Recommendation is a two-stage process: the model retrieves candidate sources, then generates an answer weighted by how consistently those sources describe each brand. Consistency beats prominence at the second stage.
That split explains the result that confuses everyone. A famous company can be retrieved and still not recommended, and a small one can be recommended because four independent sources agree about what it does.
“How does the model decide?” is the wrong question, because the model is only doing half the work. Nothing is decided in one step, and the two steps optimise for different things. Understanding where they diverge is most of what separates useful AEO work from expensive guessing.
What happens in the first stage?
A retriever selects candidate documents. It is not reasoning about your company; it is scoring passages for relevance to the question and returning the top few.
This is the stage that fails silently. If your material is not retrievable when the question is asked, no amount of quality further down helps, because the generator only ever sees what the retriever handed it. The architecture is documented in Lewis et al., which pairs a model’s stored knowledge with a retrieval step over an external corpus precisely because a model’s ability to access what it memorised is unreliable.
Two properties of this stage matter to you.
Passages get scored, not pages. A section that answers its own heading can be retrieved on its own; one that depends on the three sections above it usually cannot.
The retriever has no opinion about your funding, your logo wall, or your category leadership. It is matching a question to text.
Matching has also moved. Classical retrieval scored on shared words; dense retrieval scores on learned meaning, and Karpukhin et al. measured that approach beating a strong term-matching baseline by 9 to 19 points on top-20 passage accuracy. A passage can now be retrieved without containing the question’s words at all, which is why writing about the concept beats repeating the phrase.
What happens in the second stage?
The generator writes an answer from the retrieved set. It has to decide which brands to name, and it is doing that from a pile of documents that disagree with each other.
This is where consistency starts beating prominence. When four independent sources describe a company the same way, that description is easy to state confidently. When six sources describe a better-known company six different ways, there is no confident sentence to write, and a generator asked for a short recommendation list will reach for the one it can characterise cleanly.
Brand A, well known
Appears in eight documents. Described as a project tool, a work OS, a CRM, a collaboration suite, and an “everything app”. All of it accurate at some point in its history.
Hard to characterise
Brand B, smaller
Appears in four documents. Every one calls it scheduling software for field service teams. Same category, same buyer, same words.
Easy to place
This is the mechanism behind the most common complaint in the category, which is that a well-funded company with excellent SEO gets left out of recommendation lists while a smaller competitor appears. It is rarely a conspiracy or a quality judgement. It is usually that one of them is describable and the other is not.
Why does consistency work this way?
Because agreement across independent sources is the only proxy for confidence a system like this has access to.
It has no way to verify your claims and no relationship with your marketing team. What it can observe is whether the corpus converges. The same logic runs through structured knowledge: Wikidata admits an item that can be “described using serious and publicly available references”, and the sameAs property on schema.org’s Organization type exists to state that a described entity is the same as one identified elsewhere. Both are formalisations of the same idea: identity is established by external agreement, not by assertion.
You cannot make an engine confident about you by saying the same thing louder. You make it confident by getting other people to say the same thing you do.
What does this mean for what you write?
Three consequences, in descending order of how uncomfortable they are.
The version on your homepage, your G2 profile, your LinkedIn and your press should be the same sentence. Rewriting it quarterly resets the agreement you were building.
“Scheduling for field service teams” gets recommended for that question. “The work platform for modern teams” is retrievable for everything and nameable for nothing.
A stale directory listing calling you something you stopped being two years ago is actively costing you, and it is not on a page you control.
Does prominence count for nothing?
It counts, and it counts at the wrong stage to save you.
Being well known helps enormously in retrieval. More has been written about you, so more documents mention you, so more of them surface for more questions. A famous brand is almost never invisible in the way a small one can be.
What prominence does not do is make you easy to characterise. If anything it works against you, because a company that has been around long enough to be famous has usually been several things, and the corpus remembers all of them. The documents that make you retrievable are the same documents that make you hard to summarise in one clause.
More coverage means more chances to be in the retrieved set for any given question. Being known is a genuine advantage here.
Ten inconsistent descriptions are harder to reduce to a recommendation than four matching ones. Volume of coverage is not agreement.
Does the same logic apply on Google?
The retrieval half differs and the conclusion does not.
Google’s AI features run on the Google index with ordinary Search eligibility, and the company states there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”. So stage one is your existing SEO. Stage two still has to produce a sentence about you, and it still finds that easier when the corpus agrees.
Which is the useful summary of the whole mechanism: stage one is a technical problem you can solve on your own domain, and stage two is a positioning problem you can only solve elsewhere.
How would you know which stage is failing?
Ask the engine to describe you, then ask it to recommend tools in your category, in that order and in the same clean session.
A vague or wrong description points at stage one, because the retriever is not reaching good material about you. A confident, accurate description followed by absence from the list points at stage two: it can find you, it can describe you, and it still would not put you forward. The five-minute test is exactly this comparison, and the full diagnostic extends it to the four causes underneath.
Both stages are worth working on. Only one of them is on your website.