Plain definitions of the terms that get thrown around in AI search, AI integration and agentic commerce — written so you can tell whether the person using one at you knows what it means.
Where a term is genuinely contested, or where the evidence behind it is thin, we say so rather than picking the version that sells better. Every figure is sourced.
Where the answers appear
The places a buyer now meets a recommendation, and the behaviour that comes with them.
- AI Overviews
Google’s AI-generated summary above the traditional results. Only 38% of pages it cites rank in the top ten. - Google AI Mode
Google’s dedicated AI search experience, in its own tab. Not the same thing as AI Overviews. - Local pack
The map and the three names Google shows for a local search. Still the highest-intent surface there is. - Zero-click search
A search that ends without a click. 68% of US Google searches in early 2026.
What the work is called
The labels agencies sell. Two of them overlap heavily and one of them is not what most people think it is.
- Generative Engine Optimization (GEO)
Getting a business named, quoted or linked inside an AI-generated answer, rather than ranked in a list of links. - Answer Engine Optimization (AEO)
Optimizing a page so a system can extract a direct answer from it. Related to GEO, and the boundary is blurrier than most agencies admit. - E-E-A-T
Experience, Expertise, Authoritativeness, Trustworthiness. Google states plainly that it is not a ranking factor.
How the systems actually work
The mechanisms underneath. Worth understanding, because they explain why the old tactics stopped paying.
- Retrieval-augmented generation (RAG)
The model searches first and writes from what it finds. Being retrievable now matters more than being memorable. - Query fan-out
Google’s own term for breaking one question into many simultaneous searches. The mechanism behind most of the surprises in AI search. - Semantic search
Matching on meaning rather than the exact words typed. Why a page gets retrieved for a phrasing that appears nowhere on it. - Grounding
Tying an answer to retrieved sources. The opening a local business has against a national brand. - Hallucination
A confident statement the sources do not support. The costly version is mundane: wrong hours, wrong service area. - Knowledge graph
A database of things and their relationships, rather than of documents. A recommendation is a claim about a thing. - AI crawlers
Training, indexing and live-fetching bots do different jobs. Blocking the wrong one removes you from answers.
What you actually change
The work itself. Unglamorous, mostly cheap, and where the results come from.
- Entity resolution
How a system decides that scattered references to a business are all the same thing. The cheapest high-leverage work there is. - Structured data
Machine-readable labelling. Google is explicit that it is not required for AI search and is not a ranking factor. - NAP consistency
Name, address and phone identical everywhere. No published study measures the effect, and we say so. - Google Business Profile
The free listing that controls how you appear in Search and Maps. The most load-bearing asset in local. - llms.txt
A proposed file listing your important pages for AI systems. Useful discipline, unproven effect, and nobody should sell it as a ranking factor.
Putting AI to work
The vocabulary of AI working inside a business, rather than answering questions about one. Most of it is plumbing, and the plumbing is where projects succeed or stall.
- AI integration
Connecting a model to the systems you already run, so it works inside them. Most of the job is permissions, not the model. - Large language model (LLM)
The engine under ChatGPT, Claude and Gemini. Fluent by design, factual only when grounded. - AI agent
A model that decides its own next steps and tools. More capable than a workflow, harder to predict, and needs an off switch. - Workflow automation
A defined sequence of steps that runs when something happens. Predictable, auditable, and the right answer more often than an agent. - Human-in-the-loop
A person reviews or approves the AI’s output before it takes effect. The cheapest risk control there is. - Model Context Protocol (MCP)
An open standard for plugging AI applications into tools and data. Build the connection once; many assistants can use it. - Prompt injection
Instructions hidden in content an AI reads, hijacking what it does. First on OWASP’s list of LLM application risks. - AI governance
The written rules for how a business uses AI: which tools, which data, who reviews what. Usually the first thing a cautious client asks for.
When an agent does the shopping
Commerce terms that barely existed two years ago and now decide whether a store makes the shortlist.
- Agentic commerce
Shopping where an AI agent finds, compares and sometimes buys for the shopper. The agent reads your catalog, not your homepage. - Agentic Commerce Protocol (ACP)
OpenAI and Stripe’s open standard for agents to transact with merchants. Launched for in-chat checkout; now mostly about discovery. - Universal Commerce Protocol (UCP)
The open standard Google co-developed with Shopify and major retailers. Powers checkout in AI Mode and Gemini for eligible merchants. - Product feed
A structured file of everything you sell. Once a Google Shopping chore; now the first thing AI shopping assistants read.
How any of it gets measured
Two honest instruments. Neither produces a score worth celebrating on its own.
- Share of AI voice
The proportion of a fixed set of buying questions where you get named. The only honest way to measure a non-deterministic system. - Rank grid
Local rank sampled across a map instead of at one point. A coverage map, not a score.
Missing something you have been asked to pay for? Email jimmy@foundfirstsearch.com and it gets added.
Longer reads on all of this in Insights, or see the method applied to your own business with the free AI Visibility Snapshot.