2.How AI answers actually form.

To change what AI says about you, it helps to know where the words come from. The short version: an assistant answering a buying question is drawing on two supplies of information, and blending them into one confident voice.

Two supplies: memory and reading

The first supply is training knowledge: what the model absorbed about the world, and about your market, when it was built. It is broad, it is opinionated, and it ages. A model trained last year describes last year’s version of you.

The second supply is live retrieval: many assistants now search the web while answering, read a handful of pages, and cite them as sources. This supply is fresh but narrow. The engine reads a few pages, not the whole internet, and which pages make that reading list decides the answer. The bots that do this reading are AI crawlers, and they visit your site more often than you think.

Different engines mix the two supplies differently, which is the first reason the same question gets different answers from different assistants.

Same question, eight answers

People imagine “what AI says” as one thing. It is at least eight things. ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek, Copilot, and Google’s AI Overviews each have their own training, their own retrieval habits, and their own taste. They disagree far more than most people expect. In the CiteHawk AI Index’s July 2026 edition, the eight engines could not agree unanimously on the number-one brand in roughly 40 percent of categories measured. This engine disagreementis normal, permanent, and decisive: a brand can be ChatGPT’s favorite and absent from Gemini entirely.

The practical consequence: measuring one engine tells you almost nothing about the other seven. Rankings built from one or two engines look complete and stable while being wrong in ways the tracker cannot see. We call this the coverage illusion, and it is the reason serious measurement covers every major engine rather than sampling one.

Same question, same engine, different day

Ask one assistant the same question twice and you will often get two slightly different answers. The names shuffle, one brand drops off, another appears. This is not the market moving; it is the model sampling. A single ask is one draw from a distribution, which is why screenshots of one good answer prove very little, in either direction.

Real measurement handles this the way pollsters do: repeat the question, across engines, on a steady cadence, and watch the rates rather than the individual answers. When a rate moves and stays moved, something real happened.

Location changes the answer too

The same question asked from Sydney, London, and Chicago can produce three different shortlists. Engines localize aggressively, and regional divergence is large enough that a brand dominant at home can be invisible one market over. If you sell in more than one country, each market is its own scoreboard.