PondralWhen an AI assistant recommends a brand, it feels like a verdict, as if the model knows, from some vast store of judgement, that this is the best option. That mental picture is wrong. The gap between how people think assistants choose and how they actually choose is exactly where most brands lose visibility they could have won.
This post is the mechanics. No tactics, no checklist. Just a clear walk through what happens when ChatGPT, Claude, Gemini, or Perplexity decides which names make it into an answer, and which get left out. Once you can see the machine, the moves become obvious.
First, the distinction that explains everything: retrieval vs generation
There are two very different ways an AI assistant can produce an answer, and knowing which one you're up against changes everything.
Generation from memory.A language model is trained on an enormous amount of text. Some facts get baked into its parameters during training: call this its memory. Ask it something stable and well-established and it can answer straight from that memory, with no live lookup. The catch: this memory is frozen at training time, it's fuzzy, and it skews toward whatever was abundant and repeated in the training data. Big, long-established brands tend to live in here. New entrants, niche players, anything local, and anything that changed recently mostly do not.
Retrieval, then generation.For anything current, specific, local, or fast-moving, modern assistants don't rely on memory. They go and fetch, searching the live web, pulling back a set of pages, and then writing an answer grounded in what they just read. This is often called retrieval-augmented generation, and it is the mode that matters most for AI Visibility, because it's the mode triggered by exactly the questions that drive new customers: “best X for Y,” “compare the top options,” “who should I use for Z.”
Here's why the distinction is the whole game. If an answer comes purely from memory, you mostly can't influence it in the short term: it reflects years of accumulated training text. But if the answer comes from retrieval, then it's being built, right now, from pages the assistant can find and read. Pages you can write. Pages you can structure. That's the influenceable surface, and it's bigger than most people assume, because assistants reach for retrieval precisely when the question is the kind a buyer asks.
So the first thing to internalise: for the questions that matter most, you're not fighting the model's memory. You're competing to be one of the pages it retrieves and trusts.
The pipeline, step by step
When an assistant answers a retrieval-style question, it runs something close to this sequence. The exact engineering varies by product, but the shape is consistent.
Step 1: It decomposes the question. “Best [category] for [use case]” isn't searched literally. The assistant breaks it into sub-questions: what makes an option good here, who the candidates are, what each offers, what others say. This already tells you something. It's hunting for pages that answer specific sub-questions, not pages that just repeat the headline phrase.
Step 2: It retrieves candidates.It runs searches and pulls back a shortlist of pages. This stage leans on familiar search signals: relevance, authority, links. This is the one place where classic SEO and brand strength do real work. They help you make the shortlist. If you're not retrieved here, nothing downstream can save you. But making the shortlist is necessary, not sufficient.
Step 3: It reads and extracts. Now the assistant actually readseach candidate page and tries to pull clean, specific facts out of it. This is the quiet filter that eliminates most brands. A page with clear headings, direct answers, and machine-readable facts is easy to extract from. A page of promotional prose, adjectives, slogans, “world-class,” “trusted by thousands,” no specifics, gives the assistant very little it can lift and reuse. Assistants cannot quote what they cannot parse. Plenty of pages get retrieved in step 2 and then quietly dropped here.
Step 4: It scores the sources. Surviving candidates get weighed against each other on things like specificity, internal consistency, how directly they answer the sub-questions, and how trustworthy the source looks. A big domain helps a little here, but a small site with a precise, well-structured, genuinely useful page can, and often does, outscore it.
Step 5: It writes the answer and chooses who to name and cite. Finally, the assistant composes the answer, naming the brands its best sources support and, where the product shows citations, linking back to those sources. The names that make it in are the ones attached to the clearest, most useful, most extractable pages from the steps above.
Map your brand against those five steps and you can usually locate your own failure point precisely. Most strong brands sail through step 2 and stall at step 3: retrieved, then unread, because the page gave the assistant nothing specific to work with.
Why citations matter more than they look
It's tempting to treat the source links under an AI answer as a footnote. They're not. They're three things at once.
They're the audit trail.A citation is the assistant showing its work: “here's where this came from.” For a retrieval-grounded answer, the cited pages arethe evidence the answer rests on. If your page is one of them, you didn't just get mentioned. You helped build the answer.
They're the traffic. A description of your brand with no link is a compliment. A description witha link to your domain is a visitor. Citations are the bridge from “the assistant talked about us” to “the assistant sent someone to us.” If a review aggregator or a competitor's comparison page gets the citation instead of you, they capture the click off the back of your category.
They're a trust signal that compounds.Pages that get cited are, by definition, pages the assistant judged worth standing behind. That tends to reinforce itself: clear, citable, frequently-referenced pages keep getting reached for. Being citable isn't a one-time win. It's a position that builds.
This is why, in any serious measure of AI Visibility, “were you linked” is tracked as its own distinct factor, separate from whether you were merely named. Being talked about and being sourced are different outcomes, and only one of them sends you a customer.
Competitive share: visibility is relative, not absolute
Here's the last piece people miss. AI answers almost never name a single option. Ask for “the best X” and you typically get a short list: three names, five, sometimes more, often with brief reasons for each.
That means your visibility is never just about you. It's about your slice of a shared answer. If an assistant names five brands and you're one of them, you have roughly a fifth of that answer's attention, even if you were “present.” If a competitor is named first, described in detail, and cited with a link while you get a passing mention, you are both technically in the answer and losing it badly.
This is what competitive share captures: of all the brands named in the answers your buyers see, how much of that conversation is yours? It reframes the goal. The question isn't only “do I show up?” It's “am I leading this answer, or just standing in the back of it?” A brand can be present in every relevant answer and still be quietly dominated in all of them. You only see that when you measure share, not just appearance.
It also reframes the opportunity. Because the answer is shared and finite, share you take is share a competitor loses. In categories where nobody is paying attention to AI Visibility yet, which is still most of them, that share is sitting there unclaimed.
Putting it together
Strip away the mystique and the picture is clear:
For the questions that matter, assistants retrieve and read pages, then build an answer from them. They don't recite a verdict from memory.
You compete first to be retrieved (where brand and SEO help), then to be read and extracted (where page clarity and structure decide it), then to be named, positioned, and cited (where you either lead the answer or fade into it).
Citationsaren't decoration. They're the evidence trail, the traffic, and a compounding trust signal.
Visibility is relative: your share of a shared answer, not a solo verdict.
You don't have to out-spend a category to win here. You have to give the assistant the clearest, most specific, most citable page on the question, and to know, by measuring, where in this pipeline you're currently falling out.
That last part is the practical starting point. The pipeline is invisible until you instrument it. Once you can see whether you're getting retrieved, read, named, and cited, and how your share compares, the work to improve stops being guesswork.
Pondral measures and improves AI Visibility: how brands appear in AI-assistant answers across ChatGPT, Claude, Gemini, Perplexity, and Grok. You can run a free check at pondral.com/analyze to see exactly where you stand.