Public research exploring how AI systems discover, understand, trust, recommend, and influence commercial decisions.
Atom Foundry Research is an independent research initiative focused on AI Commerce Intelligence™. We study AI behavior through real experiments, transparent methods, and public research.
- 50 public research publications
- 40,000+ AI recommendations analyzed
- 1,490 brands observed
- 100 buyer intent prompts
- 5 ecommerce categories
- 66,090 ecommerce stores analyzed
- 100% real world data
- No synthetic datasets
- Public methodology
Every report is based on recorded AI responses, real ecommerce stores, and controlled experiments.
Our main cross category research report.
- The State of AI Recommendations Across Commerce 2026
Controlled experiments that test how AI recommendation and selection work.
- Web Search Rewrites 77% of AI Product Recommendations
- The Fame Study, Corrected
- AI Knows Your Website. It Still Won't Recommend You.
- 29,633 Reasons. 26,812 Unique. The Model Confabulates.
- Search Changes the Vocabulary, Not Just the Brands
- Candidacy vs Selection
- Nothing About Your Brand Predicts Recommendation. The Model's Own Past Behavior Does.
- Two Months Later, the Model Still Agrees With Itself
- Hand It a Rating, and It Follows Every Single Time
- We Invented a Brand With Zero History. Reviews Got It Picked Anyway
- We Widened the Fame Signal Four Ways. It Barely Moved
- The Model Hedges Most When It's Most Sure
- The Model Knows 6 Facts About Your Brand. It Uses One
- It Recommends You First. By Turn Four, It's Moved On.
- Give the Model the One Fact It's Missing. The Brand Goes From Invisible to Everywhere.
- Tell the Model You Saw an Ad. It Recommends That Brand 88% of the Time
- The Model Is Almost Never Wrong About Your Brand. It Just Doesn't Say Much.
- The Model Had Real Web Search. It Never Once Reached for It.
- We Described One Brand to the Model, Real or Invented. It Recommended That Brand Anyway.
- Ask Like You're Googling It, and the AI Recommends the Brand 15.6 Points Less
- Say It Yourself, and the Model Picks You 67.5% of the Time. A Third Party Only Gets 51.1%.
- Specific Claims Add 15.7 Points in Spec Driven Categories. They Cost Nearly 5 in Trust Driven Ones.
- Give It the Better Rating, and It Wins 91% of the Time
- Mention It Was Featured Somewhere, and It Wins 85% of the Time. Bring In a Rating, and It Nearly Disappears.
- Claim the Brand Is Widely Known, and It Wins 80% of the Time. Bring In a Rating, and the Edge Is Nearly Gone.
- The Cited Source Changes. The Winner Almost Never Does.
- Two Signals Get You Most of the Way There. A Third Barely Helps, and Format Doesn't Move It at All.
- We Turned the Same Facts Into Bullets. The Model Cited Fewer of Them.
- Authority Adds 9 Points in Trust Categories. Familiarity Adds 14.
- Reword the Question Three Ways. The Winner Mostly Stays the Same.
- Rating Recommendations Don't Move When You Reword the Question. Price Recommendations Do.
- A Bad Rating Barely Touches Candidacy. It Wrecks the Final Round.
- Mark the Listing Sponsored, and It Loses 18.4 Points in a Head-to-Head Pick
- Match the Sponsored Listing to the Search Term, and the Penalty Overshoots by 21 Points
- Rank First in the Cited Source, and the Model Always Agrees. Fourth Place Wins 1.2% of the Time.
- Under 4.1 Stars, the Model Ignores an Unknown Brand. At 4.9 Stars It Picks It 85% of the Time.
- In a Ranked Source, the Higher-Ranked Brand Wins 100% of the Time, Even Against the Model's Favorite.
- Press Coverage Moves the Model Only When It Answers the Exact Question Asked. Generic: 1.9%. Matching: 80.6%.
- Describe a Company by Its Side Feature and It Falls Off the AI Shortlist: 82.8% to 24.0%. The Model Still Names Its Category Correctly.
- Reviews Four Years Old Count as No Reviews at All: 0%. Five Hundred Fresh Ones Beat Thirty Thousand Old Ones, 72.8% to 1.4%.
Repeated recommendation studies across five ecommerce categories.
- Beauty
- Supplements
- Coffee
- Pets
- Home & Living
Each category report uses the same core research method so results can be compared across categories.
Founder Lab is our public research laboratory.
We also build and document an AI native ecommerce brand in public.
Current publications:
- Founder Lab: Day Zero
- Founder Lab: Research Log
How AI Decides is the living map of the research program.
It describes the working path from memory and retrieval through understanding, candidacy, evaluation, recommendation, stability, confidence, and purchase. It is a research model, not a claim that every AI system follows the same fixed process.
Founder Reality Check documents the practical reality of building Atom Foundry and the research program in public.
Atom Foundry researches how AI systems:
- discover businesses
- understand websites and products
- build candidate sets
- evaluate alternatives
- select brands
- generate recommendations
- explain decisions
- respond to web search
- use information they already have
- deploy only some of the information they know
- change behavior across a conversation
- respond to controlled changes in information and context
- influence commercial outcomes
The goal is not only to measure visibility.
The goal is to understand why an AI system chooses one brand over another.
Every report follows the same principles.
- Real AI responses
- Real ecommerce stores
- Controlled experiments
- Public methodology
- Reproducible analysis
- Independent analysis
- No paid placements
- No sponsored conclusions
- No synthetic datasets
We separate measured behavior from interpretation.
If we cannot measure it, we do not publish it as a finding.
Our research uses several methods.
- AI recommendation experiments
- AI Commerce Score™ measurements
- Website analysis
- AI readability evaluation
- AI understanding analysis
- AI trust analysis
- Cross model comparison
- Web search experiments
- Controlled signal changes
- Longitudinal studies
- Candidate set analysis
- Selection analysis
- Multi turn conversation studies
The research is based on recorded observations and controlled comparisons.
This repository supports the AI Commerce Intelligence™ Framework.
Core concepts include:
- AI Readability™
- AI Understanding™
- AI Trust™
- Recommendation Intelligence™
- Recommendation Share™
- Recommendation Confidence™
- AI Commerce Score™
More information:
https://atomfoundry.dev/framework
Research begins.
First ecommerce stores are scanned.
AI Commerce Intelligence™ is introduced.
Cross category recommendation research begins.
20,000 AI recommendations are collected.
The first mechanism studies are published.
Controlled mechanism research expands.
Research adds studies on stability, selection, fame, search, and signal changes.
Founder Lab research also expands.
Research expands into recommendation intelligence, AI visibility, agentic commerce, and AI decision science.
New studies examine facts, retrieval, hidden context, multi turn conversations, accuracy, brand legibility, attribution, and prompt shape.
Research turns to the signals behind selection itself: claim specificity, star rating, third-party authority, brand familiarity, cited-source stability, and how those signals rank against each other once rating is removed from the comparison. A final study in this run tests whether structuring the same facts as bullets instead of prose changes how much of them a model cites. Two closing studies test whether category type changes how much authority and familiarity signals move the recommended brand, and whether recommendation stability holds up when the same buying intent is worded three different ways instead of repeated verbatim.
Late September adds five more studies: whether recommendations stay stable when the same facts are phrased five ways, what a bad rating does to candidacy versus the final selection, what the sponsored label costs a listing in a head-to-head pick, whether matching a sponsored listing to the search term removes that cost, and how a brand's rank inside a cited source decides the pick.
Research returns to the cold-start problem and to the evidence an unknown brand can bring.
New studies test how star rating and review count move an unknown brand against an entrenched incumbent, whether source rank still decides the pick against the model's own favorite, what a press mention must contain to matter, whether describing a company by a side feature removes it from the AI shortlist, and whether the age of reviews changes how much they count.
The public research inventory reaches 50 studies.
reports/
flagship/
mechanisms/
categories/
founder-lab/
framework/
methodology/
datasets/
observations/
timeline/
Each research report can contain its own summary, methodology, experiment data, and charts.
The first generation of AI commerce focused on visibility.
The next generation will focus on recommendation.
Being discovered is not the same as being recommended.
Being recommended is not the same as being chosen.
Atom Foundry exists to understand the difference.
Main website:
Research Library:
https://atomfoundry.dev/research
AI Commerce Intelligence Framework:
https://atomfoundry.dev/framework
Founder Lab:
https://founder.atomfoundry.dev
If you reference this research, please cite the original report or the Atom Foundry Research repository.
Atom Foundry Research. (2026). Atom Foundry Research (Version v1.0.0) [Computer software]. Zenodo.
https://doi.org/10.5281/zenodo.22753415
For individual studies, cite the original study title, publication date, and the corresponding research repository.
Example:
Atom Foundry. "The State of AI Recommendations Across Commerce 2026." Atom Foundry Research, 2026.
Published by Atom Foundry
Advancing AI Commerce Intelligence™ through open research.