Generative engine optimization and how to show up in AI answers
Every few years the industry invents a new acronym and declares everything that came before obsolete. This time it is GEO, generative engine optimization, and the pitch is that ranking in Google is yesterday's game because the real battle is being cited by AI Overviews, AI Mode, ChatGPT, Claude, and Perplexity.
Watch AI referral data, crawler logs, and citation patterns across enough sites and an honest read emerges: the destinations changed more than the discipline did. Most of what earns AI citations is recognizable SEO. But there is a genuinely new layer, and pretending otherwise is as lazy as the hype. Here is how these systems actually pick sources, what to do about it, and which parts of the GEO conversation are rebranded fundamentals.
How AI answer engines actually choose sources
The first thing to internalize is the difference between training data and retrieval. A model's training data shapes what it knows in general, and you have almost no near-term influence over it. Retrieval is what happens at question time: the system runs searches, fetches a set of candidate documents, and composes an answer grounded in them, ideally with citations. Retrieval is where visibility is won and lost, and retrieval looks a lot like search.
AI Overviews and AI Mode sit directly on top of Google's index and ranking systems. Google has been explicit that its AI features lean on core search infrastructure, and the systems fan a question out into multiple sub-queries, then draw on documents that rank for them. You cannot be cited by AI Overviews if you are invisible in the underlying index. The practical implication: ranking well, especially across the cluster of related queries a question decomposes into, remains the entry ticket.
ChatGPT search pulls from a search layer with its own crawler ecosystem: GPTBot gathers training data, OAI-SearchBot powers search indexing, and ChatGPT-User fetches pages live when a user's question triggers browsing. Perplexity is retrieval-first by design, running its own index fed by PerplexityBot and leaning heavily on sources it considers authoritative for the topic. Claude searches and fetches the live web when the question calls for it, with Anthropic publishing its own crawler user agents, including ClaudeBot. The systems differ in flavor, but the shape is the same everywhere: a search-like candidate selection step, then a language model deciding which passages to actually use and cite.
That second step is the genuinely new part. Traditional search ranks documents. These systems extract claims. A page can rank fifth and dominate the answer because one paragraph stated the key fact more clearly than anything above it.
Being quotable is the new on-page skill
When a model composes an answer, it wants passages it can lift cleanly: a direct claim, a definition, a number with context, a step-by-step that stands alone. Pages that bury their point across twelve paragraphs of throat-clearing give the model nothing to grab.
What to push for in content reviews:
Lead sections with the answer. State the claim in the first sentence, then elaborate. Inverted pyramid writing, which journalists have done forever, is ideal extraction material.
One idea per section, under a heading that says what the section claims, not a teaser. Clever headings are camouflage against retrieval.
Concrete, attributable statements. A model can quote a specific claim with a number and a date. It cannot quote vibes.
Structure that machines parse cheaply: real heading hierarchy, lists for genuinely list-shaped content, tables for comparisons, HTML that contains the content rather than assembling it client-side.
Consistent entities. Your company name, product names, and author names should be spelled and described identically across your site, your About page, and your external profiles. Retrieval systems reconcile entities across sources, and inconsistency reads as uncertainty.
Information the model cannot get elsewhere: original data, first-hand testing, real experience. Generic summaries of consensus are exactly what a language model already produces without you.
None of this harms human readers. That is the tell that it is good practice rather than a trick.
Technical access means letting the right bots in
This is the most concrete, most auditable piece of GEO, and it is broken constantly. If AI crawlers cannot fetch your pages, nothing else in this post matters.
Check your robots.txt for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, and Google-Extended. Blanket-blocking these was a popular defiant gesture a couple of years ago, and plenty of sites still carry those blocks like an old tattoo while their marketing team asks why they never appear in ChatGPT answers. Blocking Google-Extended keeps your content out of Gemini training but not out of AI Overviews, which follow ordinary Googlebot rules. Decide policy deliberately, per bot: training crawlers are a philosophical and legal question, but blocking the search and user-triggered fetchers is declining referral traffic on purpose.
Then check what actually happens beyond robots.txt. Audits keep turning up CDN bot protection silently serving challenge pages or 403s to AI crawlers that robots.txt welcomes. Grep your server logs or CDN logs for these user agents and look at the status codes they receive, and verify the traffic is genuine, since AI user agents are widely spoofed. Also remember most of these fetchers do not execute JavaScript. If your content only exists after client-side rendering, you are invisible to them regardless of policy. Server-side render anything you want quoted, and confirm with curl using the relevant user agent string.
The llms.txt debate, briefly and honestly
llms.txt is a proposed convention: a markdown file at your site root that curates your most important content for language models. It is cheap to implement and the idea is sympathetic. It is also, as of mid-2026, not confirmed as consumed by any of the major systems in a way that changes retrieval or citations. Google has said it does not use llms.txt, and server log analyses shared by practitioners show the major AI crawlers overwhelmingly fetching regular pages, not manifest files.
The pragmatic position: if it takes an hour, fine, ship one, it cannot hurt. But treat it as a lottery ticket, not a strategy. Any hour spent on llms.txt instead of fixing server-side rendering or unblocking a misconfigured bot rule is an hour spent on the wrong problem. Do not let a vendor sell you an llms.txt package as GEO.
Measuring AI visibility without fooling yourself
Measurement here is genuinely immature, so build a modest, honest stack. Start with referral traffic: chatgpt.com, perplexity.ai, claude.ai, and gemini referrers are visible in your analytics, and tagging them into a dedicated channel group takes minutes. The volumes are usually small relative to organic search but growing steadily, and their conversion quality is often striking, because someone who clicks through from an AI answer arrives pre-briefed.
Beyond referrals: monitor AI crawler activity in your logs as a leading indicator of inclusion. Run a recurring panel of the questions you care about through the major assistants and record whether you are cited, by page and by claim, accepting that answers are non-deterministic and personalized, so trends matter and single runs do not. Several tracking tools now automate this sampling; treat their absolute numbers skeptically and their trend lines as directionally useful. And watch Search Console for impressions that hold while clicks sag on informational queries, the familiar signature of AI Overviews absorbing the click.
What is genuinely new versus rebranded
Worth calling out plainly, because the discourse profits from blurring it.
Genuinely new: optimizing for passage extraction and quotability. Managing per-bot crawler policy as a business decision. Citation tracking across assistants as a KPI. Answer non-determinism, meaning you monitor distributions rather than positions. Referral analysis for AI platforms.
Rebranded: crawlability and server-side rendering, which is technical SEO. Entity consistency and structured data, which the industry has preached for a decade. Authority, original research, and being cited by other trusted sites, which is E-E-A-T and digital PR. Clear writing under descriptive headings, which was always just good content.
The overlap is not a coincidence. These systems retrieve from indexes built by crawlers ranking documents by relevance and authority. Of course the same fundamentals dominate. In most audits, the site that ranks well and answers clearly is already the site the AI engines cite, and the correlation between strong organic visibility and AI citations is obvious the moment you start tracking both.
What to actually do this quarter
Audit robots.txt, CDN rules, and server logs for every major AI crawler, and fix unintended blocks.
Confirm your money content is present in the initial HTML, not assembled by JavaScript.
Rewrite your ten most important pages so each section leads with its claim under a heading that states it.
Tighten entity consistency: Organization and Person schema, a real About page, identical naming everywhere.
Set up AI referral tracking and a monthly citation check for your priority questions.
Keep doing the boring work: earning links, publishing original data, improving the pages people actually cite.
GEO is not a replacement for SEO. It is SEO with a new front end and a stricter editor. The systems answering questions in 2026 reward the same things search always rewarded, plus clarity of claim and cleanliness of access. Be crawlable, be quotable, be consistent, and be genuinely worth citing. The acronym on the invoice matters much less than that.