Concepts

Grounding

Grounding is the retrieval step in which an AI system pulls outside pages into its working context before it writes an answer. Joe Ashta, a GEO / AEO consultant for law firms, treats it as the first stage of the sequence that ends with a firm being named. This page explains the difference between an answer written from memory and one written from retrieved pages, why the buyer cannot see which they received, and what a law firm can do to be in the retrieved set.

Two kinds of answer

A language model holds what it learned in training as parametric memory: patterns and facts encoded in the model itself, fixed at the point training ended. Asked "who is a good DUI lawyer in Vancouver" with no retrieval, the model answers from that memory. It may name a firm that was written about often enough to be remembered, it may name one that has since closed, or it may produce a plausible name that belongs to no firm at all. Nothing in the answer distinguishes the three cases.

A grounded answer starts with a search. The system forms one or more queries from the buyer's question, retrieves a set of pages, reads them, and writes the answer from what it read. The glossary calls the pages it retrieved the grounding set. The firms such an answer names are firms that appeared in that set, either on their own pages or on pages about them. The citations shown beneath the answer are the visible trace of which pages were used.

The two kinds of answer name firms for different reasons. An answer from memory names a firm because it was well known when the model was trained. A grounded answer names a firm because a page retrieved today says so. A firm with no presence in training data can be named in a grounded answer next week. A firm that dominated training data can be missing from one, because the pages retrieved today do not mention it.

Why grounding is invisible to the buyer

The buyer sees a paragraph and, on most surfaces, some source links. The buyer does not see the queries the system ran, the pages it retrieved and discarded, or whether it retrieved anything at all. Two answers to the same question can look identical, one grounded and one from memory, and the buyer reads both with the same confidence.

This has a consequence for a law firm. A firm cannot tell from the answer why it was left out. It may have been absent from the grounding set, present in the set and discarded at ranking, or present and used without being named. The remedies differ, and the AI search visibility page lays out the stages at which each failure happens. Grounding is the first of them, and it is the one to rule out before anything else, because a firm whose pages are never retrieved cannot be cited, and is rarely named except from memory.

Getting into the grounding set

Retrieval needs relevance. The system is looking for pages that answer the buyer's question, and it selects from what it can reach. Four conditions follow.

The page exists. A firm that defends impaired driving charges and has no page that says so, in those words, has nothing to be retrieved. A practice area described in one line on a services page is close to nothing.
The page is crawlable. It is indexable, loads without a script that hides the text, is not blocked to the crawlers the platforms use, and is linked from the rest of the site. Retrieval systems read the web the way search engines do, and a page a search engine cannot index is a page an answer engine cannot retrieve.
The page is specific to the question. A page on the immediate roadside prohibition review process in British Columbia matches a question about a DUI in Vancouver. A general page on criminal law does not. Specific pages are retrieved for the questions they answer; general pages are retrieved for nothing in particular.
The page is attributed. It names the firm, the lawyer who wrote it, and the city, in the text and in structured data. An attributed page carries the firm's name into the context along with the answer. An anonymous page contributes the answer and leaves the firm behind.

Pages about the firm count too. A law society listing, a legal directory entry or a news article that states the firm's practice and city can be retrieved for a question the firm's own site would never match. The firm does not write those pages. It can make sure the facts on them are correct.

Necessary and not sufficient

Being grounded gets a firm into the room. It does not get the firm named. Once the pages are retrieved the system still has to weight them against competing pages, resolve which firm they belong to, and decide whether that firm is the one to put forward. A firm can be in the grounding set for every relevant question in its city and never appear as a recommendation, because the retrieved pages describe the law well and describe the firm inconsistently.

The entity record on this site exists for the stages after grounding: it gives a system that has already retrieved a page about Joe Ashta one canonical statement to resolve it against. For a law firm the equivalent is a single, consistent record of name, place and practice that every retrieved page agrees with. Grounding gets the pages read. The record gets the firm recognized.

AI citations covers the trace grounding leaves in the answer. Entity authority covers the stage after grounding, where a retrieved page has to be tied to one firm. AI Mode describes a surface that grounds on every turn of a conversation.