The GEO glossary.

GEO (Generative Engine Optimization)

GEO is the work of getting a brand or a product to appear, with accurate information, in the answers of AI assistants. It extends SEO: the goal is no longer a position in a list of links, but a place in a written answer.

For an online store, GEO is done product by product: the assistant recommends specific models, with a price and specifications.

AEO (Answer Engine Optimization)

Another name for the same work, centered on answer engines. GEO, AEO and "AI visibility" mean the same thing in practice.

AI visibility

The measured presence of a brand or a product in the answers of AI assistants, over a given set of queries. It is expressed as a proportion, not as a single position.

AI assistant

An application that answers in natural language from a language model: ChatGPT, Claude, Gemini, Perplexity, Grok, Mistral's Le Chat, DeepSeek, Qwen. For a shopping question it relies on its model's memory and, when web search is on, on what it reads as it answers.

Shopping query

A question a customer asks an assistant before buying: a category, a use, a budget. "Which electric bike for commuting, under €2,500?"

A neutral query names no brand. It measures whether the assistant thinks of you without being prompted.

Query panel

A stable set of shopping queries tracked over time for a brand and a market. Keeping the same panel from one week to the next is what makes the numbers comparable.

Run

One execution of a query on an assistant, producing one answer. The same assistant can answer the same question differently, so a measurement relies on several runs per query.

Shelf share

The number of answers that cite a product, divided by the number of answers analyzed for the same query, the same market and the same period. A product cited twice in one answer counts once.

It is the equivalent, in an assistant's answer, of the space held on a store shelf.

Average position

The average rank at which a product appears in the answers that cite it. The rank kept for an answer is the first one at which the product appears.

Confidence interval

The range that very likely contains the true value of a share measured on a sample. At 95%, with 256 answers and a 34% share, the interval is about ± 6 points.

A change smaller than the interval is not a movement: it is noise.

Cited source

A web page an assistant gives as the origin of a piece of information in its answer. The sources that keep coming back on your shopping queries are the ones feeding the recommendation: comparisons, reviews, media, marketplaces.

Hallucination

Information stated by an assistant that exists in no source: a price never charged, a made-up specification, an imaginary stock-out.

Accuracy verdict

The result of comparing what an answer says about a product with the store's catalog. skoup uses four: accurate, missing, distorted, wrong. A verdict always quotes the words of the answer.

Web search

An assistant's ability to query the web as it answers, instead of relying only on what its model learned. For a shopping question, it is what provides today's prices and stock.

AI assistant robots

Programs that read web pages on behalf of an assistant. Each vendor runs several, with distinct roles: training models, feeding search, or fetching a page at a user's request. At OpenAI: GPTBot, OAI-SearchBot and ChatGPT-User.

Blocking a training robot does not keep you out of the answers. Blocking a search robot does.

robots.txt

A file at the root of a site that tells robots what they may read. Its format is described by RFC 9309: for a given robot the longest rule wins, and "Allow" wins a tie.

Content-Signal

A robots.txt directive proposed by Cloudflare in 2025. It states three preferences on how content is used: "search" (search index), "ai-input" (an assistant's answer) and "ai-train" (model training). It is a declared preference, not a technical block.

llms.txt

A Markdown file at the root of a site that summarizes the site and lists its important pages for language models. It is a proposal, not a standard: its effect on answers is not established.

Structured data

A description of a page in a machine-readable format, most often JSON-LD with the schema.org vocabulary. For a product, the "Product" and "Offer" types carry the name, the price, the currency and the availability.

Product feed

A file listing the products of a catalog with their attributes: identifier, title, price, stock, variants, barcode (GTIN). The most common format is Google Shopping's, in XML or CSV.

Agentic commerce

A purchase carried out by an assistant on behalf of a customer: it searches, compares, then places the order without the customer leaving the conversation. It requires the store to expose its catalog and checkout in a format the assistant can use.

ACP (Agentic Commerce Protocol)

An open standard designed by OpenAI and Stripe and published in 2025. It describes how an assistant places an order with a merchant; it is the one used for checkout in ChatGPT.

UCP (Universal Commerce Protocol)

An open standard published by Google in 2026, designed with commerce players including Shopify. A merchant declares what it supports in a profile published at /.well-known/ucp on its site.

MCP (Model Context Protocol)

An open standard published by Anthropic in 2024 that lets an assistant connect to a tool or a data source. skoup's MCP server gives your assistant access to your measurements, alerts and tasks.

Write-back

A fix written directly into the store from the tool that found the problem. In skoup, write-back changes the description, the SEO title, the meta description and the barcode of a Shopify product. Never the price.

Last-touch attribution

The rule that credits an order to the last source of the visit that converted. For AI assistants it is read from the session's referrer, then from the utm_source parameter. Mobile apps do not always send a referrer: the resulting figure is a minimum.