What is Generative Engine Optimization (GEO)? The ecommerce guide

What is generative engine optimization?

The term comes from a 2023 paper by researchers at Princeton, Georgia Tech, IIT Delhi and the Allen Institute for AI, titled exactly that: GEO: Generative Engine Optimization. The authors (Aggarwal et al., no relation, though I did check) defined a generative engine as a search system that synthesises an answer from multiple sources rather than ranking them, then measured which content-side changes made a source more visible inside those answers. Their headline result: the right changes lifted source visibility by up to 40% in generative engine responses. Citing sources, adding statistics and quoting authorities all moved the number. Keyword stuffing, the oldest SEO reflex there is, did roughly nothing.

That result matters because it was the first controlled evidence that visibility inside AI answers is earnable. It is not a lottery, and it is not a mirror of your Google rankings. It responds to specific, unglamorous work on the page: evidence, structure, attribution. This guide is that finding applied to a product catalogue.

A definition you can paste into a planning doc: GEO is the practice of increasing the rate at which AI engines retrieve your pages, trust what they read, and cite your brand in the answers they generate. Three verbs, three failure points, three workstreams. Most brands fail at the middle one, because trust is built from evidence the model can check, and marketing copy is not that.

GEO, AEO, LLMO: is this one discipline or three?

The naming situation is a mess, so let us settle it. GEO is the umbrella and the term the research literature uses. AEO (answer engine optimisation) names the same discipline from the engine's side and tends to emphasise question-and-answer mechanics: being the source an answer engine quotes. LLMO, AI SEO and GAIO are the same thing again, coined by people who arrived at the same problem independently. The work underneath every acronym is identical.

On this blog the tactical layer lives under the AEO name, because that is what we called it when we started publishing. The answer engine optimisation playbook covers the four levers (JSON-LD coverage, llms.txt, conversational PDP structure, a verified-review evidence layer) in working detail, with a 90-day programme attached. Treat GEO as the strategy word and AEO as the playbook word. Buy one, get the other.

A spelling note, since someone always asks. The term of art arrives with an American z, so Generative Engine Optimization it stays, even on a blog that otherwise optimises with an s. Consistency lost, findability won.

How is GEO different from classic SEO?

The workflow looks familiar: audit, fix pages, measure, repeat. The rules underneath changed. Classic SEO optimises a page to rank on a results page and earn a click. GEO optimises a page to be read by a language model and quoted inside an answer the buyer may never click away from.

The deepest difference is what authority means. Google's classic model inferred it from links. Generative engines behave more like a cautious journalist: they quote what they can defend. In our 90-day study of 1,200 prompts across five engines, third-party domain authority barely predicted citations on Perplexity, while verified-review depth was the strongest single predictor everywhere we looked.

None of this makes SEO obsolete. Retrieval still runs on an index, and for Google AI Overviews the candidate set still leans on organic results. GEO sits on top of SEO the way conversion optimisation sits on top of traffic acquisition. If your real question is how to split effort and budget between the two, that gets its own decision piece.

Why are UGC, reviews and structured data the ecommerce GEO inputs?

Engines weight evidence by how hard it is to fake. Brand copy is free to write, so it earns almost nothing. A verified-purchase review with an author, a date and a rating distribution is falsifiable, so it gets quoted. Across our prompt panel, PDPs carrying complete Review JSON-LD earned citations at roughly fourteen times the rate of pages showing text-only review counts, the largest multiplier we measured. The full argument is in why AI engines treat verified reviews as evidence.

UGC does a related job: third-party corroboration. A customer photo or video of the product in use is a claim the brand did not author, which is exactly the kind of source a model reaches for when it wants to say something concrete about fit, texture or size. For that content to count it has to be rights-cleared, tagged to the product it shows, and server-rendered so a crawler actually receives it. Client-side-only galleries are invisible to every engine we have tested.

Structured data is the contract that makes all of it machine-legible. Product, Offer, AggregateRating, Review and FAQPage JSON-LD hand the engine facts it does not have to infer: price, availability, rating, answers. A system whose worst failure mode is hallucinating a price will always prefer the page that states it in a schema block.

How do ecommerce brands get cited by ChatGPT, Perplexity and Google AI Overviews?

The uncomfortable finding from our panel: the engines disagree with each other. Perplexity cited a specific brand in 52% of shopping prompts, ChatGPT in 38%, Copilot in 18%. Optimising for one engine and assuming the rest follow is the most common first mistake, and it fails quietly, because nobody is watching the engines they did not optimise for.

Perplexity rewards structure and freshness over link equity, which makes it the friendliest engine for smaller brands doing the schema work properly. ChatGPT is the most volatile, and the one where retrieval hygiene (a feed, an llms.txt, server-rendered PDPs) pays off most directly. Google AI Overviews draw their candidate set largely from pages that already rank, then quote the most liftable passage; the Google-specific path, including the schema shapes the Overview actually pulls, is covered in how to appear in Google AI Overviews.

The common core across all of them is the same four-lever work from the AEO playbook. Do that once, properly, and the per-engine differences become tuning rather than separate programmes.

How do you measure GEO?

Citation rate is the KPI. Build a panel of 100 to 300 prompts phrased the way shoppers in your category actually ask, run it weekly across the major engines, and log four things: whether you were cited, where in the citation list you sat, whether the quoted facts were accurate, and what share of all cited brands you took. A flat spreadsheet is fine for the first quarter.

Pair the prompt panel with a page-level audit so you know what to fix when a prompt goes wrong. The 42-signal AEO scorecard rates any PDP for citation-readiness in about ten minutes; the median PDP we audit scores 18 out of 42, which tells you how much room the average catalogue still has.

Where to start: the first 30 days

1. 1Baseline. Run 25 category prompts across ChatGPT, Perplexity and Gemini. Record every brand cited. This 45-minute exercise usually settles the internal debate about whether GEO matters.
2. 2Audit one money page. Score your top-traffic PDP against the 42-signal scorecard. Fix the reds in schema coverage first.
3. 3Ship llms.txt. Six sections, one afternoon. Point it at your catalogue, your review index and the URLs you want quoted.
4. 4Surface verified evidence. Get the first five reviews server-rendered on the PDP with Review JSON-LD behind them, not hidden in a tab behind JavaScript.
5. 5Restructure the opening. Rewrite the first 100 words of that PDP to answer "what is this and who is it for" before anything else.
6. 6Re-run the 25 prompts. Log the movement. Schedule the panel weekly from here on.