Generative engine optimisation, without the hype

Where the term came from, and what it means

The phrase comes from a November 2023 research paper on arXiv titled "GEO: Generative Engine Optimization", written by a group of academic researchers and later presented at the KDD conference in 2024. It gave a name to an obvious problem: if an AI system answers the question and cites a handful of sources, how do you become one of the sources?

The paper tested a set of content changes against a benchmark of queries and reported that adding citations, quotations from credible sources and relevant statistics improved a page's visibility inside generated answers, while keyword stuffing did not. Worth reading, worth treating as directional — it was measured on a constructed benchmark, not on live commercial traffic, and the systems have changed several times since.

Then the term got picked up by the marketing industry, which is where the trouble started. You'll now see GEO, AEO (answer engine optimisation), AI SEO and "search everywhere optimisation" used interchangeably, each with a price attached. The concepts underneath overlap almost completely. If you want the distinction spelled out, we've compared GEO against ordinary SEO directly.

The part that is just SEO with a new name

Be honest about the overlap before paying for the difference. Almost everything on the standard GEO checklist is something a competent SEO was already doing in 2019.

The part that's genuinely different

There is a real difference, and it's narrower than the pitch decks suggest. It comes down to what happens after retrieval, when a model has your page in context and has to decide whether to use it.

A ranking system asks: is this page relevant and trustworthy enough to show? A generative system asks something harder: can I lift a defensible sentence out of this page and attribute it? Those two questions reward different writing.

• Extractability. A complete answer inside two or three sentences, near the top, that survives being pulled out of the page. Prose that only makes sense after 800 words of context is invisible to an extraction step.
• Self-contained blocks. Sections that stand alone. "As mentioned above" is fine for a reader and useless to a system that only retrieved that chunk.
• Attributable claims. Specific, checkable statements — a number, a date, a named source, a price. Generic advice gets absorbed into the summary without credit; a specific claim has to be attributed to somebody.
• Information that exists nowhere else. A synthesiser blends five sources into one paragraph and cites none of them individually. It cannot do that with a figure only you have published. This is the strongest single lever and it's a content decision, not a technical one.
• Bot access as a deliberate choice. Crawlers like GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google's Google-Extended can each be allowed or disallowed by name. Blocking them all is a legitimate position with a cost attached: you also stop appearing in the answers your competitors appear in.

How a generative engine actually picks a source

Understanding the pipeline stops you buying magic. There are three gates, and different work matters at each one.

1. Retrieval. The system runs one or more searches and pulls a candidate set of pages. If you aren't in an index it can reach, the process ends here and nothing else you do matters. This gate is pure classic SEO.
2. Selection. From the candidates, it picks which pages to ground the answer in. Corroboration matters here — a claim confirmed in several places is safer to repeat than one that appears only on your own site. So do the usual credibility signals, because they shaped which pages ranked in the first place.
3. Attribution. The model writes an answer and decides what to link. This is where extractability pays. Pages that state something specific and quotable get named; pages that restate the consensus get blended into it without a citation.

Measuring citation share, honestly

This is the weakest part of the whole discipline, and it's where most GEO products oversell. There's no stable equivalent of rank tracking, for a structural reason: generated answers vary by phrasing, by session, by user context, by location and by model version. Ask the same question twice and you can get two different source lists.

What you can do is sample deliberately and accept the noise. Pick 20 to 40 questions your buyers genuinely ask. Run them across the assistants that matter to you, on a fixed schedule, from a clean session. Record whether your brand is mentioned, whether you're cited with a link, and which competitor got named instead. Repeat monthly. You're tracking a trend, not a position.

Two supporting signals are more reliable than any citation tracker. First, referral traffic from assistant domains in your analytics — small in volume, usually high in intent. Second, branded search volume, because the most common outcome of an AI mention isn't a click, it's someone Googling your name afterwards.

• Treat smooth dashboards with suspicion. A precise "AI visibility score" trending neatly upward is a modelled number. Ask how it's sampled and how much it moves on a re-run.
• Log the prompt, not just the result. Slight rewordings change the answer. Without the exact prompt, the measurement isn't repeatable.
• Watch who else gets cited. The competitor set in AI answers is often different from the one in the SERP, and that's genuinely useful information.

Does it deserve its own budget line? Not yet.

Here's the position we'll defend. For almost every Indian business we talk to, GEO is not a separate service to buy. It's a change to how the pages you're already paying for get written and structured.

The reasoning is arithmetic. If four-fifths of the work is identical to your existing SEO, buying it twice means paying twice for the same crawl fixes, the same schema, the same digital PR. And the remaining fifth — extractability, self-contained answers, publishing information nobody else has — costs almost nothing to adopt because it's a brief and an editing standard, not a workstream.

There's one exception worth naming. If your buyers demonstrably research through assistants, and you can see it in referral data or hear it on sales calls, then dedicated measurement and a deliberate content programme start to earn a line item. For most people that day hasn't arrived, and paying for it early funds a category rather than a result.

The practical move today is smaller than the pitch: answer the question in the first two sentences of every page, publish at least some information that doesn't exist anywhere else, keep your brand facts consistent, and decide on purpose which crawlers you allow. That's most of GEO, and it makes your ordinary rankings better too. The tactical detail sits in how to get cited by AI assistants.