Generative Engine Optimization: The Operator's Guide

Generative engine optimization — GEO — is what search optimization becomes when the result is a written answer instead of a list of links. This is the operator's version of the guide: we run 3 production authority builds, and every tactic below is tagged by the evidence behind it — our own logs, a named study, platform documentation, or honestly labeled speculation. Nothing here is a guarantee of rankings, citations, or traffic, because no one controls answer-engine output.

What is generative engine optimization?

GEO is the practice of structuring content so AI answer engines can retrieve it, quote it, and cite it as a source. The term comes from a Princeton-led research paper (arXiv 2311.09735, posted November 2023, presented at KDD 2024) that also built the first benchmark for measuring visibility inside generated answers.

The one-sentence model: engines answer questions by retrieving passages, so the page that wins is the page whose passages survive being lifted out of context. That reframes the familiar SEO contest — rank the page — into a passage-level contest: be the quotable chunk. The full definitional treatment, including what GEO is not, lives in what is GEO, and the point-by-point comparison with classic search is in GEO vs SEO.

How do AI answer engines decide what to cite?

Every major answer engine runs some version of the same pipeline: retrieve candidate documents, chunk them into passages, select passages relevant to the question, and synthesize an answer that cites some subset of them. You can act on three of those four stages — retrieval (be crawlable and indexed), chunking (write self-contained sections), and selection (make passages direct, evidenced, and quotable). Synthesis is the model's alone.

Two data points bound what is winnable. Kevin Indig's analysis found 24% of ChatGPT answers were generated without fetching any web content at all (Growth Memo, 2026) — no page wins those. And crawling is wildly disproportionate to referral: Cloudflare measured crawl-to-refer ratios of roughly 38,066:1 for Anthropic, 1,091:1 for OpenAI, and 195:1 for Perplexity, against about 5.4:1 for Google (Cloudflare, July 2025). The mechanics of the pipeline, stage by stage, are in how AI search works.

What has actually changed for publishers?

The measurable change is the click, and the numbers deserve to be read together. Ahrefs found CTR for top-ranking pages fell 34.5% on queries with AI Overviews (March 2025), with follow-ups reporting −58% by December 2025. Pew Research found users clicked a traditional link in 8% of visits with an AI summary versus 15% without, and clicked a source cited inside the summary in roughly 1% of visits (March 2025 data, published July 2025).

The counter-current: Seer Interactive measured organic CTR roughly 35% higher for sites cited in the AI Overview than for sites absent from it (2025). These findings reconcile cleanly. The total click pool shrinks; the remaining clicks concentrate on cited sources. Fewer, better-qualified visitors is the honest description of the new economics — and the reason the citation is worth competing for even when almost nobody clicks it.

Which GEO tactics actually work?

Every tactic we use or get asked about, tagged by its evidence tier — proven in our fleet, supported by a named study, platform-documented, or unproven:

The table is the guide. Work down from the top: the tactics with the strongest evidence are also, conveniently, the cheapest — writing a direct answer costs nothing but discipline.

What doesn't work — or can't be proven yet?

The discipline's junk drawer is large, and naming its contents is part of the job. Schema-as-citation-engine leads the list: Ahrefs tested 1,885 pages and found no causal lift in AI citations (2025). Schema still earns its keep for other purposes — the nuance is in schema for AI search: the evidence.

The llms.txt proposal (llmstxt.org) — not an adopted standard, and no major engine has committed to reading it — is the clearest case of deployment outrunning evidence. We serve it on all 3 builds and log every request: 0 from any major engine crawler to date [our data]. The running log is in our llms.txt watch.

Beyond specific tactics, distrust unsourced precision. This niche launders invented numbers — "content quality scores" with no methodology, conversion multipliers with no cohort definition — between vendor blogs until they sound established. Our rule: if a number has no primary source and is not in our own logs, it does not appear on this site.

How do we run GEO in production?

Our operating loop, across an insurance lead-gen brand, an auto-finance authority rebuild, and our own network [our data]: demand research to find the questions people actually ask; site architecture that gives every question one page; production with a fact-audit gate — 400+ published pages have passed it — so no unsourced statistic ships; then measurement, which is where most GEO programs go vague.

We measure three ways. A GA4 channel we built segments AI-assistant referrals from ordinary traffic; a Search Console mining pipeline approximates AI Overview exposure that Google does not report directly; and server-log analysis tracks which AI crawlers fetch what, which is how the llms.txt result above exists at all [our data]. None of this requires paid AI-visibility tools — a point we make against our own interest, since audits and builds are what we sell.

The output of the loop is receipts: a documented AI Overview citation win recorded from dated manual checks, negative results published with the same prominence as wins, and every page on this site carrying its sources in the open.

Where should you start?

Start with the shortest path through the evidence table, in order:

1. Verify access. Fetch your key pages as an AI crawler would and confirm the answers are in the raw HTML.
2. Restructure your top 10 question pages answer-first: direct answer up top, question-shaped headings, sections that stand alone.
3. Source your claims. Add real citations and concrete numbers — the Princeton-benchmarked tactic class (up to +40%, KDD 2024).
4. Instrument before you spend. Segment AI referrals in your analytics and watch your logs before buying any tool or retainer.

Then read this library's deeper layers as you hit them — definitions, platform mechanics, measurement, and costs are all one internal link away from this page. Everything here obeys the same contract: evidence tiers declared, sources in the open, and no promises about a system nobody controls.