How LLMs “Choose” Which Brands to Cite

Ask ChatGPT for the best project-management tool for a forty-person agency and it names three brands. Ask again five minutes later and one of them has been swapped out. Ask Gemini and you get a different shortlist with one overlap. For anyone raised on search, where position 3 was a fact you could screenshot, this feels broken. It isn’t. It’s what brand selection looks like when it stops being a ranking problem and becomes a sampling problem. At no point does a large language model “choose” a brand the way a search engine ranks a page. Two very different probabilistic processes produce that shortlist, and almost every claim you’ll read about generative engine optimization (GEO) is really a claim about one of them. Understanding which one, precisely, is the difference between a GEO strategy and GEO folklore.

First, dissolve the word “choose”

A language model is a next-token predictor. A brand name, to the model, is nothing more privileged than a short token sequence with unusually strong statistical gravity. When the model has already generated “For an agency that size, the strongest option is”, what happens next is not a lookup or a query: it’s a softmax over the entire vocabulary, conditioned on everything that came before. Your brand gets named when its first token outscores every alternative at that position, in that context.

That conditioning context is where all the leverage lives, and it comes from exactly two places: what was baked into the weights at training time (parametric memory), and what was injected into the context window at answer time (grounded generation). Every engine (ChatGPT, Gemini, AI Overviews, Perplexity, Copilot) is some blend of the two, and they behave nothing alike.

Fig. 1 · Two routes from prompt to brand mention. Parametric memory answers from frozen training-time associations; grounded generation pulls live passages from the web and cites them.

Parametric memory: the slow consensus

When no retrieval fires, the brand that gets named is the one most strongly associated with the query’s concepts across the training corpus: review sites, forums, documentation, news, listicles, Reddit threads, all compressed into weights. Three properties follow. First, it rewards consensus, not recency: a brand mentioned consistently across thousands of independent sources for years outweighs one brilliant recent page. Second, it’s frozen: nothing you publish today touches it until the next training run, which is why parametric visibility trails reality by quarters. Third, it’s shaped by post-training: models are tuned to hedge on commercial questions, offering a diverse shortlist of two to four options rather than one winner, which is why “category default plus challengers” is the standard answer shape.

You can move parametric memory by changing what the public corpus says about you, everywhere, patiently. But it’s corpus-scale work measured in months. The fast game is elsewhere.

Grounded generation: where citations actually come from

Citations, the linked sources under an AI answer, come almost exclusively from the retrieval path, and the pipeline is more brutal than most GEO writing admits. The engine first rewrites your question into a fan-out of synthetic sub-queries (“best CRM small agency 2026”, “CRM pricing comparison”, “competitor alternatives”). Each hits a search backend: Bing for ChatGPT and Copilot, Google for Gemini and AI Overviews, Perplexity’s own index. Retrieved pages are fetched and chunked into passages of a few hundred tokens, which are embedded and reranked for relevance, freshness and authority. A handful survive into the context window. The model then synthesizes an answer and attributes claims back to surviving passages. Attribution is itself generated, which is why citations occasionally point at the wrong survivor.

Fig. 2 · From one prompt to three-to-five citations: each stage is a knockout round. A page can rank well in the backend search and still lose at chunking or reranking: attrition your analytics never sees.

The strategic consequence is easy to state and widely ignored: the engine does not cite your site; it cites a passage. A 300-token chunk either survives every round of that tournament on its own (self-contained, attributable, information-dense) or your beautifully ranked page contributes nothing to the answer.

The engine doesn’t cite your site. It cites a 300-token passage that survived a knockout tournament your analytics never saw.

What makes a passage win

The best controlled evidence remains the Princeton–Georgia Tech–IIT GEO study that coined the term: across 10,000 queries, adding quotations, statistics and source citations to a page improved its visibility in generative answers by up to ~40%, while classic keyword stuffing was flat to negative. That result makes mechanical sense given the funnel. Rerankers score passages for information density and authority signals; generators preferentially attribute claims that are concrete and quotable. A passage that opens with a direct answer, carries a number, names its source and survives out of context is simply a better candidate at every stage.

Two additions from what we see in Ayan scans, week after week. Entity coherence: the same brand name, category framing and claims across your site, third-party profiles and review platforms. Retrieval hits many of these at once, and contradictions get hedged or dropped. And cross-regime consistency: when a retrieved passage says X but parametric memory learned Y, the model waffles. The brands that dominate AI answers are the ones where both machines agree.

Position was a fact. Probability is a distribution.

Everything above compounds into one methodological rupture: sampling temperature, non-deterministic retrieval and index churn mean the same prompt, on the same engine, on the same day, yields different citations. A rank was a fact; a citation is a rate. The only honest way to measure GEO visibility is repeated, structured sampling (the same prompt set, run across engines, surfaces and markets, again and again) and reading share of voice out of the distribution.

Fig. 3 · Ten identical runs, four engines, one prompt (illustrative). No single run tells you anything; the citation rate per engine is the measurement. Multiply by hundreds of prompts, surfaces and markets to get a real visibility picture.

That loop is exactly what we built Ayan to run: the same prompt sets replayed continuously across ChatGPT, Gemini, AI Overviews, Perplexity and Copilot, across surfaces and markets, so citation share becomes a trend you manage rather than a screenshot you argue about.

Playing both machines

The playbook falls directly out of the mechanism:

• Slow game (parametric): shift the corpus-wide consensus: consistent entity framing everywhere your brand is written about, third-party validation, patience measured in training cycles.
• Fast game (retrieval): win at the passage level: answer-first structure, quotable claims carrying statistics and named sources, chunks that survive out of context. This moves in weeks, not quarters.
• Ground truth (measurement): treat visibility as a distribution: sample the same prompt set repeatedly across engines, surfaces and markets, and manage citation share the way you once managed rankings.

Generative engine optimization isn’t a bag of tricks for gaming a mysterious algorithm. It’s the discipline of making both machines, the frozen one and the live one, agree that your brand is the answer, and of measuring honestly whether they do.

References

Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of KDD ’24. The study that introduced the term and measured content-side interventions across 10,000 queries.