor dashboard scoring you against invented metrics.
What doesn't work — GEO myths
Keyword stuffing 'for the AI' fails for the same reason it failed for search: retrieval scores semantic relevance, and stuffed passages read worse, score worse, and get extracted less. Spinning out hundreds of thin 'AI-optimised' pages fails because retrieval favours the one page with the complete answer over ten pages with fragments. And claims of guaranteed AI rankings should end the conversation — these engines do not publish ranking factors, their retrieval behaviour changes without notice, and nobody outside those companies can promise placement. The unglamorous truth: the same qualities that made content rank well — direct answers, real expertise, clean structure, crawlability — are what get it cited. GEO is a shift in mechanics, not a loophole.