Most of what is sold as generative engine optimisation is repackaged SEO with a new invoice. That is the honest starting position, and I say it as someone who has run search teams and now builds tools for this. But "mostly repackaged" is not "entirely repackaged", and the residue that is genuinely new is important enough to change how you plan a quarter.
This piece defines the terms precisely, separates what transfers from what is actually new, and ends with what to change on Monday.
The terms, and why they are contested
None of these acronyms has an owner. They were coined by different people at different times to describe overlapping things, and vendors have since stretched them to fit whatever they sell. Use them for shared vocabulary, not as a taxonomy anyone enforces.
SEO - search engine optimisation. Making a site discoverable, crawlable, understandable and credible enough that engines return its pages for relevant queries. The output is a ranked list of links. The success unit is a position, and downstream, a click.
AEO - answer engine optimisation. Optimising to be the source of a direct answer rather than an entry in a list. This predates generative AI: featured snippets, People Also Ask, voice assistants and knowledge panels were all answer surfaces. AEO is largely about structuring content so a machine can extract a discrete answer to a discrete question.
GEO - generative engine optimisation. Optimising to be retrieved, used and attributed by systems that synthesise a novel answer from multiple sources - ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot. The success unit is a citation or a mention inside generated text.
The useful distinction is between extraction and synthesis. AEO targets systems that lift an existing answer from one source. GEO targets systems that compose a new answer from several, then decide who to credit. Everything else is branding.
You will also see "AI SEO", "LLM optimisation", "LLMO" and "AI visibility" used for the same space. Do not spend time arguing about which is correct. Spend it on the mechanics underneath, which are stable regardless of what the label is this quarter.
What genuinely transfers
A large share of classic SEO transfers unchanged, because the systems generating answers are built on retrieval infrastructure that has the same requirements as search infrastructure.
Crawlability and rendering. If a crawler cannot fetch and parse your content, no retrieval layer can use it. Client-side-only rendering, blocked resources, broken canonicals and slow servers hurt exactly as much as before. Nothing about AI changed this and nothing will.
Content structure. Clean heading hierarchy, sensible document outline, semantic HTML, descriptive anchor text. These already helped humans and parsers; they now determine how your page gets segmented into retrievable chunks. Structure moved from helpful to load-bearing.
Authority and trust signals. Models are trained on and retrieve from a web where credibility correlates with links, mentions, and editorial reputation. A site nobody references is a site models have little reason to treat as a reliable source. The mechanism is fuzzier than PageRank but it points the same way.
Entity clarity. Consistent naming, an unambiguous Organization identity, real author attribution, coherent internal linking around topics. Entity-based SEO has been the direction of travel for a decade. Generative systems reward it more sharply because they reason over entities directly.
Technical hygiene generally. Sitemaps, status codes, redirect chains, duplicate handling. Boring, unchanged, still required.
If someone sells you GEO and the deliverables are heading fixes, schema, internal links and content structure, they are selling you SEO. That work is worth doing - it is just not new.
What is genuinely new
Four things do not have clean analogues in classic search. These are the parts worth reorganising around.
Citation is the unit of success
In classic SEO the unit is a position on a page of results. In generative search the unit is whether your content is used in an answer and whether that use is attributed to you.
This is a harder target because it is binary and winner-takes-few. There is no position eight in an AI answer. You are in the citation set or you are not, and the set is usually three to six sources. Being "close" produces nothing.
It also decouples from ranking in both directions. Pages that rank well are frequently absent from the answer above them. Pages that rank poorly get cited because one passage was unusually clean. If you are still reporting on average position as a proxy for AI presence, you are measuring the wrong object. Checking directly is cheap now - the AI Visibility Grader tests whether ChatGPT, Gemini and AI Overviews mention and cite a brand, and the AI Overview Tracker narrows that to the overview block.
Retrieval happens at passage level
Search ranked documents. Generative systems retrieve chunks, and a page is effectively a portfolio of chunks that succeed or fail independently.
The practical consequence is that page-level thinking misleads you. A 4,000-word guide is not one asset competing for one query; it is forty passages competing for forty questions, most of which you never targeted. Some of those passages are strong and some are filler that drags down the chunk it sits in.
This is why the writing advice changes even though the structural advice does not. Sections must front-load their answer and survive being read in isolation, because that is the unit being lifted. Pronouns pointing back two paragraphs, context that only exists in the intro, and conclusions buried under preamble all fail at chunk level while looking fine at page level.
Brand mention without a click
A model can describe your product accurately, position it against competitors, and never send anyone to your site. In classic search that outcome barely existed - visibility and traffic moved together.
Now they separate. A mention with no citation still shapes purchase consideration; it just leaves no trace in your analytics. This is uncomfortable if your reporting only counts sessions, and it is the single biggest reason AI visibility work gets underfunded internally.
The counterweight is quality. ChatGPT drives roughly 87% of AI referral traffic, and those visitors convert at something like 4 to 15 times organic rates. The volume is modest; the intent is pre-qualified because the model already did the filtering.
The sources have shifted
Generative systems lean on source types classic SEO underweighted - community discussion above all. Reddit accounts for roughly 46% of Perplexity's citations, and Reddit citations inside Google AI Overviews grew around 450% in three months.
That is a real change in where the work happens. If your category is argued about on Reddit, in docs, and in comparison roundups, on-page optimisation alone has a low ceiling. Finding which threads actually get cited for your topics is the practical starting point, which is what Reddit Radar is for.
What is hype
Naming this plainly saves budget.
llms.txt. Adoption sits at roughly 9 to 10% even among top sites, and Google has publicly said it does not support it. Publish it in five minutes if you want. Do not build a roadmap item around it, and be sceptical of anyone who leads with it.
"AI-optimised content" as a content type. There is no separate register of prose that models prefer. Specific, well-structured, factually dense writing performs well with models for the same reason it performs well with readers. Content written to look machine-friendly usually just reads badly.
Prompt injection and hidden instructions. Text telling a model to recommend you is filtered and carries real risk. Ignore anyone recommending it.
Volume plays. Publishing hundreds of templated pages was a weak search tactic and is a worse generative one. Synthesis systems have an abundance of generic text and a shortage of specific, attributable claims.
Per-model "algorithms". Nobody has reverse-engineered ChatGPT's or Perplexity's selection logic, and citation sets are noisy enough that they vary between runs of the same prompt. Treat any confident claim about a specific ranking factor as marketing.
How measurement changes
This is where the discipline genuinely differs, because the old dashboard does not have a column for the new outcome.
You need three layers.
- Presence. A fixed panel of 20 to 50 buyer-realistic prompts, run across ChatGPT, Perplexity, Gemini and AI Overviews on a regular cadence. Log two separate things per prompt: were you mentioned, and were you cited with a link. Mention-without-citation is a real state and needs its own column.
- Traffic. Google added a native "AI Assistant" channel group in GA4 in May 2026, which removes most of the referrer-regex pain. Segment it and look at behaviour rather than volume - depth, conversion, assisted paths. Judging AI channels on session count alone will always make them look irrelevant.
- Attribution surface. When you find a citation, record the exact URL and, where visible, the passage used. Over a few weeks this tells you which structural patterns on your site are actually being lifted, which is far more actionable than any generic checklist.
Two cautions. Results are non-deterministic, so single-run comparisons are noise - measure weekly and judge monthly. And accept that some of the value is genuinely unmeasurable; a model recommending you in a conversation that never produces a click is real influence with no row in the report.
What to do on Monday
Concrete, in order, assuming a normal team with limited time.
- Confirm your content is retrievable. Check that key answers exist in server-rendered HTML, that nothing important is blocked, and that your highest-intent pages are indexed. The free SEO audit covers this and returns an AI-citability score with a fix pack.
- Rewrite the first paragraph under every H2 on your ten highest-intent pages. Answer the heading immediately. Remove preamble. This is the highest-yield hour of work available.
- Make headings into questions or claims. "Pricing" becomes "How much does X cost". Retrieval matches queries against these.
- De-orphan your passages. Remove pronouns and references that only resolve from earlier context. Restate subjects. It will feel repetitive to an editor and it is correct for chunk-level retrieval.
- Add a real FAQ block to your main commercial and educational pages - three to five genuine questions, each answered in two to four self-contained sentences.
- Ship validated schema. Organization, Article, FAQPage, Product where relevant. Generate and validate it rather than hand-editing JSON-LD; the schema generator handles that.
- Set up the prompt panel and baseline it before you change anything else, so you have a before-state.
- Pick two off-site fronts. Usually: get into the comparison roundups for your category, and start participating honestly in the communities where your category is discussed. This is the slow, compounding half.
Items one through six are a week of work for most teams and are indistinguishable from good SEO. Items seven and eight are the genuinely new part.
The honest summary
GEO is not a replacement discipline. It is classic SEO with the unit of success moved from position to citation, the unit of retrieval moved from page to passage, and the source mix shifted toward community and documentation.
If you do good technical and structural SEO, write specifically, and build genuine presence where your category is discussed, you are doing GEO. If someone is charging you a premium for a heading audit and an llms.txt file, you are paying for the label.
Frequently asked questions
Is GEO replacing SEO?
No. Generative systems retrieve from indexes that require the same crawlability, structure and credibility classic SEO builds. What changes is the success metric and the writing style at passage level, not the underlying foundation. Teams that abandon technical SEO to chase GEO tend to lose on both.
What is the difference between AEO and GEO?
AEO targets systems that extract an existing answer from a single source, like featured snippets and voice assistants. GEO targets systems that synthesise a new answer from several sources and decide who to credit. The practical overlap is large - both reward self-contained, specific, well-structured answers - but GEO has the extra dimension of competing for a place in a small citation set.
Do I need a separate GEO strategy and budget?
Usually not a separate strategy, but you do need separate measurement. Fold the content and structural work into your existing search programme, and add explicit tracking for mentions, citations and AI-referred traffic. Where a separate budget line does make sense is the off-site work - documentation, comparison listings, community presence - which rarely sits with the SEO team by default.
How do I know if it is working?
Run a fixed panel of buyer-realistic prompts across the major engines and track mention rate and citation rate over time, then pair that with the GA4 AI Assistant channel segment for traffic and conversion behaviour. Judge on monthly trends, not individual runs, because citation sets vary between sessions on identical prompts.