How AI Models Rank Brands in Search Results

No vendor publishes a formula for how its AI answers rank brands. OpenAI says only that ChatGPT "ranks search results using multiple factors intended to help users find relevant, reliable information," and that "placement is not guaranteed." What we do have is each vendor's documentation of its search step, plus research on how language models weigh evidence. This post puts the two side by side.

If the AI describes an older version of your company, that's a different problem, covered in our guide to semantic drift and running an AI "memory reset" after a pivot. To measure where you rank, use the formulas in our GEO metrics framework.

Where an AI Brand List Comes From

Every brand list in an AI answer comes from one of two places, or both.

The first is the model's trained memory: associations between brands and categories, learned from the web before its knowledge cutoff. If the engine doesn't search, this is all it has.

The second is retrieval. The engine searches, reads pages and writes from them. Google says its AI features rely on core Search ranking systems to retrieve pages, and may use query fan-out, "a set of concurrent, related queries generated by the model." Microsoft says Copilot "centers its response on high-ranking content from the web." ChatGPT rewrites your question into targeted queries for its search providers, and Perplexity searches its own index.

So when people ask how AI models rank brands, the honest answer has two halves. Ranking inside the model is invisible and slow to change. Ranking in the search step follows familiar SEO rules, plus a few new ones.

A rank is a rate, not a slot

An AI answer has no fixed positions. When SparkToro had 600 volunteers run 12 prompts 2,961 times, the odds of the same list twice were under 1 in 100, and the same order came up about once in 1,000 runs. A 2026 paper, "Don't Measure Once", concludes that visibility is "a distribution rather than a single-point outcome." Everything below describes forces that shift the odds, not rules that fix a slot.

The Five-Filter Stack: How AI Models Rank Brands

We call our model of this the Five-Filter Stack. A brand passes through five filters on its way into an answer, and each one can promote or drop it. The model is ours. The evidence under each filter is not.

Filter 1: Prior, or what the model already believes

Before it reads anything, a model already has a view of how to rank brands in your category. Research shows that view leans toward the famous.

• In a study presented at ECIR 2024, models used to rank items "can be biased by popularity or item positions in the prompts."
• In an EMNLP 2024 paper, models were found disproportionately associating global brands with positive attributes.
• In a 2025 test of 50 unbranded cola prompts, major brands made up 56.3% of ChatGPT's brand mentions and 67.9% of Perplexity's. Niche brands got 12.3% and 5.8%.

The prior also resists new evidence. In Xie et al. (ICLR 2024), when GPT-4 saw evidence for both its memorized answer and a rival one, it kept the memorized answer 80% of the time on the most popular entities. For a small brand, that cuts both ways: a model that knows little about you is more open to what it retrieves.

Filter 2: Retrieval, or which sources get read

When an engine searches, its source ranking decides which brands it even sees. Perplexity describes the pipeline behind its search API, which it says is built together with its own products. It retrieves by both keywords and meaning, filters out "clearly non-responsive or stale content," then narrows the set with faster scorers before "more powerful cross-encoder reranker models" make the final cut. It scores passages as well as whole pages.

The sources that win tend to be third-party. In Chen et al.'s 2025 study, a web-enabled GPT answering US software queries drew 72.7% of its sources from earned media, 26.7% from brand sites and almost none from social platforms. Google's results for the same queries split 45.4% earned, 43.7% brand and 10.9% social. For ChatGPT recommendations specifically, Ahrefs found "best X" lists made up 43.8% of cited page types.

So the pages that rank brands for an AI engine are mostly roundups, reviews and publications, not the brands' own sites. Your site still counts: brand sites made up about a quarter of GPT's sources in that study.

Filter 3: Consensus, or how many sources agree

Once the engine has its sources, agreement decides how it will rank brands among them. Xie et al. found that "LLMs generally provide answers backed by the majority of evidence," and that "the higher the proportion of evidence supporting a particular answer, the more likely LLMs will return that answer."

Large correlation studies point the same way. Across 75,000 brands, Ahrefs found branded web mentions correlated 0.664 with ChatGPT visibility, against 0.266 for domain rating. That's a correlation, not proof of cause, but it fits the majority finding: brands named on many pages win more answers.

Conflict works against you. Microsoft's Bing team wrote in May 2026 that when sources contradict each other, a grounding system must "register that conflict," and that abstention is valid "when support is missing, stale, or conflicting." Inconsistent facts about your brand can cost you a place in the list.

Don't try to fake consensus. Google warns that seeking inauthentic mentions "isn't as helpful as it might seem," because its AI features depend on the same systems that block spam.

Filter 4: Position, or where your name sits

Language models don't read their input evenly. In "Lost in the Middle", accuracy was "often highest when relevant information occurs at the beginning or end of the input context." Xie et al. found order effects that differed by model: "When evidence is presented first, ChatGPT tends to favor it," while two other models leaned toward later evidence.

For product rankings, Pfrommer et al. (EMNLP 2024) found that "different LLMs vary significantly in prioritizing product name, document content, and context position." And inside the pages engines read, placement seems to help: Ahrefs found a correlation between ranking high on third-party "best" lists and being recommended by ChatGPT.

These studies tested models from 2023 and 2024, and no vendor documents how position affects the way today's engines rank brands. Take the lesson, not the exact effect: small placement differences move lists.

Filter 5: Personal context, or who is asking

Engines increasingly rank brands for a person, not a query.

• ChatGPT may use saved memories when it rewrites a search query, and an approximate location from the user's IP address.
• Google's AI Mode can use Personal Intelligence, which draws on Gmail and Photos. Google's own example suggests coats that fit "your preferred look," and it expanded to free US users in March 2026.
• Perplexity stores structured preferences, "like favorite brands", and turns memory off in incognito mode.
• Copilot can store facts users ask it to remember.

You can't optimize for one person's memory, but you can keep it out of your tests. OpenAI's temporary chat has an Unpersonalized option that skips memory and custom instructions.

What Doesn't Change How AI Models Rank Brands

A few levers look promising and aren't.

• Ads. OpenAI says ads "do not influence ChatGPT's answers" and that advertisers "have no ability to shape, rank, or alter" responses.
• Special markup. Google says structured data isn't required for its AI features and there's no special schema to add.
• Hidden instructions. Pfrommer et al. showed that prompt injection can push low-ranked products up, even on perplexity.ai. Bing's guidelines say attempts to manipulate its language models "may result in reduced visibility or removal."
• Third-party tools that claim inside access. Google says "no third-party tool has access to our internal ranking or AI systems."

Find the Filter That Drops Your Brand

The Five-Filter Stack is most useful as a diagnosis of why engines rank brands the way they do in your category. Run a small prompt panel twice: once in each app with search on, in an unpersonalized session, and once through each vendor's API with no search tool, which shows the prior alone. Then match your pattern to a row.

A worked example

Plannora, a made-up project planning tool, runs 10 buyer prompts five times each, 50 answers per mode. The numbers are illustrative.

• Search on: Plannora is named in 22 of 50 answers, a 44% visibility rate. Every answer names at least one brand, and Plannora comes first in 3 of them, a first-mention rate of 3 ÷ 50 = 6%. Across its 22 mentions, its average answer position is 3.4.
• Search off: it's named in 6 of 50 answers, 12%.

The 32-point gap says retrieval carries Plannora while its prior is weak, which is normal for a young brand. The 6% first-mention rate points at Filter 4. So Plannora works on placements near the top of the roundups engines cite, states plainly who it's best for, and rechecks monthly.

The first-mention rate and average answer position formulas are in our metrics framework. To follow rivals week by week, see how to track competitor rankings in AI search. For a one-off score, see benchmarking AI citations against competitors, and for ChatGPT's shortlist, how to rank on ChatGPT.

Where Rankbox Helps

Rankbox doesn't track AI citations, mentions or rankings today, so run the diagnosis above by hand or with a tracker. Its role is in Filters 2 and 3. Answer-Space Research maps the questions buyers ask AI in your category, and the Citation-Ready Writer writes source-backed articles, such as clear comparison and use-case pages, that reach your site through Rankbox's API. The Business plan is $49.50 a month with a 7-day trial. See pricing.