Glossary of AI search and shopping terms

Search marketing, machine learning and retail each have their own vocabulary, and AI shopping uses all three. The same word often means different things in a vendor deck and in a paper. These definitions are written to be checked: where a term is one the Dopbot Index measures, the entry gives the rule it uses, and where a term belongs to another company's product, the entry paraphrases that company's public documentation as of October 2026.

Terms A to B

Agentic commerce

Shopping in which an AI system acts for the shopper, for example by comparing offers, filling a cart or completing a checkout, rather than only answering. Examples are protocols such as the Agentic Commerce Protocol and the Universal Commerce Protocol, described below. The index samples recommendations in answers. It does not measure agent-led purchases.

Agentic Commerce Protocol (ACP)

An open standard that OpenAI documents as the connective layer between merchants and ChatGPT users. Its documentation says it lets ChatGPT ingest structured catalog data, understand merchant inventory and surface relevant products in context (as of October 2026).

Agentic search

Search in which the model plans its own queries, reads what comes back and decides whether to search again, instead of sending one query to a search tool and passing along what it returns. OpenAI's API documentation lists three kinds of actions in this mode: running a search, opening a page and searching within an opened page.

AI crawler user agents

Names that a bot sends with each request so that a site owner can allow or block it in robots.txt. As of October 2026 the vendors document different roles:

• OAI-SearchBot (OpenAI) is used to surface websites in ChatGPT's search features.
• GPTBot (OpenAI) crawls content that may be used in training OpenAI's models.
• ChatGPT-User (OpenAI) visits pages when a user asks; OpenAI says robots.txt rules may not apply to those actions and that it is not used to decide whether content appears in Search.
• PerplexityBot (Perplexity) is designed to surface and link websites in Perplexity search results.
• Perplexity-User (Perplexity) supports user actions within Perplexity.

Because the roles differ, one blanket rule does not mean one thing. The index does not measure crawling.

Alternative prompt

The fifth question in each category of the October 2026 edition, asking for an alternative to a named brand. It produced 316 answers that recommended 5.0 brands on average, against 2.3 for the other four questions (both engines together). The named brand was found in 296 of those answers and was recommended in 21 of the 296. Because the question names a brand and inflates counts, brand-count measures exclude it. In the 40 US categories updated from November 2026 it is replaced by a buying-advice question.

Answer

One reply from one engine to one prompt. It is the unit the index counts. October 2026 attempted 1,600 answers and collected 1,587, because 13 Perplexity answers failed.

Answer engine optimization (AEO)

A term for work aimed at being the answer an AI search experience gives. Vendors use it in different ways. Google's guidance describes AEO and GEO as terms used for work focused on improving visibility in AI search experiences, and says that from Google Search's perspective this is still SEO.

API versus app

The index samples developer APIs because that allows identical questions on a fixed schedule and a record of sources. Apps may differ in model version, instructions, search tools and personalization. Perplexity's documentation says its API uses the same search system as its interface with differences in configuration, so outputs may differ. See API answers vs app answers.

Average position

Over the answers that recommend a brand, the mean place the brand takes among the brands that answer recommends, where 1 is first. A brand that an answer only mentions takes no place. The index shows it only for brands recommended in 3 or more answers, with the count, and it never decides a rank.

Brand extraction

How the index turns text into counts. A model reads each answer and lists the brands it names, in order of first appearance, marking each as recommended or only mentioned. Retailers and marketplaces are excluded, at most 15 brands are kept per answer, names are normalized, and a hand-kept list merges names that belong to one brand. A brand counts once per answer.

Buying-advice prompt

The fifth question in the 40 US categories updated monthly from November 2026. It describes a buying situation and asks what to look for and what to buy, naming no brand and no retailer. It replaced the alternative prompt.

Terms C to G

Category

A product group the index samples, such as carry-on luggage or mattresses, with five prompts each. The October 2026 edition has 160: 100 in the United States, 30 in Mexico (questions in Spanish) and 30 in Canada. See the index.

Citation

A link an assistant attaches to a statement in its reply. OpenAI's API returns it as a url_citation annotation holding the URL, title and location of the cited source. A citation is not the same as a retrieved source: see that entry. In its own figures the index says "source URLs" and counts per engine.

Edition

A fixed sample design. The October 2026 edition asked each prompt once per engine, about 10 answers per category, in all 160 categories. From November 2026, 40 US categories are re-sampled monthly with 40 answers each, and the other 120 stay published as the October 2026 edition. Numbers from the two designs are not like for like.

Engine

An assistant sampled through its developer API: OpenAI's Responses API with web search (gpt-6-luna, 800 answers in October 2026) and Perplexity's Agent API with the Sonar model (perplexity/sonar, 787 answers).

Engine overlap

How much two engines' brand lists agree. Per category, brands recommended by both divided by brands recommended by either (Jaccard), averaged over categories with equal weight. Excluding the alternative prompt it averaged 47.0% in October 2026 (53.6% with it), with 3.9 brands from both engines and 5.3 from only one per category. It is not answer-by-answer agreement. See OpenAI vs Perplexity.

Generative engine optimization (GEO)

A 2023 research paper by Aggarwal and colleagues introduced the term as a framework for helping content creators improve the visibility of their content in generative engine responses, tested on a benchmark it calls GEO-bench. Google's guidance treats it as a term for work that is, from Google Search's perspective, still SEO.

Google-Extended

A robots.txt token, not a separate crawler. Google documents it as something web publishers can use to manage whether content Google crawls may be used for training future Gemini models and for grounding in some Gemini products, and says it does not affect inclusion in Google Search and has no separate HTTP request user agent string (as of October 2026).

Grounding

Giving a model content at the time of the prompt, such as search results, so its answer can rest on current sources and not only on its training. Google describes grounding as providing content from the Google Search index to the model at prompt time to improve factuality and relevancy. Both engines are called with web search turned on, and the index's OpenAI requests require the model to search. Grounding does not make an answer correct.

Terms H to P

Intent

The kind of shopping question a prompt asks. The index uses best overall, budget (with a price ceiling), premium, a specific use case, and a fifth type: alternative in October 2026, buying advice from November 2026. Brands per answer by intent in October 2026, both engines together: 2.4 for best overall, 2.2 for budget, 2.2 for use case, 2.6 for premium and 5.0 for the alternative question.

Jaccard overlap

For two sets of brands, the number in both divided by the number in either, as a percentage. Two identical lists score 100%, and two lists with nothing in common score 0%. The index uses it for engine overlap and for the overlap of budget and premium shortlists (8.2% on average per category, both engines together, October 2026; see budget vs premium).

Large language model (LLM)

A model trained on large amounts of text to predict and generate text. An assistant is an LLM plus instructions and, often, tools such as web search. What an LLM knows from training and what it reads from search results at answer time are different inputs.

llms.txt

A proposal, published at llmstxt.org, to add a markdown file at /llms.txt that gives LLM-friendly background and links. Google's guidance on AI features says you do not need such files to appear in Google Search. The index has not measured whether any assistant uses them.

Market

One of the three countries the index covers: the United States, Mexico or Canada. A market sets the country location sent with each web search and the language and wording of the questions. Mexico's questions are in Spanish and its categories differ from the others'.

Mention

A brand named without being presented as a pick, for example in a comparison or a warning. The index does not count a mention as a recommendation.

Merchant Center

Google's tool for merchants to submit product data. Google's specification says the data is used to match products to queries and as an input for ads in its AI-powered formats, and its guidance on AI features says responses can include product listings and product information.

Model Context Protocol (MCP)

An open-source standard for connecting AI applications to external systems such as data sources and tools.

Non-determinism

Asking the same question twice can return different answers. The index's methodology says answers vary from run to run, even for the same question. It is why a count has a margin of error and why one answer is a weak basis for a conclusion.

Own-site match

An answer whose source list holds a domain whose first label equals a recommended brand's name with spaces and punctuation removed (samsung.com for Samsung). It misses country domains, parent companies and other store names. In the October 2026 edition, OpenAI's answers linked one in 46.6% of 800 answers; in Perplexity's retrieved lists one appeared in 62.0% of 787 answers (all prompts).

Personalization

Tailoring an answer to an account, history, memory or location. The index's requests carry no account, history or memory, only the market's country. Apps may personalize, so an app answer can differ from an API answer.

Product feed

A file or data stream of a merchant's product attributes, such as title, price and availability, sent to a platform. Google's Merchant Center has a product data specification. OpenAI's help center addresses merchants interested in providing a direct product feed to OpenAI. The index does not measure the effect of feeds.

Prompt

The exact question text sent to an engine. Each category has five, and the page for each category lists them as they were asked. Changing a prompt starts a new series, because the counts are no longer like for like.

Terms Q to Z

Query fan-out

A technique where one user question becomes several related searches. Google's documentation says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources, and OpenAI's help center says ChatGPT search typically rewrites a question into one or more targeted queries. The query a search system sees can therefore differ from the shopper's wording.

Rank

Position by count of answers recommending the brand and nothing else. Brands with the same count tie and share a rank: two brands tied for first are both 1 and the next is 3.

Recommendation

An answer presents the brand as a pick. It is what the index counts, at most once per answer.

Retrieval-augmented generation (RAG)

A design in which a system first retrieves documents and then writes an answer from them. A 2020 paper by Lewis and colleagues describes RAG models as combining pre-trained parametric and non-parametric memory for language generation. Google's guidance uses RAG, also called grounding, and describes relying on its Search ranking systems to retrieve relevant, up-to-date web pages from its Search index.

Retrieved source

A page that an engine's search retrieved while answering. For Perplexity the index records, almost always, every page its search retrieved, and the API does not say which ones the answer relies on. In October 2026, Perplexity's retrieved lists held 27.3 source URLs per answer on average (15.2 distinct domains, 787 answers); OpenAI's answers linked 2.2 pages (2.0 distinct domains, 800 answers). A domain in a list does not show that it influenced a recommendation.

robots.txt

A file at a site's root that tells automatic clients, called crawlers, which parts of the site they may access. It is specified in RFC 9309, which describes it as a way for service owners to control how content may be accessed by crawlers. It is a request that compliant crawlers follow, not access control, and OpenAI documents that robots.txt rules may not apply to some user-initiated fetches.

Sample

One answer to one prompt on one engine; n is the number of them. October 2026 had about 10 answers per category (5 per engine). The monthly update has 40 (20 per engine).

SEO

Search engine optimization: work that helps pages be found in search. Google says that the best practices for SEO remain relevant for AI features in Google Search.

Share of answer

Answers that recommend a brand divided by answers sampled, for a stated engine, question set and period. Report it as "k of n" with an interval. It is a rate, so shares across brands add up to more than 100% when answers recommend several brands. To count it yourself, see the free manual method.

Source list

The web pages recorded for one answer. For OpenAI it is the pages its answer links. For Perplexity it is the pages its search retrieved. The index reports the two per engine and never adds them together, because they record different things.

Source type

How the index classifies a domain in a source list: retailer, forum or video, or own site, among others. A hand-made list decides which domains are retailers, so it misses some. In October 2026 (all prompts), retailer domains appeared in the source lists of 17.6% of 800 OpenAI answers and 76.7% of 787 Perplexity answers; forum or video domains in 1.3% and 3.3%.

Structured data

Markup on a page that describes its content to machines. Google says it uses structured data it finds to understand a page, and that most of the structured data it uses relies on the schema.org vocabulary. For products, Google separates product snippets, for pages where people cannot buy directly, from merchant listings, for pages where customers can purchase from you. The index has not measured whether markup changes recommendations.

Universal Commerce Protocol (UCP)

An open standard for agentic commerce. Google's support page says it works across the shopping journey, from discovery and buying to post-purchase support, and that the protocol will be available in phases, with checkout available for select merchants at the time of that page. Google's developer page says adopting it enables agentic actions in AI Mode and Gemini, starting with direct buying (as of October 2026).

Visibility

Short for AI shopping visibility: how often a brand is recommended in answers to shopping questions, per engine and question type. See What is AI shopping visibility?.

Web search tool

The tool that lets an API model search the web. In OpenAI's Responses API the model can choose whether to search based on the prompt, and a request can carry an approximate user location. The index sets the location to the market's country, turns the web search tool on for both engines, and requires OpenAI's model to use it.

Wilson interval

A range for a true rate, given a count out of a sample, that works well for small samples and stays between 0% and 100%. The index uses the 95% Wilson interval: for example, a brand recommended in 20 of 40 answers could have a true rate of about 35% to 65%. For how many answers a change needs, see how many AI answers to trust a change.

Zero-brand answer

An answer that recommends no brand. The index counts these answers but does not record what they offer instead. In October 2026, excluding the alternative prompt, 4.8% of OpenAI answers (640) and 1.9% of Perplexity answers (631) recommended no brand. Among the five question types, best overall had the highest share at 9.1% of 318 answers (both engines together).