The honest starting point
Any article that tells you the ranking factors for an AI assistant is guessing. These systems do not publish a list of signals, there is no documented weighting, and the behaviour changes as models and retrieval systems are updated. We are describing a mechanism, not a formula, and we would rather say so than sell certainty.
What makes the question answerable at all is that the architecture is broadly understood, and that observed behaviour is consistent enough to reason about. You can work with tendencies without pretending they are rules.
We also use one assistant’s name here because it is the one people ask about. The mechanism described applies broadly to systems that generate an answer rather than return links, and the platforms differ in how heavily they retrieve live sources. AI Says is not affiliated with any of them.
Step one: the question is interpreted, not matched
A traditional search engine can succeed by matching a query to documents. An assistant first works out what is being asked: what kind of business would solve this, what constraints are implied, what the person probably means by "best" in this context, and whether location matters.
This is why phrasing changes the answer so much. "Cheapest option near me", "someone I can trust with a heritage building" and "a firm that handles commercial work at scale" are aimed at the same market but set up three different selection problems. A business whose material addresses one of those framings precisely will fare better on that framing than a business whose material addresses the category in general.
Step two: two sources of knowledge, with different behaviour
The assistant then draws on what it knows, and frequently on what it can fetch. The distinction matters because the two layers respond to your work at completely different speeds — the same split described in our explainer on generative engine optimisation.
- Internal knowledge from training. Confident, wide-ranging, and potentially out of date. It may still describe a service you stopped offering or a location you left.
- Retrieved sources at the moment of asking. Often a search of the live web, sometimes specific pages or connected data. More current, more specific, and the part your recent changes can reach.
Step three: what appears to earn a place in the shortlist
From observed behaviour across many answers, some tendencies recur. Treat these as informed observation, not documentation.
- Agreement between independent sources. A claim that appears only on your own site is weaker than one several unrelated sources support.
- Specificity of fit. An assistant building a shortlist needs a reason to name you; a documented specialism matching the question gives it one.
- Clarity of the entity. If it cannot tell confidently which business you are — because of a name collision, inconsistent details or thin information — it is likelier to name someone it is sure about.
- Something quotable. Self-contained, plainly written passages that answer the question directly are easier to build an answer from than marketing prose.
- Recency signals. Evidence that the business is currently operating and the information is current.
- Location cues, for anything local. A clearly stated service area, consistent with third-party listings.
- Nothing that contradicts the rest of the picture. Conflicting information is a reason to be cautious about naming you.
Why the same question gives different answers
Generative models are probabilistic: they sample from a distribution of plausible continuations, so two runs of the same prompt can differ in wording and sometimes in which businesses are named. This is a property of the technology, not a malfunction, and it cannot be eliminated by anything you do to your website.
On top of that sit several other sources of variation: which platform and which model version is answering, the user’s location, what the platform remembers from earlier conversations, whether the system chose to retrieve live sources for that particular question, and the exact wording used. Two people asking what feels like the same thing can reasonably get different shortlists.
The practical consequence is that a single screenshot proves almost nothing — in either direction. Appearing once is not a win; being absent once is not a verdict. Our article on measuring AI visibility explains how to get a usable signal out of a system that behaves this way.
It has an organisational consequence too. If a director runs one prompt, sees a competitor named and asks why, the only defensible reply is that one run is an anecdote. Agreeing in advance that decisions will be based on sampled rates rather than individual answers saves a great deal of unproductive argument.
What this means for what you should actually do
If selection is driven by interpretation, corroboration and clarity, the work follows: be unambiguous about what you do, be consistent about it everywhere, give the system plainly written material on the specific questions you want to be the answer to, and make sure something other than your own website supports your claims.
That is a slower programme than a tactic, and it is the only version of this advice we would stand behind. The step-by-step form is in our guide to appearing in AI search, and the sequencing — what to fix first when you cannot do everything — is in how to improve your AI visibility.
Claims to treat with suspicion
Anyone offering guaranteed inclusion in AI answers, a published list of ranking factors for a named assistant, a way to submit your business into training data, or a fixed timeline for a change to appear is describing something they cannot control. The same applies to quoted percentages about how often AI recommends particular businesses unless you are shown the prompt set, the platforms, the dates and the sampling method.
AI Says is not affiliated with or endorsed by any AI platform; platform names are used descriptively to identify the services being assessed.
Last reviewed 4 August 2026. AI Says is not affiliated with or endorsed by OpenAI, Google, Anthropic, Perplexity or other AI and search providers; platform names are used descriptively.