How AI Assistants Decide Which Brands To Recommend

Assistants pull brand names from three different places: what a model absorbed during training, what it fetches live, and what its tools return. This page explains that mechanism layer as observable behaviour, and says where the evidence stops.

Use the manual snapshot method · Use the change-briefing template

Everything below describes behaviour that can be observed from outside a provider or that providers document. Nobody outside the providers can see the ranking function itself, and provider behaviour changes without notice, so check current provider documentation before acting on any detail.

In short

When someone asks an assistant for the best option in a category, the names in the reply come from some mix of three sources: recall from training, live retrieval of documents, and tool calls made during the answer. Which mix applies depends on the assistant, the surface, the phrasing, and the moment. This page separates those mechanisms, compares how the major assistants source and display citations, explains why AI Overviews behave differently from a conversational assistant, and explains why two people asking the same question get different brands. It also states plainly what cannot be known from outside a provider.

Who this is for

In-house SEO and organic growth leads who need to explain AI answers to a leadership team that still thinks in blue links
Founders and heads of marketing who were told the company was missing from an assistant's answer and want to know what that means
Content and PR leads deciding which publications and communities are worth pitching for AI-answer presence
Agency strategists who have to teach this mechanism layer to clients before selling any measurement work
Technical SEOs who want the retrieval model straight before touching crawler access or page structure

When to use this

Before you interpret any AI-visibility number, so you know what the number is a measurement of
When a colleague says an assistant recommends a competitor and you need to explain the range of possible causes
When you are choosing whether to treat AI Overviews and conversational assistants as one metric or two
When someone asks what your position is in an AI answer and you need a better question to replace it

Three places a brand name can come from

A brand name in an AI answer arrives by one of three routes, and the practical consequences differ sharply. The first is training recall: the model produces the name from patterns absorbed when it was trained, with no document fetched at answer time. The second is live retrieval: the system searches, fetches pages, and conditions the answer on what it read. The third is tool use, where the assistant calls a search index, a browsing tool, or a connected data source and works from what comes back.

You can often tell them apart by what the answer shows you. An answer with linked sources almost always involved retrieval or a tool call. An answer with no sources, confident specifics, and stale facts is more consistent with recall. That is an inference from observable output, not a confession from the system, and the honest framing is: this is what the surface behaviour suggests, observed on a stated date.

The distinction matters because the levers differ. Content you publish today can enter a retrieval path quickly. Getting into a model's trained recall is slower, indirect, and outside anyone's control.

Most real answers are a blend, and the blend is not disclosed to you at answer time.

How the major assistants differ in sourcing and citation display

ChatGPT, Perplexity, Gemini, Copilot and Claude do not behave as one system. They differ in when they retrieve at all, which index or search partner they draw from, how many sources they read, whether they show citations by default, and whether a shown citation is the document that actually shaped the sentence or a supporting link attached afterwards.

Because of that, a brand can be visible in one assistant and absent in another on the same day for the same question, without anything on the brand's site having changed. Treat each assistant as a separate measurement surface with its own baseline rather than averaging them into a single headline figure.

A statement of this kind is only useful with the assistant, the model version where one is exposed, and the observation date attached. Provider behaviour changes without announcement, so an undated comparison is worthless within months.

AI Overviews are a different surface, not a smaller assistant

A search-embedded summary and a conversational assistant are different products with different triggers. One appears above results for a query the search engine chose to summarise; the other responds inside a session that may carry earlier turns, a stated persona, and follow-up questions. The intent that produces each is different, and so is the shape of the answer.

Collapsing them into one number is a common measurement error. Being summarised on a search results page and being named as a recommendation inside a chat are separate outcomes, and a change in one does not imply a change in the other.

Why two people get different answers to the same question

Assistants are non-deterministic by design: the same prompt can produce different text on consecutive runs. On top of that sit personalization, prior conversation turns, account settings, region, language, device surface, and whichever model version is being served to that user at that moment.

So when a colleague sees your brand and you do not, neither of you is wrong and neither observation is evidence. A single answer is an anecdote. The way out is not a better screenshot; it is sampling — repeated runs across a defined prompt set, with the spread reported alongside the average.

This is the hinge between the fundamentals and the measurement work. If you accept that variation is intrinsic, every claim about AI visibility becomes a statistical claim, and it has to be built and reported like one.

One answer is not a measurement. That sentence does more work on this site than any other.

There is no position, and that is not a technicality

Search trained a generation of marketers to ask for a position. An assistant does not publish an ordered list of eligible brands with your slot in it. It produces prose in which some brands appear, some are described favourably, some are cited with a link, and some are named only to be dismissed.

The replacement questions are countable: how often are you named across repeated runs, how often is your own site cited as a source, what share of named brands are you, and how are you framed when you do appear. Those are the metrics the measurement pillar defines, with formulas and failure modes.

What nobody outside the providers can tell you

No one outside OpenAI, Google, Microsoft, Anthropic or Perplexity can see the ranking function. There is no published weighting that says a review site counts for this much and a documentation page for that much. Anyone who tells you otherwise is describing a guess, a vendor's model of the world, or an inference from output presented as an internal fact.

What can be established is behaviour: run a defined prompt set enough times against named model versions on stated dates, record what came back, and report the spread. That is an observation about output, not a claim about internals, and it is the only kind of claim this site makes.

Where RankRift publishes a figure, it ships with the prompt set, the model versions, the run dates, the sample size, and the variance, so that a sceptic can re-run it and disagree with evidence.

Where we do not know, the page says so rather than filling the gap with a confident sentence.

The fundamentals cluster, grouped by what you are trying to do

These posts sit under this hub. Start with the group that matches the question you actually arrived with.

Frequently asked questions

Does an assistant look at my website when it answers?

Sometimes, and you usually cannot tell from a single answer. If the reply carries linked sources, something was retrieved at answer time and your pages were eligible to be among them. If it carries none, the answer may have come from training recall, in which case nothing on your site was read. This is why the readiness work and the content work are different jobs with different clocks, and why a behavioural observation needs a date rather than being stated as a permanent rule.

Why does the same assistant name us one day and not the next?

Because these systems are non-deterministic and the surrounding context changes. Temperature-driven variation, a different served model version, a different retrieval result set, a different region or session history can each change which brands appear. Day-to-day movement in a single answer is expected and is not a signal. The only way to tell a real change from noise is to run a fixed prompt set repeatedly and compare distributions across dates, which is what the measurement pillar sets out.

Is an AI Overview the same as ChatGPT recommending us?

No. They are different surfaces with different triggers, different sourcing behaviour, and different reader intent. A search-embedded summary appears against a query the engine chose to summarise; a conversational answer arrives inside a session that may carry earlier turns and preferences. Measure and report them separately with separate prompt sets. Combining them into one score hides which surface moved and makes any later diagnosis much harder.

How do you know any of this if the providers do not publish their ranking logic?

Nobody outside those companies can see the internals, and this page does not claim to. Mechanism statements here describe observable output or what providers document; RankRift does not publish measured runs of its own. Where a plausible explanation exists but no evidence supports it, the page says the evidence is missing. That distinction between documented behaviour and inference is the whole point of publishing the method.

Which assistant should we focus on first?

That depends on where your buyers actually are, and you can measure it rather than guess. Build a prompt set that reflects real buying questions, run it across the assistants your audience plausibly uses, and look at where you are absent relative to the competitors you care about. Focusing on the assistant with the most press coverage is a common and expensive error; focusing on the one your category's buyers use produces a defensible starting point.

Do I need to optimise separately for each assistant?

Mostly no, at the content level. The work that plausibly helps across assistants is the same: being clearly identifiable as a specific kind of company, being present in the sources assistants draw from, and being retrievable and quotable. Where the assistants genuinely differ is in sourcing and citation display, which affects measurement more than production. Optimise once, measure per assistant, and let the measurement tell you where a gap is assistant-specific.

Is AI search actually sending anyone to websites yet?

Some referral traffic is observable and some assistant sessions never produce a click at all, which is exactly why we refuse to quote a share figure we cannot reproduce. Numbers circulating in this category usually come from vendor marketing with no published method. What we can do is show you how to detect assistant referrals in your own analytics and logs, and treat your own data as the evidence rather than a market-wide estimate.

How current is this page?

Provider behaviour changes without notice, so treat any behavioural detail here as a hypothesis to re-check against the provider's current documentation before you act on it. The page describes mechanisms rather than dated observations. A claim found to be outdated should be corrected in place with a dated note rather than quietly edited.

See where you actually stand

The manual snapshot method shows you how to run a sampled prompt set across assistants yourself and record the sample size, models, dates, and confidence range beside the result. It is a starting point, not a verdict.

Use the manual snapshot · Read the methodology first