How ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews actually build a response — and what it means for communications.
AI Communications, defined
AI Communications is the discipline of building authority across AI answer engines — ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews — alongside earned media, digital, and influencer channels. It combines public relations, Generative Engine Optimization (GEO), and AI-visibility research to influence the answers anyone asking now begins with.
Most communications professionals treat the AI answer as a black box. A buyer asks a question. The engine produces a response. Some brands are named, some are not, and the reasons are mysterious.
The reasons are not mysterious. They are observable, repeatable, and increasingly well-understood.
This is the ninth piece in the AI Communications series, and the explainer most communications teams have been missing: how the AI answer is actually assembled, what determines which entities get named, and what that means for the work.
The two modes of AI answer
AI engines build answers in one of two modes — or in a hybrid of both.
Trained knowledge. The engine has read a vast corpus during training, including books, articles, websites, and other published content. The answer is assembled from what the engine "knows" from that training. Fast and broad but cannot include information published after the training cutoff.
Live retrieval. The engine queries the live web at the moment of the question, retrieves the most relevant current sources, and assembles the answer from them. Slower but current.
ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews all use some mix of these two modes. The mix differs by engine, by query type, by user settings, and by the engine's confidence in its trained knowledge.
What this means for communications: an entity needs to be present in both layers — the historical training data (which trains future models) and the current retrievable web (which serves live answers today). Both layers require deliberate work. Both layers can be measured.
What happens between question and answer
Six steps, roughly.
Step 1: Parse the question. The engine identifies what is being asked, the type of answer required, and the entities or concepts involved.
Step 2: Decide the answer mode. Trained knowledge, live retrieval, or hybrid.
Step 3: Identify candidate sources. From either trained knowledge or live retrieval, surface sources considered credible and relevant.
Step 4: Rank the sources. Authority signals — citation patterns, established publication trust, source consistency, schema clarity — determine which sources rank highest.
Step 5: Synthesize the answer. Combine information from top-ranked sources into a single response.
Step 6: Render and return. Format and return to the user — sometimes with footnoted citations (Perplexity, Google AI Overviews) and sometimes without (older ChatGPT, Claude).
Every one of these steps is a point of influence. Every one of them favors entities with strong Identity, Authority, Earned Media, structured authorship, and consistent presentation across owned and earned surfaces.
What makes a source rank high
Five factors repeat across engines.
Source authority. Established publications, peer-reviewed research, encyclopedia entries, and government or institutional sources rank higher than blogs, anonymous posts, or aggregator sites.
Source consistency. When multiple high-authority sources say the same thing, the engine's confidence increases. An entity described identically across Wikipedia, the Wall Street Journal, the company's own site, and the founder's LinkedIn is described that way in the AI answer. An entity described differently across those sources gets averaged.
Source recency. Recent sources are favored for fast-moving topics. For stable topics — biographical facts, company history, foundational concepts — older sources with strong citation patterns often outrank newer ones.
Source structure. Sources with clear schema markup, defined authorship, and machine-readable formatting are retrieved more reliably than equivalent content without those signals.
Topical relevance. A source is more likely to be cited for a query if its content is closely aligned with the query's topic. Tier-one consumer titles often lose to category-specific trades because the trade is more relevant, even if the consumer title is more prestigious.
What gets named in the answer
When the engine assembles the response, it names the entities most heavily represented in the top-ranked sources for the query.
If a query asks for "the top three brands in category X," the engine examines what its top-ranked sources say about brands in that category, identifies the brands most frequently mentioned with positive or neutral framing, and names them.
The brand named in the answer is not necessarily the brand with the largest market share. It is the brand most heavily represented in the source graph for that specific query.
This is the single most important insight in AI Communications: the answer is not a popularity contest among brands. It is a popularity contest among sources, with brands as the indirect beneficiaries.
The brand wins by winning the source graph. Not by winning the consumer mindshare.
What this means for the work
One. Earned media is not dying — but the publications that drive Citation Share are not always the publications that drive impressions. The earned-media program has to be designed for the source graph, not the prestige list.
Two. Owned content is more important than it used to be. The company's own site, the founder's bylines, the company's primary research — all of these are training data for future models and retrieval candidates for current answers. The owned-content layer is a multi-year investment that compounds.
Three. Structured authority compounds harder than any other layer. A clean Wikipedia entry, a Wikidata record, a glossary inclusion in a major reference site — these are durable, machine-readable assets that the engines weight heavily and that competitors cannot easily displace.
Four. Defense matters because bad sources compound just like good sources do. A hostile article from 2018 that has been heavily backlinked is a Citation Share liability for the next decade unless the source graph is rebalanced with countervailing authoritative sources.
Five. Measurement is non-negotiable. Without continuous Citation Share monitoring, the team has no visibility into whether the work is moving the answer.
The simple version
The AI answer is built from sources. The sources are observable. The entities named in the answer are the entities most heavily represented in the top-ranked sources for the query.
Win the source graph and you win the answer.
That is the entire game.
