How an AI answer is assembled from public evidence

A buyer who once typed a few keywords into a search box now asks a full question. Which analytics platforms should a UK fintech shortlist? Which vendors integrate with our data warehouse? What are the alternatives to the market leader for a team of our size? The answer arrives as a few confident paragraphs, often with a short list of names and a handful of sources underneath.

For the brands in that category, the answer matters more than its length suggests. It can shape a shortlist before anyone visits a website. Yet most brands have never read the answers given about them, and fewer still know how those answers were put together. This piece walks through that assembly in plain terms: what goes in, what comes out, and where a brand is won or lost along the way.

It starts with the question, not the keyword

Everything that follows depends on the words the buyer uses. A question carries context that a keyword never did: the buyer's sector, their size, their constraints and the job they need done. "Best CRM" and "Which CRM suits a regulated UK lender with a small sales team?" are different questions, and they produce different answers with different names in them.

That is why the starting point for understanding your position is the set of buying questions that shape a shortlist in your category, written in the words a buyer would actually use. If you only ever look at the questions you would like to be asked, you learn very little about the answers buyers are actually given.

What the assistant does before it answers

In broad terms, an assistant draws on two kinds of knowledge. The first is what the underlying model absorbed when it was trained: a compressed impression of a very large body of public text, fixed at a point in time. The second, for many assistants and many questions, is fresh material gathered for the question itself, by searching the web and reading a selection of what comes back.

When an assistant searches before it answers, it does not read everything. It turns the buyer's question into one or more searches, takes a selection of the pages those searches return, and reads them as evidence. The answer is then written from that evidence and from what the model already holds. Pages that were read and used may be listed as sources. Pages that were never found play no part at all.

Very little of this is visible from the answer alone, and the details differ between providers and change over time. Nobody outside a provider knows exactly how it weighs one source against another, and we do not claim to. What can be observed is the outcome: what the answer said, which sources it cited and which brands it named.

Where the evidence comes from

The raw material is public. It is your own website, your documentation, your product pages, your integration guides and the material you have published about the work you do. It is also what other people publish about you: comparison articles, analyst notes, directory listings, partner pages, reviews, community threads and press coverage.

Two things about this evidence are worth holding on to. First, it is patchy. A brand can have a strong product and a thin public record of it: an integration that exists but is described nowhere, a security standard met but never written down, a sector served well but never named in a buyer's words. Second, it moves. Pages are published, rewritten and removed every day, and an answer built from them shifts as they do.

An assistant can only work with what it can find and read. If the evidence for a claim does not exist in public, or exists only in a sales deck, the quality of the product will not carry it into the answer on its own.

Three outcomes, never one

When a brand appears in an answer, it can appear in three quite different ways, and they should never be blended into a single score.

A mention

Your name appears in the answer. It might sit in a list of options, in a passing comparison or in a sentence about the category. A mention tells you the assistant connects you with the question. On its own, it rarely wins a shortlist.

A citation

Your own evidence is used as a source. The assistant read a page you published, or a page about you, and drew on it to write the answer. A citation means the answer is being built, at least in part, on what has been said about you in public.

A qualified recommendation

You are suggested for the buyer's stated need: not simply named, but put forward as a fit for their sector, size or problem, ideally with the reason attached. This is the outcome that moves a shortlist, and the hardest one to earn.

The three come apart all the time. A brand can be mentioned without being cited, when the assistant names it from general knowledge but builds the answer on someone else's evidence. It can be cited without being recommended, when its documentation is used to explain the category while another vendor is put forward for the need. Treating any one of these as proof of the others is how teams end up celebrating the wrong result.

Why the same question gives different answers

Ask the same question twice and you may get two different answers. Ask it of two assistants and the difference can be larger still. Answers vary between providers, between the models a provider runs, from one week to the next, and with small changes in the wording of the question.

This is not a fault waiting to be fixed; it is how these systems behave. It means a single observation can tell you very little. One answer that names you is not evidence that you are recommended, and one answer that leaves you out is not evidence that you are invisible. What carries weight is a pattern: the same agreed questions, observed repeatedly, with the provider, the date and the sources recorded each time, so that a real change can be told apart from ordinary variation.

What this means for a brand

Put the pieces together and a plain picture emerges. The answer a buyer reads is assembled from public evidence, selected by a system nobody outside can fully inspect, and shaped by the exact question asked. A brand ends up in that answer when there is clear, specific and findable evidence connecting it to the buyer's need. It ends up missing when that evidence is thin, scattered or written in language no buyer uses.

That has three consequences. The first is that the starting point is reading the answers, not guessing at them. The second is that the useful question is never only whether you appear, but how: mentioned, cited or recommended, and on whose evidence. The third is that the lever a brand actually holds is its public evidence, because that is the raw material every answer is built from.

None of this promises a result. We do not control what a model recommends, and no one outside the provider does. What a brand can do is make the evidence for its real strengths easy to find, easy to verify and easy to recommend, then measure what changes on the same questions over time, and say plainly when nothing does.