AI engines don't have opinions about your brand. They have retrievals.
The model pulls from a citation graph — the body of text it was trained on, the live web it can fetch from, and the sources it has learned to weight. When a buyer asks "Is [brand] reliable?" the model isn't typically reasoning from first principles. It's synthesizing what its weighted sources have said.
Knowing how that synthesis works is how you change the output.
Training corpus — the floor
Every engine starts with a frozen training corpus. ChatGPT and Claude were trained on overlapping but distinct slices of the public web, books, and licensed content. Gemini weights Google's index. Perplexity sits closer to live retrieval. Google AI Overviews synthesizes inside search.
What got into the corpus during training is the floor. It's the baseline picture the model formed before anyone asked it anything. If the corpus saw your brand mostly through a 2021 crisis cycle, that crisis is the floor — until repair work moves the model above it.
Retrieval — the live signal
Most modern engines combine the frozen corpus with live retrieval. Ask Perplexity about a brand and it tends to fetch current articles in real time. Ask ChatGPT with browsing on and it often does the same.
Live retrieval is where current PR work tends to show up first. A new tier-1 article, a Wikipedia update, a high-authority owned page — these often enter the model's synthesis on the next prompt, not the next training run.
Ranking — what gets weighted
The corpus and live retrieval together produce a candidate pool. The model then weights sources by authority signals the engine has learned to value: domain authority, recency, citation density, editorial signal, schema. A brand mention in Reuters typically outweighs a brand mention on a low-authority blog by a wide margin. A Wikipedia citation tends to outweigh both for certain prompt types.
Synthesis — the answer
The model compresses the weighted candidate pool into a paragraph. That paragraph is what the buyer reads.
Synthesis is where two failure modes typically appear. Compression failure — the model picks the wrong three attributes to summarize. Hallucination — the model fills gaps by inventing plausible-sounding detail. [Read: Why AI Hallucinates About Your Company]
Where opinion enters
There is no opinion. There is a citation graph plus a ranking function plus a compression step. The brand's job is to shape each of those: get into the training corpus through tier-1 earned media and structured owned content; get into the live retrieval through current authoritative coverage; get weighted up through domain authority and citation density; get compressed favorably by making the strongest attributes the most-cited ones.
Brands that tend to win in AI engines don't argue with the model. They engineer the inputs.
See also: Signals That Move AI Reputation · AI Reputation Glossary
No communications firm can guarantee specific outputs inside third-party AI systems. The discipline is shaping the inputs the engines retrieve from — not directing the engines themselves.
