Optimizing for AI Search (GEO / AEO)

Make a website something AI search engines and assistants can reach, read, and cite. The goal of classic SEO was rank #1, earn the click; the goal here is be one of the few sources the model quotes in a synthesized answer (often zero-click).

The one honest principle

~90% of GEO is rigorous technical SEO + making content readable without JavaScript + factual density. There is very little that is genuinely "AI-specific." Google's own guidance says there are no special requirements or optimizations for AI Overviews/AI Mode beyond solid, crawlable, textual content. Resist vendor hype — especially "magic schema percentages" and llms.txt miracles. Lead with the deterministic, evidence-backed wins below.

How AI engines pick what to cite (why this works)

Every citing engine runs a RAG pipeline: rewrite the query → retrieve passages/chunks (from a crawler index and/or vector store) → re-rank → synthesize an answer → attach citations to the passages actually used. Two consequences drive everything:

1. If your content isn't in plain server-rendered HTML, it isn't retrieved. Most AI crawlers do not execute JavaScript.
2. You compete at the passage level, not the page level. Each section must be self-contained, quotable, and fact-dense.

The workflow

Run the audit, then fix in priority order (rendering first — it dwarfs the rest).

Step 1 — Audit (deterministic checks)

Run the script on a representative URL. It needs curl (optional jq/python3):

./scripts/geo-audit.sh https://example.com/a-real-content-page

It reports robots.txt directives for the crawlers that matter, whether content exists in raw HTML (server-rendered), whether an AI UA gets a 200, JSON-LD presence/validity, and llms.txt presence. Read what it CAN'T check: content quality and off-site authority are judgment calls (Steps 4–6).

Step 2 — Fix rendering (the #1 lever)

Most AI crawlers fetch HTML but never run JS. A client-rendered SPA shows them an empty <div id="root">.

• Server-render or statically generate every content page (Next.js, Nuxt, Astro, SvelteKit, Remix, etc.) so content + <title>/meta + canonical + JSON-LD are all in the initial HTML response.
• Stuck on a SPA? Pre-render to static HTML (e.g. prerender.io, static export, islands).
• Verify, don't trust: curl -A "GPTBot" <url> (or the audit script) — if a fact isn't in that output, AI crawlers don't see it. Browser DevTools shows the hydrated DOM and will lie to you here; use View Source / curl.

Step 3 — Fix crawler access (and avoid the classic traps)

Vendors run three tiers of bot. Blocking the wrong one silently kills AI visibility. Full table + a ready-to-paste robots.txt: references/crawlers.md.

• Never block search/retrieval crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot, Bingbot) — these are how you get cited.
• Google-Extended does NOT control AI Overviews. Those are served from the normal Googlebot index. Google-Extended only opts out of Gemini training.
• Only block training crawlers (GPTBot, ClaudeBot, CCBot, Google-Extended, Applebot-Extended) if protecting content from training is a deliberate goal — it carries a real visibility cost.

Step 4 — Fix content for citation (evidence-backed)

The Princeton GEO study (Aggarwal et al., KDD 2024; ~10k queries) measured what raises a source's visibility in generative answers by up to ~40%. The reliably winning levers — all forms of verifiable factual density — and the things that don't work: references/content-tactics.md. The essentials:

• Answer-first. Lead each section with a direct, self-contained answer in the first 1–2 sentences, then elaborate. Make every H2/H3 a "mini-article" that survives being lifted out of context.
• Add statistics (concrete numbers, not vague claims) and cite credible sources inline and add expert quotations — the three highest-impact tactics (~+30–40%; the source-citing gain is largest for lower-ranked pages).
• Question-style headings + a real FAQ matching how people prompt assistants.
• E-E-A-T: real author bylines/bios/credentials, primary-source citations, explicit dates (AI citations skew heavily toward fresh content).
• Machine-parseable formatting: lists, comparison tables, clear definitions.
• Do NOT keyword-stuff, pad, or artificially simplify — measured at ~0% or negative. Classic keyword tactics do not transfer.

Step 5 — Structured data (useful, but don't over-claim)

Add JSON-LD (server-side) for entity clarity and classic rich results; AI-citation benefit is plausible but weakly evidenced. Which types still matter, which Google rich results are dead (FAQPage retired 2026; HowTo removed 2023–25), and the iron rule that markup must match visible content: references/structured-data.md.

Step 6 — Off-site authority (not editable on the page)

LLMs synthesize across many sources, so a large share of citation probability lives off your site: mentions on high-authority third-party sites, Wikipedia, Reddit, review platforms; breadth of referring domains; consistent brand narrative (digital PR). Name this explicitly to the user — on-page fixes alone have a ceiling.

Step 7 — Measure (honestly)

Three independent layers — prompt-based visibility monitoring, AI-referral traffic in GA4 (heavily undercounted as "direct"), and server-log crawler analysis — plus the tool landscape and what is genuinely measurable vs. speculative: references/measurement.md.

Quick reference

Red flags — stop and reconsider

• "We'll just add llms.txt and we're done." → No engine has confirmed using it; it's not a substitute for server-rendered HTML. Low priority, unproven.
• "Schema markup will get us cited." → Over-claim. Markup helps parsing/rich results; it is not a proven citation lever, and mismatched markup risks a manual action.
• "Block GPTBot to protect our content." → That's a training bot; it doesn't affect citations. Blocking OAI-SearchBot/PerplexityBot is what removes you from answers.
• "Block Google-Extended to opt out of AI Overviews." → It can't; AI Overviews use the main Googlebot index.
• "The content is right there" (in a SPA). → Check raw HTML via curl, not DevTools. Hydrated DOM ≠ what crawlers see.
• "Let's add more keywords." → Keyword stuffing is inert-to-negative for GEO.

Common mistakes

• Auditing the homepage only — audit a real content page (article/product).
• Trusting DevTools' rendered DOM instead of curl/View Source for the SSR check.
• Treating GEO as a replacement for SEO — it's additive; crawlability and E-E-A-T are shared prerequisites.
• Promising citation numbers. Engines are non-deterministic and personalized; report trends, never single snapshots, and state confidence honestly.
• Misreading the content-volume signal on CJK (Chinese/Japanese/Korean) or other non-space-delimited sites — naive word counts undercount badly. The audit script counts CJK characters as content units; if you eyeball raw HTML, judge by characters, not spaces.