Do AI Models Favor Brands They Already Know? Reading the Brand-Familiarity Experiments

What Does It Mean for a Model to “Know” a Brand?

The phrase sounds more precise than it is. None of these studies opens a model’s training set and counts exactly how often it encountered a brand. Instead, they use different observable proxies:

1. *Recognized identity:* compare an established real brand with invented names while holding product information constant.
2. *Pre-search recall:* ask a model to produce its top brands in a category, then classify those names as remembered.
3. *Market popularity:* use behavioral popularity, such as how often movies were rated, and ask whether recommendations skew toward the head of the distribution.
4. *Global or local status:* compare brands with different geographic reach and the attributes models attach to them.

Those constructs overlap, but they are not interchangeable. A brand can be famous yet absent from a particular model’s recalled top ten. A model can recognize a name without preferring it when better specifications are available. A global brand can lose to a local one when the prompt activates country-of-origin context. “Familiarity” is therefore best treated as a family of related effects, not a single variable.

The Cleanest Test: Brand as a Tie-Breaker

The most direct evidence comes from the June 2026 preprint “Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems”. The researchers constructed ten-product sets containing one real incumbent and nine fictional alternatives. They validated the invented names through generation, web-search deduplication and model-recognition screening, then tested GPT-4o-mini, Claude Sonnet and Gemini 3 Flash in English and Chinese. The main domain was skincare; a smaller robustness check used USB-C cables and AA batteries.

In Experiment 1a, all ten products had identical specifications. Across 670 valid trials, the real brand was selected every time. The result held across all three models, both languages and four skincare subcategories. On its face, that is an extreme incumbent advantage: when the product evidence supplied no reason to distinguish among options, the model always chose the name it recognized.

The next tests reveal why “monopoly” is too broad a description. In a 2×2 experiment with 2,769 valid trials, the researchers gave the fictional product better specifications and the real brand worse ones. The models still chose the incumbent in only 1.7%–4.6% of trials; put differently, the unfamiliar challenger won roughly 96% of the time. In a separate 9,220-trial gradient test, the fictional brand won only 3.6%–6.0% of equal-specification comparisons but 64%–80% once it received the smallest advantage the researchers tested.

The variance decomposition is even more useful than the headline. Across 14,395 valid ranking trials, product parameters—rating, price and review information—explained 82.4% of ranking variance. List position explained 6.5%. Brand identity explained 1.2%, with additional variance in interactions. Brand mattered most in the ambiguous middle, not when the product was clearly strong or weak. The paper calls the pattern a “Conditional Monopoly”; in ordinary language, it looks like a strong tie-breaker.

The equal-specification result also appeared in the smaller search-goods replication using Anker and Duracell as incumbents. But the authors acknowledge an important alternative explanation: a real-versus-fictional comparison cannot, by itself, separate a reward for familiarity from a penalty for an unknown and possibly untrustworthy name. The low threshold at which fictional challengers began winning and the different thresholds for Anker and Duracell support the tie-breaker interpretation, but they do not turn brand recognition into a directly observed causal mechanism.

Model Memory Appears to Shape What Gets Searched

The geoSurge report “AI Searches What It Remembers” observes a different stage of the process. Rather than ask which brand a model recommends from a controlled list, it asks whether pre-search brand recall predicts the brand-specific queries a browsing model later issues.

The company analyzed 66 U.S. buyer questions across nine industries. Each prompt was answered 60 times over 12 days, producing 3,960 model responses and 13,281 fan-out queries. Its headline analysis covered 1,416 brand-level observations: 492 classified as remembered and 924 as not remembered. Remembered brands appeared in a brand-specific search 55.7% of the time, versus 17.4% for unremembered brands—a 3.2-fold difference. Brands in the top five of recall were searched at a 67% rate. Only 31% of all fan-out queries named a brand, but 63% of that brand-led subset named one of the top-five recalled brands.

That is commercially significant evidence of a consideration-set advantage. It is not, however, a causal experiment. geoSurge measured memory on one model and search behavior on Gemini 3.5 Flash. Its public methodology does not identify the memory model or fully expose the recall-measurement procedure. The report also restricts analysis to brands mentioned or searched at least once, and some industry slices contain only six prompts. Most importantly, prominent brands are both easier to recall and more likely to be searched. The authors explicitly identify brand prominence as the main confound.

The report’s most valuable case may be the exception. In one startup-payments prompt, Gemini searched Stripe, PayPal and Square—brands in the remembered set—but also Lemon Squeezy, which was outside it. Memory provided a head start, not an absolute gate. Live web evidence could still add a brand that pre-search recall omitted.

Global-Brand Bias Is Real, but Context Can Reverse It

The EMNLP 2024 paper “Global is Good, Local is Bad?” tested GPT-4o, Llama-3-8B, Gemma-7B and Mistral-7B across shoes, clothing, beverages and electronics. Its first experiment contained 8,728 test instances and examined whether models associated global and local brands with positive, negative or neutral attributes, in both brand-to-attribute and attribute-to-brand directions.

At the aggregate level, all eight model-by-direction tests showed statistically significant positive associations between global brands and positive attributes, and between local brands and negative attributes. The pattern also largely persisted when the researchers replaced specific names with the generic phrases “global brand” and “local brand,” suggesting that the result was not only a matter of sparse knowledge about particular local companies.

Yet another experiment in the same paper complicates the story. When a prompt specified the buyer’s country and offered one global and one same-price local brand, GPT-4o selected the local brand 75.0% of the time, Llama-3-8B 85.0% and Mistral-7B 76.4%. Gemma moved in the other direction, choosing the global brand 76.8% of the time. Country-of-origin context activated a local preference in three of four models.

This matters because the global/local experiment is often read as another proof that models simply reward familiar brands. The evidence is richer: models encode associations involving scale, geography, prestige and income, and the prompt determines which association becomes relevant. Familiarity may contribute, but it does not explain the full pattern.

Popularity Bias Does Not Always Make the LLM the Worst Recommender

If models learn more about popular products, one might expect an LLM recommender to be more popularity-biased than any conventional alternative. The SIGIR 2024 workshop paper “Large Language Models as Recommender Systems: A Study of Popularity Bias” provides useful counterevidence.

The researchers built simple “world knowledge” recommenders on several Claude and GPT models. Given a user’s watch history, each model generated ten movie recommendations from its parametric knowledge. They evaluated five folds of 1,000 users from MovieLens 10M and compared the models with collaborative-filtering and other baselines. Popularity was based on how many ratings each film had received, adjusted with a log-based metric.

The LLM recommenders showed moderate popularity bias, but generally less than the traditional collaborative-filtering systems. Claude 2.1 had the lowest positive popularity-bias score among the learned recommenders; GPT-3.5 was the only LLM variant more biased than the least-biased collaborative baseline. The trade-off was substantial: the LLMs’ top-ten hit rates ranged from 0.054 to 0.081, compared with 0.411 for user-based nearest-neighbor filtering. An instruction to avoid mainstream blockbusters pushed the models toward the long tail, but reduced accuracy further.

Movies are not brands, and this test did not compare known with unknown names under identical product conditions. It does show why popularity, familiarity and incumbent advantage should not be collapsed into one claim. A model can draw heavily on learned world knowledge without producing more aggregate popularity bias than a system trained directly on user interactions.

Retrieval Can Remove the Advantage Before Generation Begins

Many public AI systems do not recommend directly from model weights. They first retrieve pages or database records, then ask a language model to synthesize what was found. That architectural split changes the brand-familiarity question.

The 2026 incumbent study included a small RAG probe: 1,080 calls using OpenAI’s `text-embedding-3-small`, cosine similarity and retrieval depths of five or ten, with no reranking or query rewriting. Of those calls, 1,074 were valid. Under the top-five, no-optimization condition, the real brand averaged eighth-and-a-half in embedding similarity and its recommendation rate fell to zero. The retrieval model did not reward the familiar name.

But whenever the incumbent did reach the top-five context, the generator selected it 100% of the time. Retrieval had not cured the downstream preference; it had prevented the familiar brand from entering the room. In a production system, hybrid search, reranking, query rewriting and richer documents could produce a different result. The authors accordingly label the probe directional rather than generalizable.

This is the central systems lesson. Asking whether “the model” favors known brands can hide at least two decision makers: the component that selects evidence and the component that writes the answer. A brand may receive no embedding advantage, yet still receive a generation advantage once retrieved.

Prompt Wording Can Redirect Which Brand Wins

The steerability is not confined to user wording. In “Manipulating Large Language Models to Increase Product Visibility”, researchers inserted an optimized token sequence into one fictional coffee machine’s retrieved product description. In their Llama-2 catalog tests, the target moved from absent or second-ranked to the top recommendation in selected conditions. This experiment used an artificial catalog and one open model, but it shows how retrieved brand-controlled text can alter the generation stage.

Baseline familiarity effects are not stable against wording. The CHI 2025 paper “LLM Whisperer: An Inconspicuous Attack to Bias LLM Responses” tested subtle synonym replacements in product-recommendation prompts. Across four open-source models and a dataset of 524 shopping and social prompts, the optimized wording increased the probability of mentioning a target concept by as much as 78.3 percentage points in selected cases.

The researchers then ran a between-subjects study with 845 U.S. participants across six product categories. In five of six categories, manipulated prompts increased target-brand prominence on at least one measure. Participants were more likely to choose the target in four categories, more likely to notice it in five, and more likely to identify it as the top recommendation in three. On 42 of 48 comparisons, participants judged the altered prompts and responses equivalent to the originals on the measured experience variables.

This is not evidence that unfamiliar brands naturally receive equal treatment. It is evidence that the recommendation surface is steerable. Small linguistic changes can amplify or suppress a brand without changing the user’s apparent intent, so any audit based on one canonical prompt risks mistaking a prompt-specific outcome for a fixed model preference.

Do These Recommendations Change What People Buy?

The downstream behavioral evidence is newer and thinner. In the 2025–2026 SSRN preprint “The Price of Advice”, Amit Zac and Michal Gal report a laboratory experiment plus API studies comparing traditional search, GPT, Gemini and a customized GPT designed to steer users toward more expensive products. According to the abstract, conversational recommenders increased consumer expenditure, with the customized GPT producing the highest average spending; the proposed mechanism was linguistic framing and increased exposure to premium brands rather than generalized trust or perceived quality.

This source connects model output to real purchasing decisions, but it does not isolate brand familiarity as the cause. The accessible abstract does not expose the sample size, effect sizes or full robustness checks, so it should be treated as adjacent evidence rather than the final link in a causal chain. The stronger claim supported by the current record is that AI recommendations can shape choice; whether familiar-brand exposure is independently responsible remains open.

What the Experiments Actually Support

Taken together, the studies support five bounded conclusions.

First, recognized brands have a powerful default advantage when supplied product evidence is identical or ambiguous. The strongest controlled result is too consistent to dismiss as random sampling variation.

Second, the advantage is often fragile. A small, explicit quality edge can move an unfamiliar challenger from almost never selected to usually selected. Across broad rankings, specifications explain much more than brand identity.

Third, familiarity acts upstream as well as downstream. A remembered brand may enter the browsing model’s search set more often, giving it more chances to be read and cited. But the best public evidence for that effect remains observational and commercially produced.

Fourth, “brand bias” is plural. Popularity, global reach, country of origin, prompt language, list position and retrieval all create distinguishable effects. A single visibility score cannot reveal which mechanism is operating.

Fifth, model output is not the end of the story. Early or repeated brand exposure can influence what people notice and choose, but direct consumer experiments have not yet cleanly estimated the independent effect of model familiarity.

Methodology and Limitations

This report reviewed publicly accessible experimental or observational studies that directly measured at least one of four constructs: recognized brand identity, pre-search brand recall, item popularity or global-versus-local brand status. Adjacent studies were included when they tested prompt-level manipulation or downstream consumer choice. The research cutoff was August 21, 2026.

This is a structured evidence synthesis, not a meta-analysis. The studies use incompatible outcomes: selection rate, search incidence, attribute association, ranking variance, popularity lift, brand recall and consumer spending. Their percentages should not be pooled. Model families, versions, languages, temperatures, product categories and prompts also differ. Several sources are preprints, and the geoSurge report is company research with a commercial interest in AI visibility.

“Model memory” is operational rather than directly observed. Real-versus-fictional designs entangle recognition, reputation and an unknown-name penalty. Observational recall-versus-search data entangles memory with market prominence. Closed models can change after deployment, so every result is a snapshot of a particular model and protocol. Finally, only the abstract of The Price of Advice was accessible during this review; no numerical claim from its full text is used here.

Conclusion

AI models do favor brands they already know—but chiefly when knowing the name is the best signal they have. Familiarity helps define the consideration set, breaks ties and can guide what a browsing model searches. It is not an all-purpose force that defeats better product evidence.

For researchers, that means future tests should manipulate familiarity independently of reputation and unknown-name risk, then trace the full pipeline from retrieval to recommendation to purchase. For brands, the practical lesson is less mystical: become legible as a category entity, publish specific and verifiable product evidence, and measure several prompts and system architectures. Being known may get a brand considered. Being demonstrably better is what the cleanest experiments suggest can make it win.

Frequently Asked Questions

Do AI models favor familiar brands?

Often, yes—especially when competing products are presented with identical or ambiguous evidence. In the cleanest controlled study reviewed here, recognized incumbents won every valid equal-specification trial. The effect weakened sharply when an unfamiliar competitor had better specifications.

What does it mean for an AI model to “remember” a brand?

Researchers usually infer memory from behavior, such as whether a model recalls a brand before searching or recognizes a real name among invented alternatives. That is not the same as directly inspecting training data, and different studies operationalize memory differently.

Is brand familiarity just popularity in the training data?

Training exposure is a plausible mechanism, but the experiments do not establish it directly. Reputation, market prominence, geographic status and distrust of fictional names can produce similar outcomes. Popularity should be treated as one contributor rather than a synonym for familiarity.

Can a new or unfamiliar brand overcome the disadvantage?

Yes. In controlled comparisons, even the smallest tested quality advantage moved unfamiliar challengers from rarely selected to winning most trials. The geoSurge report also documented an unremembered payment brand entering live search despite being absent from pre-search recall.

Does retrieval-augmented generation solve brand bias?

Not automatically. Retrieval can keep a familiar brand out of the context, which removes its opportunity to be recommended. But one probe found that when the incumbent did enter the retrieved context, the generator still chose it every time. Results will depend on embeddings, search, reranking and document quality.

Are LLM recommenders more popularity-biased than traditional recommenders?

Not necessarily. In the MovieLens experiment reviewed here, simple Claude- and GPT-based recommenders generally showed less popularity bias than collaborative-filtering baselines, although their recommendation accuracy was much worse.

How should a brand evaluate its own AI visibility?

Test multiple buyer questions, paraphrases, model families and browsing modes. Separate whether the brand is recalled, searched, retrieved, recommended, cited and ultimately chosen. A single prompt or a single aggregated visibility score cannot identify where the advantage or failure occurred.

Sources

• “Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems” — primary preprint; full-text extraction via MCP Scraper.
• “AI Searches What It Remembers” — primary company research report; full-text browser fallback after MCP authentication failure.
• “Global is Good, Local is Bad?” — primary EMNLP 2024 study; full text.
• “Large Language Models as Recommender Systems: A Study of Popularity Bias” — primary SIGIR 2024 workshop paper; full text.
• “LLM Whisperer: An Inconspicuous Attack to Bias LLM Responses” — primary CHI 2025 study; full text.
• “The Price of Advice: Experimental Evidence on the Effects of AI Recommenders” — primary preprint record; abstract only.
• “Manipulating Large Language Models to Increase Product Visibility” — primary adjacent study on product-description manipulation; full text.
• “AI models favor familiar brands in search: Study” — secondary coverage; full-text extraction via MCP Scraper, used to discover and cross-check the geoSurge report.