SEO for AI search: how to optimize for citations instead of clicks

Executive summary

Search is splitting in two. Classic SEO competes for a blue link and a click on a results page. Answer engines such as ChatGPT Search, Perplexity, and Google's AI Overviews work differently: they retrieve a set of pages, then compose a single response that cites a few of them. If your page isn't among the cited sources, you don't exist in that conversation, no matter how well you rank.

Earning those citations is its own discipline, usually called generative engine optimization or GEO. Some teams treat it as a buzzword wrapped around regular SEO. The overlap is real, but the target moved. You're no longer writing for a crawler that indexes whole pages against keywords. You're writing for a retrieval layer that cuts your content into passages, ranks those passages by semantic similarity, and hands the winners to a language model that decides which URLs deserve credit.

This piece covers SEO for AI search from the inside out: how the retrieval pipeline works, what the first citation datasets reveal, and the concrete changes we made to our own publishing stack at BlockFrame Labs.

How answer engines pick their sources

When someone asks ChatGPT Search or Perplexity a question today, the system rarely answers from memory alone. It runs a live web search, often several, pulls candidate documents from an index maintained by crawlers, and scores those documents against the question. Google runs the same play inside AI Overviews, which puts this retrieval layer in front of billions of queries a day.

From query to citation

Retrieval breaks each document into passage-level chunks and embeds them as vectors, so ranking depends on meaning rather than exact keyword overlap. Top-ranked passages travel to the language model as context, URLs attached. During generation, the model attaches citations to individual claims, and that attribution step is where most sites fall out. A passage that hedges, buries its point under two paragraphs of intro copy, or states a fact without a number gives the model little worth quoting, so the citation goes to whichever competitor wrote something extractable.

What the citation data shows

Citation share is lumpy. An analysis of domains cited by ChatGPT during 2025 put reddit.com first at 27.8% and wikipedia.org second at 14.23%, with tech review and news outlets filling most remaining slots. Two properties explain the leaders. Both answer questions directly, and both accumulate edits from huge numbers of people, which keeps them current in ways a quarterly corporate blog can't match. Spamming forums isn't the takeaway. The takeaway is that extractable, community-vetted answers beat polished funnels, and that getting cited at all requires being present in the index these systems search.

Technical deep dive: the GEO stack

Retrieval and chunking

A traditional search index maps a URL to keywords plus link authority. Answer engines map chunks of text to embeddings, dense vectors that encode meaning. Your 2,000-word article might enter that index as forty independent passages, each competing on its own. Chunkers usually respect heading boundaries, which is why a page organized as question-shaped H2s with immediate answers outperforms one long narrative. Descriptive anchor text and stable heading structure help the splitter cut at sensible seams. A useful mental model: treat every H2 section as a potential standalone answer card. If it were screenshotted alone, would it still teach something? Client-side rendered content often never enters the pool at all, because many retrieval pipelines skip JavaScript execution entirely.

Freshness matters more than in classic SEO too. Several answer engines re-crawl high-value sources aggressively and favor recently updated pages for time-sensitive questions, so a stale date stamp quietly demotes you.

Content signals that raise citation rates

A Princeton-led study on generative engine optimization tested nine content tweaks against AI search visibility and found that adding statistics, quotations, and source attributions improved citation rates by up to 40% in their benchmarks. Quotations performed best of anything they tried. The mechanism is intuitive once you watch a model compose an answer: it prefers restating sentences that already carry verifiable specifics, because specifics make the citation look credible to the person reading. Keyword stuffing, the old reliable hack, showed no measurable effect. Structured data helps a different way. FAQPage and Article schema hand the retrieval layer explicit type information, and llms.txt, an unratified proposal for a markdown file at your domain root that lists canonical pages, costs nothing to adopt while standards settle.

Measuring citations

Measurement is the immature part. No analytics platform reports citation share yet, so teams roll their own. The working approach: keep a list of twenty questions your buyers ask, run them through ChatGPT Search and Perplexity every week or two, and log which URLs get cited. Track your share over time alongside referral traffic from each engine. It's manual and unglamorous, but it beats vanity dashboards, and it surfaces something useful fast: the questions where competitors get quoted and you don't.

Implementation guide

You can retrofit an existing site without a rebuild. Access comes first, since every later step depends on it:

• Audit robots.txt for the AI crawlers: GPTBot, ClaudeBot, PerplexityBot, Google-Extended. Any disallow rule removes you from that engine's citation pool completely.

• Render critical content server-side or statically. Retrieval pipelines that skip JavaScript see an empty shell where your app should be.

• Restructure pages around questions. Each H2 poses the buyer's question, and the first sentence under it answers it plainly in under 50 words.

• Add FAQPage or HowTo schema where it genuinely applies: one claim per answer, no marketing adjectives inside structured data.

• Publish primary numbers. Benchmarks, surveys, pricing breakdowns. Models cite sources they can't find anywhere else, and original data has no substitute.

• Track referrals from chatgpt.com, perplexity.ai, and copilot.microsoft.com as their own channel. Volumes start small but intent runs higher than organic search.

• Re-check monthly. Crawler user agents change names, and a renamed bot silently bypasses your old rules.

Case study: BlockFrame Labs

We applied this playbook to blockframe.cloud before packaging it for clients. The site publishes through Notion as a headless CMS and deploys on Vercel, so every post already ships as static HTML with server-rendered metadata. Crawlers had full access from day one. The weaknesses were structural, not technical.

We rewrote older posts so each major section opens with its conclusion instead of warming up for two paragraphs. We attached FAQ schema to service pages where buyers ask comparison questions. Then we started publishing benchmark material from our own operations, including measurements from Nerve, our dashboard for deploying and observing AI systems, that no aggregator carried. Within weeks we saw referral entries from perplexity.ai in analytics, and those benchmark posts began surfacing in answers about deployment monitoring. We track this in a simple spreadsheet next to organic, because no mainstream analytics tool separates answer-engine traffic cleanly yet. The referrals convert better than search even at lower volume, since someone asking an answer engine how to monitor AI deployments arrives further down the buying journey.

The honest caveat: we can't separate GEO gains from ordinary SEO gains, because both draw on the same index. What we can say is that content written for extraction gets cited, and citations behave like qualified introductions.

Future outlook

Expect the ground to move. Publishers have noticed that an answer engine consumes their reporting and sends back a fraction of the traffic a classic result would, which is already producing licensing deals, citation-payment experiments, and tighter bot gating. Standards bodies are circling llms.txt and related proposals, so plan for churn there. Regulators may eventually push answer engines to show sourcing more consistently, which would only raise the value of being citable. None of that changes what operators should do right now: models need attributable sources to stay accurate, and they will keep retrieving whichever sources are easiest to verify. Optimize for the machine that quotes you, and you're positioned wherever the business layer lands.

Agentic browsing is the next wrinkle. As companies deploy AI systems that research and buy on a user's behalf, machines will read your pages while acting for someone else, and there's no click to win because there may be no human looking. Clean machine-readable specs, prices, and comparison data decide whether those systems understand what you sell well enough to recommend it.

Key takeaways

• Answer engines retrieve passages, not pages. Every section should survive being quoted alone, with the claim and its evidence in the same paragraph.

• Citation share concentrates hard. Reddit took 27.8% of domains cited by ChatGPT in 2025 tracking and Wikipedia 14.23%. Direct, community-vetted answers outrank polished brochures.

• Access comes before content. Disallowed AI crawlers, JavaScript-only rendering, and missing schema are silent failures no rank tracker will show you.

• Original data is the durable edge. Nobody else can cite your benchmarks, so publish numbers only you have.

• Measure chatbot referrals as a distinct channel and judge them on intent, not volume.

• Expect churn in standards and crawler policy. Re-audit access monthly so a renamed bot doesn't silently drop you from an index.

Sources: Top Domains Cited by ChatGPT (2025), Recover Reputation | GEO benchmark study | OpenAI, ChatGPT Search