Why AI Recommends Some Brands and Not Others

Quick answer: AI recommends some brands and not others because two different processes decide the answer: recall, where the model repeats a pre-formed association from its training data, and retrieval, where it runs a live search and cites what it finds. A brand can be strong in one and invisible in the other. Optimizing only the pages a model might cite ignores the training-data half of the problem entirely.

Ask ChatGPT to recommend a brand in your category and you'll either get named or you won't, and the reason often has nothing to do with anything on your website today. AI systems produce a brand recommendation through one of two different processes: recall, where the model repeats an association it already formed during training, or retrieval, where it runs a live search and cites whatever it finds. Most advice on this topic treats it as one problem: get cited, get mentioned, build authority. That advice isn't wrong, but it's aimed at only half the mechanism. A brand can be well-optimized for retrieval and still get skipped because the model's recall of the category doesn't include it, or vice versa. This article breaks down how each mechanism actually works, what the evidence says about which one is doing the deciding, and how to check which one is shaping the answers your own brand is getting.

Recall vs. retrieval: the two ways AI decides what to say about a brand

AI systems decide what to say about a brand through two distinct mechanisms, and most GEO advice quietly assumes there's only one. Recall is the model repeating a pre-formed association it picked up during training, with no search involved. It's answering from memory, the way a person might name the first brand that comes to mind. Retrieval is the model running a live search (often a query fan-out, where one question expands into several sub-queries) and citing whatever it finds at answer time. The two produce different failure modes: a brand invisible to recall gets skipped in a fast, ungrounded answer even if its website is excellent, and a brand invisible to retrieval gets skipped in a search-grounded answer even if the model's training-era impression of it is fine. Treating these as one problem, "we need to get cited more," fixes only the retrieval half. That's why brands that publish more content sometimes see no change in how AI describes them.

How retrieval works: when AI is actually searching before it answers

When an AI system retrieves rather than recalls, it's doing an SEO-adjacent job: searching an index and picking what to cite. Ahrefs (the SEO data and analytics platform), which analyzed 1.4 million real ChatGPT prompts, found that roughly 88% of the URLs ChatGPT ends up citing come from its search channel, meaning retrieval-mode answers run through largely the same technical groundwork (crawlability, indexation, clear on-page answers) that governs classic Google visibility, not something baked into training data. But retrieval and citation aren't the same event. The same Ahrefs study found only 49.98% of the URLs ChatGPT actually retrieves are ever cited in the final answer. Being looked at by the model is not the same claim as being named by it, and conflating the two overstates how much any single page-level fix will move the needle.

Grounded products make this mechanism visible on purpose. Google's AI Mode and other retrieval-augmented systems search live before every answer, by design. That's a different operating mode from a free-tier chat assistant, which sometimes skips the live search step to save on inference cost and answers from its training-era impression instead. The practical result: the same brand, asked the same question, can get a retrieved answer in one product surface and a recalled one in another.

How recall works: when AI just repeats what it already 'knows'

Recall draws on a training-data corpus set at a fixed point in the past, and a brand can't edit that corpus after the fact the way it can publish a new page. What shapes that corpus is also far wider than what shows up as a citation. In the Ahrefs dataset, Reddit is cited just 1.93% of the time, yet it supplies 67.8% of all the URLs the model retrieves but never cites, a gap Search Engine Journal's write-up of the same data confirms independently. The model is reading Reddit constantly to understand a topic and form an impression, but it almost never credits Reddit as the source. That's recall being fed by material a brand can't see in any citation report.

This is also why recall can simply be wrong or stale. Whatever a model currently associates with a category was set during a training run that predates most brands' deliberate AI-visibility effort. An improved site today does not retroactively update what the model already learned. A brand's recall problem and its retrieval problem are separate diagnoses, and a citation audit alone only catches one of them.

Why citation-chasing optimizes the wrong variable

AI brand visibility is a probabilistic outcome, not a fixed position a brand can hold the way a #1 Google ranking is a position. The clearest evidence for this comes from outside the citation-tracking industry entirely. Rand Fishkin (Moz and SparkToro co-founder) and Patrick O'Donnell of Gumshoe.ai ran 2,961 real prompts across 600 volunteers on three major AI platforms and found less than a 1-in-100 chance of seeing the same brand recommendation list twice for the same category prompt. Fishkin's own conclusion, as reported by Search Engine Land: any tool claiming to show a stable "ranking position in AI" is measuring something that doesn't behave like a ranking.

That volatility is a symptom, not the disease. The underlying cause is that a model's answer is a sample drawn from everything it has recalled and retrieved about a category, and when the signal is thin or inconsistent across a brand's footprint, the sample varies more, not less. Practitioners tracking this directly describe it as "brand mentions matter more than backlinks": the discipline shifts from link-building on one's own site to consistency across every third-party place a brand gets described, because that's the material both recall and retrieval draw from. A brand that's described five different ways across the web is harder for a model to summarize cleanly, and a model that can't summarize a brand cleanly reaches for a competitor it can.

We saw this pattern directly in a recent audit. A B2B manufacturer client sold several unrelated product lines under one brand and one site, and got zero AI citations across every major engine for its own core category query, with some engines unable to confirm the company existed at all. A single-category competitor, holding a comparable certification set and no clear technical edge, won the recommendation outright. The difference wasn't product quality: the winning competitor's entire site was one thing, that same profile was mirrored consistently across half a dozen B2B marketplace listings, and it paired the whole footprint with a couple of specific, quotable claims (a "years in business" figure, a "first to use this process" line) that models could repeat as justification. A model faced with a single clean entity and a diluted one will recall and cite the clean one, independent of which product is actually better.

How to check which mode is deciding what AI says about your brand

A brand can test which mechanism is producing a given answer with a fairly simple, repeatable check: run the identical prompt in a free-tier chat mode and again in a grounded, search-enabled mode (ChatGPT Search, Google AI Mode, or similar), and compare the two answers. A materially different answer means the model filled gaps live rather than repeating a fixed recall, which tells you the retrieval-side fix (crawlability, clear on-page claims, indexation) is the one worth prioritizing. A near-identical, stale, or simply wrong answer in both modes points to a recall problem, which no amount of new page content fixes on its own. It needs the wider third-party footprint (reviews, comparison content, community discussion, press) to shift what gets learned about the brand in the next training-relevant window.

Doing this systematically, across engines and across a real prompt set rather than one lucky or unlucky query, is exactly the diagnostic work behind an AI Visibility Audit. Uygen's framing splits the problem into Access, Understanding, and Authority: checking whether the model can retrieve a brand's pages at all (Access), whether it describes the brand accurately when it does (Understanding), and whether the wider evidence ecosystem around the brand is strong enough for the model to trust and recall it (Authority). Recall and retrieval map cleanly onto that split; retrieval problems are mostly Access and Understanding, recall problems are mostly Authority. If you've already run a Share of Model check and the number is low, this recall/retrieval split is the next question to answer before deciding what to fix. For the ongoing version of this check, see how to track AI search visibility as a repeatable measurement system rather than a one-off test.

FAQ

Why does ChatGPT recommend competitors instead of us even though we outrank them on Google?

Because Google ranking and AI recommendation are gated by overlapping but different mechanisms. Outranking a competitor on Google mostly reflects retrieval-side signals: crawlability, indexation, on-page relevance. If a competitor is winning the recommendation anyway, the model's recall of the category (shaped by reviews, comparison content, community discussion, and press coverage it learned from during training) may simply favor them independent of current search rankings. Ranking well doesn't overwrite what a model already learned to associate with the category.

Can a bad AI brand recall be fixed without waiting for a new model training run?

Not directly and not immediately, but the fix isn't purely passive either. You can't edit what a specific past training run already encoded. What you can do is close the gap for the next relevant window: build the consistent, corroborated third-party footprint (accurate reviews, comparison mentions, clear category language on your own site, forum and community presence) that recall draws on, and use grounded, search-enabled answers in the meantime, since those retrieve live rather than relying on the stale association.

Is AI brand recommendation the same thing as ranking in AI Overviews?

No. Ranking in a Google AI Overview is a retrieval-mode outcome: Google's system searched, selected sources, and cited one of them. A brand can be absent from AI Overviews entirely and still get recommended by name in a ChatGPT answer that never searched at all, because that answer came from recall. They're related but separately measurable outcomes, and a brand can be strong in one and weak in the other.

How is this different from a standard AI citation audit?

A standard citation audit typically measures retrieval-mode outcomes: which prompts get your brand cited, and by which sources. That's real signal, but it's only half the picture, since it can't see what a model recalls when it doesn't search at all. A fuller diagnostic tests both modes explicitly, the same prompt in a free-tier (recall-leaning) and a grounded (retrieval-leaning) setting, and treats a gap between them as its own finding, not noise.

AI recommends some brands and not others because two different processes decide the answer, and most advice on the topic optimizes for only one of them. Retrieval is checkable and largely SEO-adjacent: roughly 88% of what ChatGPT ends up citing comes straight from its search channel, so the fundamentals that get a brand ranked also get it retrieved. Recall is the harder, less visible half: shaped by a training-era corpus a brand can't directly edit, fed heavily by sources like Reddit that almost never show up as a citation, and volatile enough that Fishkin and O'Donnell's study found under a 1-in-100 chance of the same brand list appearing twice. Chasing citations alone treats a two-variable problem as a one-variable problem. The fix starts with finding out which mechanism is actually producing the answers about your brand today.

Next step: Run the free-tier-vs-grounded prompt test from this article on your own brand. If the two answers disagree, or if both are stale or wrong, that's a recall problem no new blog post will fix on its own, and it's exactly what an AI Visibility Audit is built to diagnose across Access, Understanding, and Authority before you spend budget on the wrong half of the fix.