How AI assistants choose which brands to recommend

Two mechanisms, not one

"How does ChatGPT decide?" gets asked as if there's a single ranking formula somewhere. There isn't. Two separate mechanisms combine, and mixing them up is why a lot of advice on this topic contradicts itself.

• Recall from training data. A model's baseline sense of a category is shaped by how often, and how positively, a brand appeared across the text it was trained on. This is slow to shift and mostly outside your control this quarter.
• Live retrieval. For anything current, tools like ChatGPT and Perplexity run a live search, pull back a set of candidate pages, and generate an answer grounded in what they retrieved. This is the mechanism an ordinary business can actually influence, and it's the one this piece is about.

Retrieval itself has two parts worth knowing, per Google's own definitions for its AI features: retrieval-augmented generation, which pulls specifics from ranked pages, and query fan-out, where the model silently generates several related searches instead of just one to fill out its answer (Google Search Central). Being retrieved for the obvious query is not enough. You need to be retrievable for the several adjacent ones the model generates on its own.

Google rank still predicts AI mentions

The clearest evidence that this is not a mysterious separate system: a study by Seer Interactive analyzed over 300,000 keywords across finance and SaaS, which triggered nearly 600,000 "People Also Ask" questions. The team narrowed that to 10,000 relevant questions and queried each through GPT-4o's own API to measure which brands got mentioned. Brands ranking on Google's first page showed a correlation of roughly 0.65 with being mentioned by the model, a meaningfully strong relationship (Seer Interactive).

That number alone should reset expectations. If you already rank well on Google for a query, you are already most of the way to being retrievable for the AI answer to the same question. This is the same conclusion Google states directly about its own AI features, now with independent third party measurement behind it for a standalone model too.

The backlink surprise

The same study measured backlinks, the single most talked about ranking factor in traditional SEO, against AI mention rate directly. The result: backlinks showed minimal or neutral measured impact on whether GPT-4o mentioned a brand (Seer Interactive). The researchers note this contradicted their own expectations going in.

Read carefully, not as "backlinks are dead." Google ranking still correlates strongly, and backlinks remain one input into that ranking. The finding is narrower and more useful than that: once a page is already ranking, chasing more backlinks specifically to move an AI citation number is not where the leverage is. The leverage sits in what's covered next.

Why list position still matters

Retrieval returns a ranked set of candidate pages, and where you land in that set is itself a factor, confirmed as one of four gatekeeper properties in the largest controlled citation study to date, which ran 252,000 trials across six language models (What Gets Cited, SIGIR 2026, arXiv:2605.25517). This tracks with a broader pattern documented across language model research generally: models exhibit measurable position bias, weighting earlier and more prominent items in a given context more heavily, a pattern researchers have traced back to how attention accumulates across a model's layers (arXiv:2512.02665). Retrieval, then generation, both have a position effect stacked on top of each other.

What happens when everyone optimizes

A 2026 study testing GPT-4o-mini, Claude Sonnet, and Gemini 3 Flash on products with identical specifications found that a well known incumbent gets recommended in effectively every trial by default, a "conditional monopoly" that a challenger can break with a small, specific edge, evidenced language, or a marginal rating advantage (arXiv:2606.17443). I go through exactly how small that edge needs to be in the companion diagnostic piece.

The same research modeled what happens once every brand in a category adopts the same optimization tactics: individual payoff for each brand collapses from +0.802 down to +0.007, an almost total erosion (arXiv:2606.17443). This is the honest ceiling on any "GEO trick." Being early and specific breaks a monopoly. Everyone doing the same trick simultaneously returns the category to something close to where it started, minus the effort spent.

The format with outsized weight

A separate 2026 study tracking over 100,000 prompt responses across more than 100 brands between March and May found that a single content format carries disproportionate citation weight: best-of listicle pages accounted for roughly 21% of all citations measured, the highest share of any single format (arXiv:2606.20065). If a model is choosing between candidates, a page that has already done the comparison work, "best CMMS software," "top CRM alternatives," gives it a structure to lift directly rather than synthesizing one from scattered pages.

Where manipulation gets caught

Because this can be gamed, it's worth knowing where that stops working. GEO-Bench, a 2026 benchmark built specifically to test manipulation techniques against ranking systems, evaluated black-box prompt based attacks against white-box gradient based methods and ten defensive strategies. Its core finding: effectiveness and stealth trade off against each other, and access level, whether an attacker has deep model access or just prompt access, did not reliably predict how well an attack actually worked or how detectable it was (arXiv:2605.29107).

In plain terms: obvious manipulation, fabricated claims, stuffed keywords, artificial "mentions" bought across the web, is a live research target for people building detection systems, not a stable hack. It also matches what Google says directly about manufactured mentions in its own AI guidance: quality systems focus on genuine signals, and separate systems exist specifically to catch this kind of manipulation.

What this means for you

1. Rank well on Google first. The ~0.65 correlation means this is still the foundation, not a separate track.
2. Stop over-indexing on backlinks specifically for AI citation. They're not harmful, they're just not where this particular lever sits.
3. Compete for position within the retrieved set, not just for inclusion in it. Specificity and evidence are what move you up.
4. Publish genuine comparison and "best of" style content where you can back it honestly. It carries outsized weight in what gets cited.
5. Move now, while the category isn't saturated. The advantage from being specific and evidenced shrinks as competitors catch on.
6. Don't manipulate. It's increasingly detectable, and the honest path already works.

A proper audit walks your actual pages against all six of these. That is what I do.

Primary sources: Seer Interactive, What Drives Brand Mentions in AI Answers · Google Search Central · What Gets Cited: Competitive GEO in AI Answer Engines, SIGIR 2026 (arXiv:2605.25517) · Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems (arXiv:2606.17443) · Generative Engine Optimization at Scale (arXiv:2606.20065) · Input Order Shapes LLM Semantic Alignment in Multi-Document Summarization (arXiv:2512.02665) · GEO-Bench (arXiv:2605.29107). All sources accessed July 25, 2026.