How AI assistants decide which brands to name

This article tries to answer a question that a lot of people will answer confidently and few can answer honestly: when someone asks an assistant to recommend a business in a category, how does it choose the names?

We will separate what is observable from what is inference, and say clearly where the answer is that we do not know.

There are two different mechanisms, and confusing them causes most of the bad advice

An assistant can produce a brand name from two very different places. It can recall it from what the model absorbed during training, or it can retrieve it live from the web at the moment you ask.

These behave nothing alike. Trained-in knowledge is frozen at a cutoff, favours whatever was widely written about before that date, and cannot be influenced on any timescale you can plan around. Retrieved knowledge reflects what is on the web now, can incorporate a page published last week, and is heavily shaped by which sources the retrieval step happens to surface.

Almost all the advice worth following applies to the retrieval path. When someone tells you they got their client cited in ChatGPT within a month, they are describing retrieval. That is not a criticism — retrieval is the path that matters commercially, because it is the one that responds to work.

What we can observe about retrieval

Running thousands of scheduled prompts across four assistants gives you a lot of pattern and not much certainty. A few things show up consistently enough to be worth acting on.

Specificity of the source page beats authority of the domain

For narrow questions, a small site with a page addressing exactly that question is cited more often than a large site with a page covering the topic among ten others. This is different from classic search, where domain strength does a lot of work. It is also the most encouraging finding for smaller businesses.

Aggregators are cited more than vendors

For 'best X in Y' style questions, directory pages, review platforms and roundup articles appear as citations far more often than the websites of the businesses being named. The assistant is behaving sensibly: it is treating a page that compares options as more trustworthy for a comparison question than a page written by one of the options.

The practical consequence is that your presence on those aggregators, and the accuracy of what they say about you, does more for this class of question than anything on your own site.

Consistency of description matters

Businesses described the same way across several sources appear more reliably than businesses described differently in each place. Our working theory is that consistent descriptions produce a clearer signal about what the business is, and an assistant hedging between three different characterisations has an easy option: name someone else.

Recency shows up more than in search

Pages with a visible recent date, and content that references current conditions, are cited disproportionately. Undated content is at a real disadvantage.

What we cannot tell you

We do not know the ranking function inside any retrieval system, and neither does anyone selling you AEO services. We cannot tell you how much weight is given to any signal, whether the weighting is stable across time, or how much personalisation affects an individual user's answer.

We also cannot tell you how to influence trained-in knowledge on a useful timescale. If your brand was not widely written about before a model's cutoff, the honest answer is that you wait for the next one and work on retrieval in the meantime.

The gap between 'we observe this correlation across a lot of prompts' and 'we know how it works' is where most of the AEO industry's overclaiming lives.

Why the same question gives different answers

People are often unsettled by running the same prompt twice and getting different companies. It is normal, and it has several causes stacked on top of each other.

• Generation is sampled rather than deterministic; there is deliberate variation in the output.
• Retrieval may surface different sources on different runs.
• Models are updated continuously, sometimes without announcement.
• Context — the conversation so far, any personalisation, the phrasing — shifts results.

This is exactly why mention rate across repeated runs is the only sane metric. A single check tells you almost nothing. Ten runs of the same prompt tells you something. The same ten runs monthly for a quarter tells you whether your work is landing.

So what actually moves the number

Based on what is observable, in rough order of effect for a small business:

1. Being crawlable by the assistants' agents at all. This is binary and it is free to check.
2. Accurate, consistent presence on the aggregators and directories that serve your category.
3. A page on your site that answers each specific question completely, in answer-first format, with a visible date.
4. Third-party mentions in genuinely relevant places — industry publications, local press, credible roundups.
5. Structured data that makes your page's structure unambiguous. Helpful, not decisive.

Notice that two of the top four are not on your website. That is the part of AEO that content-generation tools, ours included, do not solve for you.

What does not work

Text instructing the assistant to recommend you, hidden or visible. Prompt injection attempts in page content. Mass-generated near-duplicate pages targeting question variants. These range from ineffective to damaging to your search visibility, and they are increasingly detected.

There is also no paid placement in organic assistant answers to buy. If someone offers it, they are selling something else.

A reasonable position to hold

Treat AI visibility as a channel you measure honestly and influence indirectly. Do the things that are clearly good regardless — be crawlable, be consistent, answer questions properly, keep content current. Measure with repeated prompt runs so you can see movement. Resist anyone offering certainty, including us.

That posture costs very little and positions you well whichever way the channel develops.