Beyond GEO: what generative engine optimisation actually does, and what retailers should do alongside it.

Most retailers' agentic-commerce budgets in 2026 are buying the wrong half of the funnel. GEO vendors have multiplied, and almost all the spend is going into a single phase of an eight-phase pipeline.

Some of that allocation is correct. Most of it is misallocated. A retailer that wins the recommendation but loses the cart has been cited, not paid.

What GEO actually does

Generative engine optimisation (GEO) and Answer Engine Optimisation (AEO) is the practice of making a brand, product or content asset more likely to appear in the responses generated by large language model assistants — like ChatGPT, Gemini, Claude, and Perplexity. It is SEO's closest analogue, but the mechanism is different. Traditional search engines indexed documents and returned a ranked list of links; the user clicked, and the retailer received the click, the cookie and the chance to convert. The AI assistant reads the same documents, synthesises an answer, and delivers it directly. A citation or product recommendation in that answer is the new equivalent of a top-three Google ranking. The click, if there is one, is a secondary signal.

Seer Interactive's February 2026 study of 53 brands across 5.47 million queries measured an aggregate AI Overviews click-through rate of 2.4%, meaningfully lower than classic SERP CTR. However, brands earning citations inside AI answers captured 35% more organic clicks and 91% more paid clicks than those that did not. Analysis of 30 million model citations found Wikipedia accounting for 47.9% of ChatGPT's top-10 source citations, with Reddit, YouTube and named press dominating the rest. The visibility surface is uneven, and the levers that actually move it are countable.

The eight phases — and which one GEO actually addresses

Agent-mediated commerce can be thought of as a sequence of eight phases. At each one, a potential order can succeed or quietly fail.

Phase 1 — Consumer to agent conversation. The buyer agent (e.g. ChatGPT) gathers context about the user's stated need, prior purchases, budget and constraints, often before any retailer is consulted.

Phase 2 — Discovery. The agent assembles its initial candidate set from training data and live-retrieval citations. This is where GEO lives: making sure the brand or product surfaces in the generated answer at all. Phase 2 is one-way — the agent reads, the retailer is read about.

Phase 3 — Selection and negotiation. The agent now actively queries retailer-side systems for the candidates it has chosen to engage with: in-stock substitutes, bundles, promotional eligibility or price-match offers. Phase 3 is two-way, and the retailer's response (or silence) shapes which candidate the agent settles on.

Phase 4 — Cart construction. The agent assembles a cart against the retailer's live catalogue: confirming availability, attributes, sizing, shipping options and the all-in price to the user's postcode.

Phase 5 — Trust verification. The agent validates that the merchant is legitimate, that the payment issuer is recognised, and that the order satisfies category-specific rules (age, region, regulated goods).

Phase 6 — Payment authorisation. The agent presents a verified credential — an AP2 mandate, a Stripe Shared Payment Token, a Mastercard Verifiable Intent token, or a Visa Trusted Agent token — and the retailer accepts or declines under SCA-compatible rules.

Phase 7 — Fulfilment. The agent receives confirmation, tracks delivery, and surfaces any exceptions back to the user.

Phase 8 — Loyalty and reorder. The agent retains the merchant in the user's repeat-purchase mandate set, and either reorders automatically or treats the merchant as the preferred default for future category queries.

GEO covers Phase 2. One phase out of eight. Phase 1 sits almost entirely on the agent side — the retailer influences it only indirectly, through the same third-party citation surface that drives Phase 2. Phase 5, 6 and 7 sit with the carrier and the platform, and disappear if the consumer checks out themselves. Phases 3 and 4, and the long tail of Phase 8 are where most agent-initiated revenue is won or lost, and where the retailer has the most direct control. As agents begin to make more autonomous purchases, phases 5-7 become increasingly critical. But right now, the focus should be on building a strong foundation in Phases 3 and 4.

Why winning Phase 2 alone is not enough

A retailer that wins Phase 2 and loses the rest has been recommended, not bought. The cart fails at Phase 4 because stock and pricing signals were stale. The order is steered elsewhere at Phase 3 because a rival exposed a credible substitute and the merchant exposed nothing. The order is declined at Phase 5 or 6 because payment-token issuance is not configured for AP2 or the relevant network credential. The customer is lost at Phase 8 because no persistent reorder mandate was established; the agent's next default is a competitor.

None of these failures is a GEO problem, and none is fixed by buying another GEO tool. Discovery answers whether the retailer is in the conversation. The other seven phases answer whether the retailer is the one paid.

Retailers in 2026 tend to over-invest in Phase 2 and under-invest in the rest, partly because Phase 2 has a clearer vendor market and Phases 3 through 8 do not. Vendor presence is not a measure of importance. None of this is an argument against GEO, which is an absoutely necessary first step. It is only an argument against treating it as the entire programme.

Six things to do across the agent pipeline — not just inside GEO

Concrete moves, mapped to the phases where the failure rate is currently highest. Three for MDs, Three for CTOs.

1. (CTOs)   Make every product page machine-readable, and pair it with llms.txt (For Phase 2)

JSON-LD structured data is the discovery foundation. Digital Applied's April 2026 audit of 5,000 sites found just 8% of retail sites with schema correctly implemented across the catalogue — meaning the typical retailer is facing a catalogue-wide retrofit, not a fresh install, and the work should be scoped top-revenue categories first. The minimum schema is Product, with Offer, Brand, Review/AggregateRating, and named attributes (size, colour, material, sustainability, country of origin). Pair this with an llms.txt file in the site root — BuiltWith now tracks more than 844,000 implementations of the convention, proposed by Jeremy Howard on 3 September 2024 — so that AI assistants honouring it have a deterministic route to the catalogue, FAQs and policy pages. This is the one piece of Phase 2 work that everything else rests on.

2. (MDs)   Adjust third-party citation surface investment (For Phase 2)

AI assistants weight third-party authority heavily. Search Engine Land puts Reddit at roughly 40% across the major AI search engines, with YouTube and LinkedIn close behind. Conventional digital PR re-pointed at AI ingestion — Wikipedia, reputable Reddit threads, named press, YouTube, Trustpilot, Reviews.io, Sitejabber — is the slowest discovery lever and the highest-compounding.

3. (CTOs)   Make the feed and stock signals agent-grade (For Phase 4)

This is where most "ranked but did not sell" failures live. An agent constructing a cart needs accurate, real-time data at SKU level: stock, all-in price including shipping to the user's postcode, returnable status, matching variant attributes. Google Merchant Center remains the most common upstream of AI assistants' product knowledge, including Google AI Mode. A feed with missing GTINs, stale prices or thin attributes will be selectively de-recommended even when the underlying site ranks well. Treat the feed and the live product API as a primary quality target, not a downstream artefact of the catalogue.

4. (MDs)   Ask the commerce platform what is switched on, and switch on what is missing (For Phase 5 & 6)

Phases 5 and 6 sit largely with the commerce platform (Shopify, BigCommerce, Salesforce Commerce Cloud, Adobe) and the payment processor. The retailer's job is not to implement the Agentic Commerce Protocol or the Universal Commerce Protocol; it is to find out what the platform supports today, ask for an end-to-end test order against an agent client, and confirm that merchant-of-record settings, fraud screening (e.g. Stripe Radar for Agents) and accepted payment-token formats (e.g. AP2, Shared Payment Tokens, Mastercard Verifiable Intent, Visa Trusted Agent Protocol) are configured for live agent traffic. This is a vendor conversation, not an engineering project, but silent declines at Phases 5 or 6 are invisible in standard reporting.

5. (MDs)   Organise at least one negotiation or substitution capability (For Phase 3)

Phase 3 is where mid-market retailers will differentiate over the next twelve months. Microsoft Research's Magentic Marketplace simulations of 100 customer agents against 300 seller agents recorded customer agents accepting the first acceptable proposal between 80% and 100% of the time — the first seller to respond credibly takes the order. Pick one capability that fits the brand: in-stock substitute recommendations, delivery options, bundle-discount logic, a competitor price-match routine, or eligibility for a loyalty perk surfaced at agent-checkout time. The retailer that exposes a real Phase-3 response first in any category sets the bar for the rest.

6. (CTOs)   Lay foundations for the persistent purchase mandate (For Phase 8)

The buyer agent will increasingly hold persistent mandates from its user — "re-order this every month", "find me one of these when stock is low". Winning a mandate is a year-long arc, and the full version (subscribable reorder cadence, agent-readable post-purchase data, returns that do not break the agent's representation of the merchant) is an architectural commitment. The mid-market move this year is to pick one sub-component and ship it: usually a simple reorder-link or subscription surface, on the grounds that it is the lightest engineering lift and the most likely to be recognised by an agent today.

Where retailers go from here

Eight phases. GEO addresses one. The retailers who treat that ratio as a roadmap will compound; the retailers who stay inside the visibility dashboard, optimising the one phase the market already sells tools against, will spend the next two years wondering why share-of-voice gains never translated into revenue.

The work is not exotic. The order of operations is the part most retailers get wrong. Start with discovery. Do not stop there.