Generative Engine Optimisation (GEO) Guide 2026

What Generative Engine Optimisation Actually Optimises For

What is Generative Engine Optimisation (GEO)?

GEO is the practice of structuring content so a large language model selects it as a source when generating a conversational answer, rather than structuring it to rank in a list of links.

GEO is the generation-layer sibling of Answer Engine Optimisation (AEO). AEO gets you into featured snippets and Google AI Overviews. GEO gets you cited when ChatGPT, Claude, Gemini, or Perplexity write a paragraph answering someone's question. The two overlap heavily in technique, but they're solving different problems: AEO extracts an answer from an index it already has, GEO retrieves candidates and generates a new sentence around the ones it trusts. If you want the full AEO vs SEO comparison, that lives in the AEO guide; this one goes deep on the mechanism that's specific to GEO.

Why GEO Should Come First

Should I focus on GEO or AEO first?

For LLM-native engines like ChatGPT, Claude, and Perplexity, GEO first. A model that can't resolve who you are won't ground an answer on your content, however well the answer is written. Google's AI Overviews are the exception: Google's own index already carries a baseline of who you are, so AEO can produce results there without GEO being mature first.

In the audits and workshops we run, this is the pattern that shows up consistently: businesses that fix entity resolution, the schema, @id, and sameAs work covered later in this guide, before touching answer formatting see faster, more durable traction than businesses that start with answer capsules and structured FAQs alone. GEO is the identity layer. AEO is the answer-quality layer. You can format the best answer in the world, but a model that can't confirm who you are has nothing to ground it on.

That sequencing isn't universal, though, and it's exactly where advice that treats AEO and GEO as interchangeable gets it wrong. Google's AI Overviews sit inside an index Google has been building for over two decades; your entity is already resolved to some degree through Search and the Knowledge Graph, independent of any GEO work you've done. That's why AEO tactics, answer capsules, question-based headings, direct extractable answers, can win an AI Overview citation on their own, even before your GEO foundation is solid. A newer LLM without that decades-long crawl relationship doesn't have that shortcut. It has to resolve who you are through retrieval and grounding signals in the moment, which is what GEO exists to make possible.

None of this makes AEO secondary in importance, only in sequence for LLM engines. Once a model can confirm who you are, what decides whether it actually cites you is the same thing that decides whether a human reader trusts you: does the answer genuinely answer the question, directly and specifically, without padding it out to look more substantial than it is. That's the real substance of AEO, and it's covered in full in the AEO guide.

GEO Checks You Can Run This Week

What's the quickest way to improve GEO?

Confirm AI crawlers can actually fetch your pages, then make sure every page describes the same entity with the same @id, working sameAs links, and factual descriptions.

These are the checks that most often turn up gaps when we audit a site for AI visibility. None of them needs a specialist tool, and most take minutes. Work through them in order: there's no point polishing content a crawler can't reach, and no point publishing original data a model can't attribute to you.

Check	Why it matters for GEO	How to test it
1. Can a bot actually fetch your pages?	Firewalls and bot protection can block AI crawlers even when robots.txt allows them. Cloudflare, for example, has prompted new domains to block AI crawlers by default since July 2025.	Run `curl -I -A "PerplexityBot" https://yoursite.com`. A 403 or a challenge page is a warning sign. Confirm in your host's bot or firewall settings, since a borrowed user agent isn't treated exactly like the real crawler.
2. Training and search crawlers treated separately	Blocking GPTBot (training) is a legitimate choice. Blocking OAI-SearchBot keeps you out of ChatGPT search answers.	Open yoursite.com/robots.txt and check each crawler name individually, not just the wildcard rule.
3. No snippet blocks on pages you want cited	A `nosnippet` or `noindex` directive stops Google using the page in AI Overviews and AI Mode.	View source on your key pages and search for `name="robots"`.
4. One @id across the whole site	If your homepage and your articles describe your business with different @id values, a model sees two entities instead of one.	Paste your homepage and one article into validator.schema.org and compare the Organisation @id.
5. sameAs links that resolve to you	A dead link, or one pointing at a similarly named company, misleads entity resolution.	Click every URL in your sameAs array. Each should load and clearly be your business.
6. NZBN in your schema	A public registry identifier is a fact a model can verify, rather than a claim it has to take on trust.	Search your Organisation JSON-LD for your 13-digit NZBN, in a `taxID` or `identifier` field.
7. Factual descriptions, not slogans	Models lift descriptions almost verbatim. "AI training and advisory for NZ businesses, Auckland" is citable; "Empowering your future" is not.	Read your meta description and schema description aloud. Keep meta descriptions to roughly 150 to 160 characters so they aren't cut mid-claim.
8. Answer first, in short passages	Retrieval works on passages, not whole pages. A section that opens with its answer survives being lifted out of context.	Read only the first two sentences under each heading. Do they answer the heading on their own?
9. One original, attributed number per key page	Original data is the clearest form of Information Gain, covered in detail below.	For each service page, find a figure only you could have published. If there isn't one, that's the gap.
10. Same name, same facts, everywhere	Inconsistent names, locations, or founders across your site, LinkedIn, and directories weaken the match between sources.	Compare your site, LinkedIn, Google Business Profile, and NZBN record side by side.

We run these checks on our own site too. In September 2026 they caught seven meta descriptions long enough to be cut off mid-sentence, a hard-coded star rating in our schema that wasn't tied to any live review source, and a claim on one tool page that didn't point to the sources further down the same page. A human visitor would never have noticed any of them. A model reading the markup would.

How Retrieval-Augmented Generation Decides Who Gets Cited

How does an LLM decide what to cite?

Retrieval-Augmented Generation (RAG) retrieves a set of candidate sources for a query, ranks them for relevance and trustworthiness, grounds the answer on the highest-ranked few, then generates a response with inline citations back to those sources.

That grounding step is the entire game. An LLM doesn't "know" your business the way it knows facts baked into its training data; for anything current or specific, it retrieves live candidates, the same way a search engine would, then writes around the ones it decides to trust. Getting into that shortlist of 2-7 trusted sources is what every tactic in this guide serves.

Why the Same Page Performs Differently Across ChatGPT, Claude, Gemini, and Perplexity

The four major engines don't retrieve the same way, so a page optimised for one doesn't automatically perform in another. Perplexity is built as a citation-first answer engine: nearly every response carries inline sources you can click through to. ChatGPT's default answers draw on training data unless a live search is triggered, in which case OAI-SearchBot's index takes over. Claude's web search grounds answers against live retrieval only when that feature is switched on. Gemini leans heavily on Google's own search index, which is one reason the technical SEO fundamentals that feed Google Search also feed Gemini's grounding.

The practical implication: don't assume one citation win transfers to every platform. Test your core buyer questions across all four periodically rather than optimising for whichever one you happen to check.

Share of Model: A Benchmark, Not the Whole Picture

Traditional SEO tracks share of voice: what percentage of search impressions you captured. GEO's equivalent is Share of Model: of all the times someone asks an AI assistant a question in your category, what percentage of answers cite you. If 100 people ask ChatGPT "who provides AI training for businesses in New Zealand" and you're cited 40 times, that's a 40% Share of Model for that query. It's the visibility metric GEO work is ultimately trying to move.

Be careful how it's measured. Most visibility tools send thousands of prompts through an AI provider's API, which makes the test repeatable but synthetic. A real customer's session carries context the API call doesn't: their location, earlier questions in the conversation, personalisation, and live web search. A buyer who has already told ChatGPT they run a small firm in Christchurch can get a materially different answer from the one a standalone prompt produces. Treat an API-based score as a benchmark for tracking change over time, not as what New Zealand customers actually see.

It also only measures the end of the chain. A low score could be an access problem, an entity problem, or a content problem, and the score alone won't tell you which. That's why the measures further down this guide include crawler activity and referral traffic as well as citations.

Maximise Information Gain

Information Gain is a concept from a 2022 Google patent that has become a core operational metric for AI citation systems. It refers to unique, valuable information your content provides beyond what competitors already cover. Proprietary data, original research, and direct expert quotes score high. Generic summaries of widely available information score near zero.

LLMs are trained on the consensus of the web. Content that simply echoes what everyone else says adds no new signal; the model already "knows" it. This is why AI-generated content that just paraphrases existing sources has near-zero information gain by definition, and why it's systematically deprioritised by retrieval and grounding systems. A Princeton study on generative engine optimisation found that adding original statistics improved citation rates by 41%, and adding expert quotations improved them by 40%, the two largest single-tactic gains measured.

Practical steps: add at least one original statistic or direct attribution per key page, publish proprietary data from your own work (client results, internal surveys, real project outcomes), and use direct quotes with named attribution rather than paraphrasing.

Entity Clarity: How a Model Confirms You're Real

What is entity clarity, and why does it matter for GEO?

Entity clarity is whether an AI system can confirm that the business or person described on your page is the same real-world entity referenced elsewhere online. It's resolved through structured data, not prose, and it's a prerequisite for being trusted as a citation source.

An LLM doesn't read your About page and take your word for it. Before it grounds an answer on your content, it needs to resolve who you actually are: is the business named on this page the same one that LinkedIn, an industry registry, and (if you have one) Wikidata all know about? That resolution happens through structured identity signals, not marketing copy.

The mechanism is your Organisation schema's @id and sameAs array. @id is a stable, self-referencing URI (typically https://yoursite.com/#organization) that anchors every piece of schema across your site to a single entity, rather than leaving each page's markup floating unconnected. sameAs is the list of external URLs, your LinkedIn company page, industry registry entry, Crunchbase profile, Wikidata item, that all point back to the same identity. Organisation and Person schema together produced a 19.72% increase in AI Overview visibility in one published entity-linking case study, on top of the general schema lift covered above. For New Zealand businesses, the sameAs list should also carry your NZBN as a taxID or identifier field, since it's one of the few trust anchors an AI system can verify against a public registry rather than taking your word for it.

Try this: view source on your homepage and on one article, then search each for @id inside the Organisation block. If the two values differ, a model sees two businesses, not one.

Wikidata: The Verification Layer Most Businesses Skip

A Wikidata item, where your business qualifies for one, is one of the strongest sameAs entries available, because it's an independently maintained, third-party-sourced record of who you are rather than something you wrote about yourself. Wikidata's bar for inclusion is lower than Wikipedia's: you generally qualify if you have a Wikipedia article in any language, a mention in a verifiable third-party published source, an official registry identifier, or a reference from another existing Wikidata item. Your own website doesn't count as a source; Wikidata items need external citations to survive.

The link only works in both directions. Your site's sameAs array needs to include the Wikidata QID, and the Wikidata item itself needs a P856 (official website) statement pointing back to your domain. Most businesses that get a Wikidata item stop after the first half and never check that the reverse link exists, which leaves the signal half-built. If your business doesn't yet meet Wikidata's notability bar, don't fabricate coverage to get one created; an uncited, thin item is fragile and at real risk of deletion, which does more harm than not having one at all.

Try this: open your business's Wikidata item and look for an "official website" (P856) statement. Confirm it exists and matches your domain exactly, including whether you use www.

Write Long-Form, Original Content

Length and depth directly correlate with higher citation counts specifically inside conversational answers. Short articles under 800 words average 3.2 citations from ChatGPT, while long-form pieces over 2,900 words earn 5.1. The correlation isn't with word count itself; it's that longer pieces tend to carry more original data, more expert quotes, and more sub-questions answered, all of which independently drive Information Gain.

Treat llms.txt as an Optional Site Summary

An llms.txt file can give systems that choose to read it a concise list of your site's purpose and key pages. We maintain one for this site: view our llms.txt as a live example.

There is no established citation benefit from this file. Google says its Search systems, including AI Overviews and AI Mode, do not use it. Put your effort first into accurate, crawlable pages, clear internal links and original evidence. Keep an llms.txt file accurate if you use one, but do not expect it to control how an AI system describes your business. See Google's guidance for AI search features.

Allow the Retrieval Crawlers

GEO cares most about the crawlers that feed an engine's live retrieval and training: GPTBot and ClaudeBot for training data, PerplexityBot for real-time retrieval. That's a slightly different emphasis from AEO, which leans more on OAI-SearchBot and Google-Extended for search-facing surfaces. In practice you want all five allowed; the AEO guide has the full robots.txt block to copy.

Don't Ignore Where Humans Already Talk About You

Do Reddit and LinkedIn mentions actually affect AI citations?

Yes. Reddit is the single most-cited domain across ChatGPT, Gemini, Perplexity, and Google AI Overviews in multiple 2026 citation-source studies, ahead of Wikipedia, YouTube, and every major news publisher. Community platforms are a real grounding source, not a vanity metric.

Generative engines increasingly treat community platforms as grounding sources in their own right, not just as backlinks. A Reddit thread where someone recommends your business, or a LinkedIn post discussing your work, can end up cited directly inside an AI-generated answer. This is a channel most GEO advice ignores, because it isn't something you control the way you control your own website copy, but it's worth deliberately encouraging: respond to relevant threads under your own name, keep a LinkedIn presence active, and treat every genuine mention as inventory a model might eventually retrieve from.

GEO Within Google's Rules, and Everyone Else's

Can generative engines penalise manipulative content?

OpenAI, Anthropic, and Perplexity haven't published a formal manipulation policy the way Google did for AEO in August 2026. But RAG grounding already deprioritises low-Information-Gain content by design, so a page built purely to bait a citation tends to fail on its own terms regardless of platform.

The AEO guide covers Google's August 2026 spam policy, which explicitly targets content engineered purely to manipulate generative AI responses. GEO doesn't have an equivalent named policy yet, but the same test applies: would you publish the page if no AI system ever cited it? Content built only to be lifted into an answer, with nothing genuinely useful behind it, is exactly the near-zero Information Gain pattern that grounding systems are already built to ignore. Structuring for extractability and manipulating for citation look similar on the surface; only one of them actually works.

Common GEO Mistakes to Avoid

Treating AEO and GEO as the Same Discipline

This is the single most common confusion in current AEO/GEO discussion. AEO is extraction from an index Google already has. GEO is retrieval and generation by a model deciding what to trust in the moment. Techniques overlap; the mechanisms don't. Treating them as interchangeable means you'll optimise for one and assume you've covered both.

Assuming One Engine's Behaviour Predicts Another's

A citation in Perplexity doesn't mean you'll be cited in Claude's grounded search or Gemini's answers. Each engine retrieves differently. Test across all four rather than optimising for whichever one happens to be easiest to check.

Treating llms.txt as a Shortcut

A site summary cannot replace pages that answer real questions with evidence. If you maintain an llms.txt file, check that its links and claims still match the visible site. Spend the rest of your time improving those pages and measuring whether people find them useful.

Adding Schema Without Anchoring It

Individual pages of Organisation or Person schema with no shared @id are disconnected islands rather than one resolved entity. And a sameAs link that's stale, broken, or points at the wrong company with a similar name actively misleads entity resolution rather than helping it. Check that every external URL in your sameAs array still resolves and still refers to you.

Measuring GEO Success

Citation frequency: Manually test your core buyer questions in ChatGPT, Claude, Gemini, and Perplexity. How often does your business appear, and in which engines specifically? Track this on a schedule, not once.
AI crawler activity: Which pages GPTBot, ClaudeBot, and PerplexityBot actually fetch, and which they ignore. These crawlers don't run JavaScript, so ordinary analytics never sees them; you need server or CDN logs, or an analytics tool that ingests them. It proves access, not understanding or citation, but a page no AI crawler reads can't be cited. We log this on our own site.
Share of Model: Of a fixed set of category questions, what percentage cite you. A useful benchmark, but if it comes from API prompts rather than real sessions, it isn't what your customers see (see above).
Brand Visibility Score (BVS): A composite metric weighting citations by placement and sentiment. A headline citation counts more than a footnote mention. Keep it alongside the other measures rather than collapsing everything into one number.
Entity resolution: Confirm your sameAs links still resolve, and that any Wikidata item's P856 field still points back to your site. Both sides drift over time; a broken link is worse than no link.
Referral traffic: Watch for visits from chatgpt.com, perplexity.ai, and claude.ai in your analytics. Not every citation drives a click, but the ones that do are worth isolating from regular traffic; we classify AI-agent visits as a separate category on this site for exactly that reason.
Industry benchmarks: Conductor's 2026 AEO/GEO Benchmarks Report gives sector averages to compare your citation rate against.

Try this: write down five questions a buyer would genuinely ask before hiring you. Once a month, run all five in ChatGPT, Claude, Gemini, and Perplexity, and log three things in a spreadsheet: were you mentioned, where in the answer, and was what it said accurate. After a quarter you have a real Share of Model trend, not a guess.

Do I Need the AEO Guide Too?

Do I need to read the AEO guide as well as this one?

Yes, if Google Search still sends you meaningful traffic. AEO and GEO share a foundation (structured data, original data, clear writing) but AEO adds Google-specific tactics: answer capsules tuned for AI Overviews, FAQ schema, and the Search Console technique for finding AI-style queries.

Read the AEO guide for the extraction-layer half of this. Most businesses need both; see "Why GEO Should Come First" above for how to sequence the two.

Frequently Asked Questions

Genuine questions we hear from businesses starting on GEO.

How do I find out if ChatGPT or Perplexity already knows about my business?

Ask them directly. Pose the actual questions your buyers would ask, "who provides AI training for businesses in Auckland", in ChatGPT, Claude, Gemini, and Perplexity, and note whether you're mentioned and what's said about you. It's free, takes minutes, and is the fastest real baseline available.

Is GEO a one-time setup, or does it need ongoing work?

Ongoing. Retrieval indexes refresh, competitors publish new content, and model providers periodically change how they ground answers, so a citation you have today isn't guaranteed to persist. Treat citation testing as a recurring check, not a launch task you complete once.

Can I do GEO myself, or do I need outside help?

Most of it, clear identity information, original data, useful answers and internal links, is doable in-house using the guidance in this article. Where businesses usually get stuck is having someone dedicated to checking per-engine citation results on a schedule and adjusting, which is the part most often outsourced.

Does GEO matter for a small NZ business, or only large global brands?

Potentially more. A small business competing for a narrow, local query faces a much smaller pool of retrieval candidates than a broad global term, so the entity-clarity and Information Gain work in this guide can matter disproportionately for smaller, geographically specific businesses.

The Bottom Line

GEO isn't a checklist you finish once. It's an ongoing bet that the model retrieving candidates for your category will keep finding genuinely useful, well-attributed, structurally clear content from your site and choose it over a competitor's. The businesses building that habit now, original data, named expertise, clean structure, will have compounding citation advantages once Share of Model becomes as contested as Google rankings are today.

Ready to get your business cited by AI, not just found on Google?

AI Innovisory's AI workshops cover both AEO and GEO: structured data, entity verification, and measuring Share of Model across ChatGPT, Claude, Gemini, and Perplexity.

Book a GEO workshop for your team

Generative Engine Optimisation (GEO): Getting Cited by AI