What white-hat GEO actually is — getting cited by AI answers, and what doesn't work

GEO optimizes for being cited, not ranked

Generative engine optimization (GEO) is the practice of getting your content quoted inside AI-generated answers — the responses from ChatGPT, Perplexity, and Google’s AI Overviews. The goal is different from search. A blue link wins by ranking above its neighbors. A GEO win is when the model pulls a sentence, a number, or a claim from your page and attributes it in the answer it writes.

That shift changes what you optimize for. In classic search you fight for position on a results page a human scans. In GEO there is often no list to climb — there is one synthesized answer, and either you are inside it or you are not. The unit of success moves from ranking to citation, and the measurement moves with it: not impressions and average position, but whether a model names you as a source when a user asks a question your page can answer.

This is not a fringe concern anymore. Gartner projects that by 2028, 90% of B2B buying journeys will be influenced by AI agents, tied to more than fifteen trillion dollars in purchases. When an agent assembles a shortlist on a buyer’s behalf, being cited in its answer carries the weight that ranking on page one used to — and the buyer may never see the page it cited, only the synthesized claim.

How it differs from traditional SEO

SEO and GEO overlap on fundamentals — crawlable pages, clear topical authority, real content — but they reward different things. SEO rewards relevance and link signals that lift a page in a ranked list. GEO rewards content a model can lift cleanly and trust enough to attribute: standalone factual statements, clearly sourced claims, and structure that survives being cut into a passage and placed into an answer the model writes itself.

The other practical split is the click. SEO assumes a human clicks through, so the title tag and meta description are optimized to win that click. AI answers frequently resolve the question in place, so your payoff is being named as the source rather than getting the visit. You optimize for the model’s confidence in quoting you, not for a tempting title. That reframes a lot of on-page work: the question is no longer “will a human click this headline” but “can a model extract a clean, attributable claim from this paragraph without the surrounding page.”

The two disciplines are not in conflict. The same crawlable, fast, well-structured page tends to do well in both, and the underlying signals — authority, accuracy, structure — overlap heavily. GEO is better understood as an additional optimization target layered on a technically sound site than as a replacement for SEO.

What the research actually shows works

The clearest evidence comes from the GEO paper presented at KDD 2024, which tested concrete content changes against AI answer engines rather than guessing at what helps. The headline result: adding statistics, direct quotations, and cited sources raised content visibility in generated answers by up to roughly 40% in their experiments. Among the methods tested, “Cite Sources,” along with adding quotations and statistics, were the strongest single moves — the pattern being that the model rewards material it can verify and attribute, not material that is merely keyword-dense.

Two details from that paper matter more than the headline number, because they tell you where the leverage is. First, the gains were not uniform across response styles or query types — the same change that helped a page in one position helped a different page differently. Second, and most useful operationally: the relative visibility boost was largest for pages that started lower in the ranking — a page sitting around the fifth position gained the most from these changes, while a page already sitting first could actually see its visibility drop. The practical reading is that GEO methods are most valuable to content that is good but not already dominant. If you are the incumbent top result, aggressive restructuring can cost you; if you are the credible challenger, sourced and quotable content is how you get pulled into the answer above where your ranking alone would put you.

So the work is concrete. State specific numbers with their source. Quote named experts or primary documents. Make each important claim self-contained, so a passage extracted from the middle of your page still makes sense and still carries its citation. Structure with clear headings and short, declarative sentences a model can lift without ambiguity. This is the inverse of writing that buries the claim three clauses deep behind qualifiers — that prose reads fine to a human and extracts badly for a machine.

Extractability is a precondition, not a tactic

None of the content work matters if the crawler can’t read the page. AI retrieval crawlers generally do not execute JavaScript. An analysis of over half a billion GPTBot fetches found AI crawlers do not render JavaScript — they fetch the raw HTML response and stop. If your content only appears after the browser runs a client-side framework, the crawler sees an empty shell, and there is nothing to cite. Server-rendered or static HTML is not a nice-to-have here; it is the difference between being readable and being invisible to the systems you are trying to get cited by.

The fix is architectural. Put the meaningful text in the initial HTML response. Frameworks like Next.js App Router render on the server by default, and Astro ships static HTML with JavaScript only where you opt into it — both put real content in the first response. Deploying that rendered output close to the request (for example via OpenNext on Cloudflare, or static hosting on Cloudflare Pages) keeps it fast as well, which matters because Core Web Vitals and general technical health still feed the search and AI systems your content has to pass through.

Then verify what the bot actually receives, not what your browser paints. The cheapest reliable check is to fetch the URL the way a non-rendering crawler would — curl the page and read the raw HTML, or drive a headless browser with JavaScript disabled — and confirm the claims you want cited are present in that response. A tool like Playwright can automate the with-JS-versus-without-JS comparison across a whole site, which is the fastest way to catch a page that looks complete in a browser and arrives empty to GPTBot.

Structured data helps the machine understand what it is reading. Google’s structured data documentation and the schema.org vocabulary let you mark up articles, authors, organizations, and FAQs so the type and provenance of your content are explicit rather than inferred. Schema is not a magic citation lever, but it removes ambiguity about what a passage is and who stands behind it — which is exactly the kind of signal a model leans on when deciding whether to attribute.

What doesn’t work

Two popular tactics don’t hold up under measurement.

The first is treating llms.txt as a citation lever. The proposed file is meant to tell AI systems how to read your site, but a study across roughly 300,000 domains found no clear effect on AI citations, with adoption only around 10% and no major AI engine confirming it uses the file. Shipping one costs little, so it’s not harmful, but treating it as something that moves citations is a misread of the evidence.

The second is keyword stuffing carried over from old SEO. Models synthesize meaning from context; they do not reward density. Repeating a phrase to hit a count doesn’t make you more quotable — it makes the passage worse to lift and signals low quality. The KDD result points the opposite direction: what raises visibility is verifiable substance (statistics, quotations, sources), not term frequency.

The white-hat line

There is a black-hat version of GEO: hidden instructions aimed at the model, content cloaked so the AI crawler sees something different from humans, and fabricated authority. The line is not subtle, and the engines have stated where it is. Google updated its spam policies to clarify that manipulating generative AI responses falls under the same spam rules as manipulating search. That puts the manipulative playbook squarely on the wrong side of the line, with the same downside as black-hat SEO once detection catches up — and detection of cloaking and content mismatch is a problem search engines have been solving for two decades.

Cloaking deserves a specific warning, because it overlaps with a real security boundary. Serving the crawler different content than humans see is the same shape as an access-control failure — the “user” who sees the privileged view is determined by something spoofable (the user agent), which is the pattern OWASP catalogs as broken access control. It is brittle, it is detectable, and when it’s discovered the penalty applies to the whole domain, not just the cloaked page.

White-hat GEO earns citations the durable way: real, verifiable content; the same page for humans and crawlers; sources you can stand behind. It tracks what is genuinely measurable — citation appearances, referral patterns from AI surfaces, the share of your sourced claims that show up in answers — and promises no guaranteed mention, because citations aren’t deterministic and any vendor who guarantees one is selling something the engines don’t offer.

Why this matters now

The buying behavior is already moving, and the Gartner projection above is the leading indicator: when most B2B journeys are mediated by agents that read raw HTML and quote sourced claims, the teams whose content is extractable, sourced, and honestly authoritative are the ones those answers will cite. The work is unglamorous — render on the server, source your claims, keep one page for humans and machines, and verify what the crawler actually receives. But it compounds, and unlike the manipulative shortcuts it doesn’t get clawed back the moment the engines update their policies.