How AI decides which brands to recommend

Ask ChatGPT "what's the best project management software?" and you get four or five names. Ask again tomorrow and you might get a slightly different four or five. Somewhere in that process, a decision is being made about which companies exist and which don't. For most businesses, nobody has ever looked at how it works.

It is not a ranking algorithm in the search-engine sense. There is no index position to climb, no single score being sorted. But it isn't random either, and the parts that aren't random are the parts you can work on.

Two different mechanisms, often confused

What produces the answer depends on whether the model is retrieving or recalling.

Retrieval is when the engine searches the live web mid-answer, reads what it finds, and writes a response grounded in those pages. Perplexity does this on nearly every query. ChatGPT does it when the question looks current or specific. You can usually tell because the answer carries citations.

Recall is when the model answers from what it absorbed during training: no live lookup, no sources. Ask a general question about a well-established category and you often get this: a fluent list of names with no links attached.

The distinction matters because the two respond to completely different work. Retrieval responds to what's on the web right now. Recall responds to what was on the web, repeatedly and consistently, over the years the model was trained. You can influence retrieval in weeks. Recall moves on a much longer horizon, if it moves at all.

Most practical work targets retrieval. That's the honest version of the pitch: you are not reprogramming a model's memory, you are changing what it finds when it goes looking.

What retrieval actually reads

Here is the part that surprises people. When an engine retrieves for a "best X" question, it rarely leads with vendor homepages.

Think about what the question is asking. "What's the best project management software" is a request for a comparison across vendors. A vendor's own site is, by construction, not a comparison; it's one company's account of itself. Models are trained on enough of the web to have absorbed that pattern, and retrieval reflects it: the pages that get pulled are the ones that compare.

In practice that means listicles and category roundups, review platforms, community threads where people describe what they actually chose and why, and independent analyses. Your own site tends to appear as corroboration once you've already been named: confirming what you do, not establishing that you're worth naming.

This is the single biggest structural difference from traditional SEO, and it's uncomfortable. In classic SEO, the asset you're optimising is one you own. Here, most of the surface that decides the answer belongs to somebody else.

The four things that appear to matter

From observing how answers change, four factors do most of the work.

Presence in the retrieved set. If none of the pages an engine pulls for a query mention you, you cannot appear in the answer. Nothing else on this list matters until this is true. This is why "get cited on the sources AI reads" is the first move on almost every fix list. It isn't a growth tactic; it's the entry condition.

Consistency across sources. A brand described the same way in six places is easier for a model to state confidently than one described six different ways. Categories, positioning, and even naming drift over the years; that drift is expensive here in a way it never was in search.

Clarity of category. Models need to know what bucket you're in to include you in a bucket question. "The operating system for modern work" tells a model nothing about which questions you're an answer to. Somewhere (your own site, your listings, your descriptions) the plain-language category needs to be stated plainly.

Structural legibility. Structured data, clean headings, and pages that answer a question directly rather than burying it under narrative all make a page easier to extract from. This is the one factor that lives entirely on your own site, which is why it's usually the fastest thing to fix.

What doesn't work

Two things get sold that don't survive contact with reality.

Keyword stuffing for models. Repeating "best CRM for startups" across your site does not make a model say it. Retrieval pulls pages that answer the question, and the pages that answer a comparison question are comparisons. Your homepage is not one.

Manufactured mentions. Spun posts, bought reviews, fabricated community threads. Beyond being the kind of thing that gets a domain distrusted, the sources most likely to be retrieved are the ones with editorial or community filtering, which are precisely the ones hardest to fake. The tactic tends to fail on its own terms before it becomes an ethics question.

Why the answer changes between runs

Ask the same question twice and you'll often get a different set. This is normal, and it's worth understanding before you read anything into a single result.

Language models sample. Retrieval returns slightly different pages depending on timing and phrasing. The same query on the same day can produce genuinely different answers. A single check tells you very little; what tells you something is the same question asked repeatedly, across engines, tracked over time.

This is why any number attached to AI visibility should come with a range rather than a single figure. It's also why a tool that shows you one confident number, once, is hiding the most important property of the thing it's measuring.

Where to start

If you want a concrete order of operations:

1. Find out whether you appear at all, across several engines, with the questions your buyers actually ask.
2. If you don't, look at what the engines did cite. Those sources are your target list.
3. Make sure your own site states your category in plain language and carries structured data.
4. Re-check on a schedule rather than once. You're looking for a trend, not a snapshot.

None of this is fast. Citations take time to earn and longer to propagate into what engines retrieve. But it is tractable, and it's considerably more tractable than the alternative of hoping the model already knows who you are.