We read the studies so you don't have to.
The supplement information landscape is dominated by SEO blogs, affiliate marketing, and anecdotal forum posts. Clinical literature exists, over 40 million peer-reviewed citations indexed on PubMed alone, but it's locked behind statistical jargon and dense methodology sections. Veraflux is a clinical research tool that turns peer-reviewed supplement literature into structured, fully-cited reports you can check against the source.
What's inside every report
Every Veraflux report follows the same structured anatomy, so you always know where to find efficacy, dosing, safety, and the evidence behind them.
Goal-specific efficacy sections
Two sections tailored to your selected goal: Cognitive Enhancement, Athletic Performance, Sleep Quality, Weight Management, or Immune Support. Each goal has its own specialized analysis pipeline, and the mechanism of action behind each effect is explained alongside the outcomes, so a sleep report doesn't read like a watered-down general overview.
Dosage and protocol
Dose-response data, formulation guidance, and timing protocols where the evidence supports them. Doses are kept separated by standardized extract (e.g., KSM-66 vs Sensoril ashwagandha), not averaged across formulations. When a supplement has no goal-specific dosing data, the report says so explicitly rather than guessing.
Evidence Limitations
A dedicated section in every report surfacing null findings, demographic gaps, baseline dependency, study quality caveats, and what the literature doesn't cover. Most products bury this. We make it a section.
Five-section safety analysis
Generated by a dedicated parallel pipeline (more on that below). Covers:
- Common side effects
- Serious or rare adverse events
- Contraindications
- Drug interactions
- Tolerable upper limit
Bibliography with clickable PubMed links
Every cited study appears in the bibliography with a direct link. You can open any citation, read the original abstract, and verify the claim yourself. This is the core trust mechanism: nothing in the report exists without a study behind it you can check. The report cover also carries a "last reviewed against PubMed" timestamp, so you always know how fresh the underlying evidence is.
Reports are personalized to your age, sex, and goal. Two users searching the same supplement see different studies prioritized, because the relevant evidence is different. See a real example on the home page →
From your search to your report: a multi-stage analysis
Every report passes through a series of specialized stages, each designed to search, evaluate, and verify the clinical evidence before it reaches you.
Systematic Name Resolution
Your supplement may have multiple scientific names, chemical forms, and standardized extract names. We resolve all of them, including chemical identifiers and nomenclature variants, to ensure comprehensive retrieval coverage across PubMed's indexed literature.
Multi-Angle Evidence Retrieval
Rather than a single search, the system dispatches parallel queries across distinct clinical dimensions: pharmacology, dosing, clinical outcomes, and safety. Each facet is tuned to surface a different angle of the evidence base. Results are deduplicated and relevance-ordered, typically yielding 100 to 200 candidate studies per report.
This is the "needle in a haystack" stage. Minority research topics (citrulline for vasodilation, for example) have to survive filtering against dominant research areas (citrulline for athletic performance). The faceted retrieval is designed so the right evidence reaches the next stage even when it's outnumbered.
Study Selection & Scoring
A dedicated selection model evaluates every retrieved abstract across multiple dimensions: study design, population relevance, outcome coverage, and methodological rigor. From the 100 to 200 candidates, roughly 25 to 35 make the cut. The algorithm enforces topic diversity to prevent tunnel vision, with minimum slots reserved for dose-response studies, null findings, and demographic-matched studies so they cannot be displaced.
Supplement reports apply demographic-aware selection, prioritizing studies that match your age, sex, and stated health goal.
Evidence-Grounded Report Synthesis
The final report is generated under strict grounding constraints. Every factual claim must trace to a specific abstract from the selected evidence set. The synthesizer cannot draw on outside knowledge, and when evidence is absent from the retrieved sample, the report says so explicitly rather than filling the gap with plausible-sounding text.
Each section opens with a plain-language paragraph stating the finding before the supporting evidence and citations, and the cover carries a 2-3 sentence summary so a reader can orient in seconds. Coverage is then verified procedurally: any selected study that fails to appear in the citations causes the report to be rejected, not served.
Citation Verification
Every citation is tracked and validated across pipeline stages. If a claim references a study that wasn't in the selected evidence set, the report is rejected rather than served with unverifiable findings.
Safety analysis runs as its own pipeline
Drug interaction worry cuts across every type of supplement user. We treat safety as a parallel system, not an afterthought tacked onto the efficacy report.
Runs in parallel
A separate safety pipeline runs alongside the efficacy analysis. Two specialized systems, one report.
Different search profile
Includes case reports, regulatory signals, and broader safety keywords. Around 100 to 150 abstracts per report.
Five safety domains
Side effects, adverse events, contraindications, drug interactions, and tolerable upper limit.
Graduated knowledge supplementation: the honesty mechanism
For drug interactions and contraindications, abstracts alone aren't enough. PubMed has no clinical trial for "warfarin plus St. John's Wort, what happens?" because no one would ethically run it. So the safety pipeline supplements retrieved studies with established pharmacological knowledge and regulatory records, but every non-study statement carries an explicit prefix.
Here's what it looks like in a real report:
Based on established pharmacological principles:St. John's Wort induces CYP3A4 enzymes, which can reduce warfarin's anticoagulant effect and raise clotting risk in patients on long-term anticoagulation.
Every non-study statement carries a prefix like:
Based on established pharmacological principles:Based on regulatory records:Based on established nutritional guidelines:Based on established clinical experience:
You always know whether a safety statement came from a study, a pharmacology reference, or a regulatory body. Most products either skip outside knowledge entirely (and miss real dangers) or blend it in invisibly. We label it.
Designed for clinical accuracy, not creative guessing
Standard AI tools optimize for a fluent-sounding answer. Veraflux optimizes for fidelity to the peer-reviewed literature.
Standard AI tool
Veraflux
Cites whatever a single search step returns, with no procedural check that each citation is grounded in a selected evidence pool.
Every citation verified against the evidence pool
Every finding traces to a specific peer-reviewed study chosen at the selection stage. The synthesizer is restricted to that selected pool (no blending in outside knowledge), and a citation verifier rejects the report if any cited PMID isn't in it.
One model attempting retrieval, synthesis, and verification at once.
Many specialized stages, not one chatbot
Each stage of the analysis (retrieval, selection, synthesis, verification) is handled by a system specialized for that task. Search, selection, and writing are separate processes, not a single tool trying to do everything at once.
Safety mentioned in passing, mixed in with efficacy claims.
Safety analysis is a parallel pipeline
Drug interactions, contraindications, and dose safety run as their own dedicated system in parallel with the efficacy analysis. Outside knowledge is allowed where it's needed and labeled where it's used, so you always know what came from a study and what came from pharmacology.
Overconfident framing when evidence is thin or contradictory.
Honest about evidence gaps
When a topic isn't covered by the available clinical literature, your report states this explicitly. Every report includes a dedicated Evidence Limitations section that surfaces null findings, demographic gaps, and what the research hasn't addressed.
One-size-fits-all answers, regardless of who's asking.
Reports adapt to your age, sex, and goal
Reports prioritize studies matching your demographics and weight population-relevant evidence higher. A report for a 28-year-old woman focused on performance looks different from one for a 55-year-old man focused on sleep, because the relevant evidence is different.
Presents evidence with even confidence regardless of study quality; no calibrated quality grading.
Calibrated evidence quality ratings
Strong, Moderate, Limited, and Preliminary ratings are calibrated against actual study power thresholds (sample size, replication, effect consistency), so a single underpowered pilot isn't read with the same confidence as a replicated meta-analysis. Thin evidence is rated thin.
How Veraflux compares to the alternatives
Most ways people research supplements have material limitations. Here's where Veraflux is different, and where it's straightforwardly better.
Depth and breadth, by default
A Veraflux report pulls 100 to 200 studies, selects 25 to 35 across pharmacology, dosing, clinical outcomes, safety, and limitations, and synthesizes all of it into a structured report. Almost every alternative is materially shallower on the same supplement. Examine.com is the only competitor with comparable depth, and accessing it still requires manually clicking through long-form articles and chasing study links.
Inline glossary on every term
Every technical term (SMD, p-value, CI, PSQI, RCT, meta-analysis, and many more) is clickable inside the report itself, with a plain-language definition in a popover. No alternative does this. ChatGPT requires follow-up questions, Examine assumes you already know the terms, Google sends you to a separate page.
Limitations and safety, accessible by default
Every report includes a dedicated Evidence Limitations section and a five-section safety analysis. Most alternatives bury this if they cover it at all, and ChatGPT often skips it entirely unless you specifically ask, which most users don't think to do.
Asking ChatGPT
Convenient, freeA fluent answer in seconds.
ChatGPT operates as a single conversation thread. Even with web search and careful prompting, retrieval and synthesis share the same model context, with no separate selection layer that filters and balances candidates before synthesis. The structural consequences: no procedural check that every citation maps to a study in a verified evidence pool; no parallel safety pipeline retrieving case reports and regulatory signals on a different keyword profile; no selection layer reserving slots for dose-response studies, null findings, and demographic-matched evidence so they aren't crowded out by popular results. The structural defaults of a research-purpose pipeline (calibrated evidence quality ratings, a dedicated limitations section, demographic-aware selection, inline glossary on every technical term) aren't on by default either; you assemble them across follow-up prompts and verify the result yourself.
Every claim traces to a specific PubMed abstract you can open and read, with the selection reason for each study visible alongside it. The synthesizer is constrained to its selected evidence pool, and a citation verifier rejects the report if any cited PMID isn't in that pool. Multi-stage architecture (retrieval, selection, synthesis, verification) runs as separate specialized stages, with topic diversity enforced so dose-response studies, null findings, and demographic-matched evidence cannot be displaced by popularity. Parallel safety pipeline, calibrated evidence quality ratings, demographic-aware selection, and inline glossary on every technical term, all on by default.
Examine.com
Editorial depthEditorially curated supplement summaries with comparable evidence depth.
Static articles edited by humans, limited to what their team has covered, so there are no on-demand reports for supplements outside their library. Same article for every reader, no personalization to your age, sex, or goal. Citations live in an end bibliography, so verifying a specific claim means scrolling back and forth between paragraph and reference list. Curation rationale isn't disclosed (you can't see why a particular study was included or omitted). Doses are aggregated across extracts with a "results varied across extracts" disclaimer, so you can't reconstruct which dose came from KSM-66 versus Sensoril versus full-spectrum ashwagandha. Limitations are inline-mentioned rather than given a dedicated section. No inline glossary.
On-demand reports for any supplement on PubMed, personalized to your age, sex, and goal. Doses stay separated by extract (KSM-66, Sensoril, and full-spectrum doses are not averaged together), so you can tell which evidence applies to the product on your shelf. Each paragraph carries its own citations, and each citation surfaces the reason the study was selected. Every report includes a dedicated Evidence Limitations section that names null findings, demographic gaps, and what the literature doesn't cover. Stack analysis covers synergy, antagonism, and pathway overlap across supplement pairs. Inline glossary makes the evidence accessible without leaving the report.
Google / Reddit
Free, familiarFree, familiar, broad coverage.
SEO-optimized blog spam, affiliate marketing, anecdotal forum posts. No synthesis. No quality control. No safety analysis. No way to verify a claim short of finding the underlying study yourself.
Synthesized directly from primary clinical literature. No SEO incentive, no affiliate links, no anecdotes. Every claim links to the study it came from.
Health influencers
Engaging, simpleEngaging personalities, simple takeaways.
Selectively cited. Often sponsored. Cherry-picks studies that support a thesis. Rarely covers safety or limitations.
Selection algorithm enforces topic diversity, includes null findings, and surfaces evidence quality honestly. You see the studies, not someone's interpretation of them. Safety and limitations are baked into every report by default.
What Veraflux can't do (and won't pretend to)
Trust starts with knowing what a tool can't do, not just what it can. These are the real constraints of the current product.
- Efficacy works from abstracts, not full papers
For efficacy, dosing, and clinical outcome sections, the evidence base is PubMed abstracts. Critical data sometimes lives in full-text tables, supplementary materials, or discussion sections we can't see. Publisher partnerships for full-text access (NEJM, JAMA, Cochrane) aren't feasible at our scale.
The safety pipeline is the exception: it draws on outside sources too, established pharmacological references and regulatory records, because dangerous interactions and contraindications often have no clinical trial. Every non-study safety statement is labeled with where it came from, so you always know what's a study finding and what's a referenced principle.
- Reports are documents, not conversations
A Veraflux report is a finished, structured analysis, not a chatbot you can iterate with. If you want to follow up on something specific, you currently generate a different report. Suggested follow-up searches are on the roadmap to bridge this.
- Sex-specific data has gaps in the underlying research
Most supplement RCTs enroll predominantly male participants. We personalize framing and study selection by sex and goal, but we cannot surface sex-specific findings that don't exist. Reports for female users say so honestly rather than papering over the gap.
We tell you these things because trust starts with knowing what a tool can't do, not just what it can.
Who builds this
Veraflux started as one engineer's tool for parsing PubMed without spending hours cross-referencing conflicting blog posts, grounded in a background in mathematics and computer science and years spent on how clinical data is structured, misrepresented, and kept out of reach of the people making daily health decisions.
The philosophy behind Veraflux comes from that foundation: treat health information as a data integrity problem, not a creative writing problem. Every design decision reflects the belief that a system processing clinical research should be constrained by default. It should only tell you what the evidence actually says.
The pipeline is designed with the same rigor you'd apply to any system where incorrect output has real-world consequences: deterministic where possible, explicitly bounded where not, and transparent about its limitations. It has been through more than a dozen rounds of structured iteration against batches of test reports, each round adding specific anti-regression rules so the next version doesn't lose ground the previous one gained.
Frequently asked questions
Common questions from people deciding whether Veraflux is the right tool for them.
How is this different from asking ChatGPT?
Two key differences. First, structural cite-grounding. Veraflux's synthesizer is restricted to a selected evidence pool, and a citation verifier rejects the report if any cited PMID isn't in that pool. ChatGPT, even with web search, has no equivalent verification step: citations may be attached to specific claims without actually supporting them, and outright fabrication is still a known failure mode.
Second, and just as important: a single LLM answer (even from an expert prompter) is materially shallower than what comes out of a multi-stage retrieval and synthesis pipeline. We know this empirically from many rounds of internal prompt iteration. No single prompt, or even a small chain of prompts, ever matched the breadth or depth that the full pipeline produces. Someone with strong prompt engineering skills can squeeze a lot out of ChatGPT, but a chat thread structurally cannot reproduce verified citations, a parallel safety pipeline with its own retrieval profile (case reports, regulatory signals), topic-diversity guarantees that reserve selection slots for dose-response studies and null findings, or demographic-aware selection that weights the studies themselves, not just the framing.
How does the AI turn 25–35 research papers into a readable report?
The synthesizer reads only the abstracts that were selected for your report (typically 25–35 from a candidate pool of 100–200) and writes each section starting with a plain-language paragraph stating the finding, followed by the supporting evidence and citations inline. Every cited PubMed ID has to come from that selected set, so a fabricated citation is structurally harder to produce.
A separate post-synthesis pass then writes a short 2–3 sentence summary at the top of the cover so you can orient before drilling in. The result reads as a digest rather than a research paper.
How current is the data?
Veraflux queries PubMed live whenever your report is first generated. After that, an automated surveillance worker re-checks the literature for new evidence on supplements you've researched, and when something material appears, your report refreshes to reflect it. Nothing is generated from training data.
What supplements are covered?
Anything indexed on PubMed with at least minimal clinical literature. The system handles common, niche, botanical, amino acid, probiotic, animal-derived, and dietary baseline-dependent supplements differently to surface the right evidence and safety considerations for each category. If a supplement has almost no research, the screener tells you up front rather than generating a confident-sounding report from nothing.
What if there's no research on the supplement I want?
The screener catches this and tells you up front. When the literature is thin, the system rates evidence as Preliminary or Limited rather than fabricating confidence. Honest "we don't know yet" beats invented certainty.
Is this medical advice?
No. Veraflux is an evidence summary tool, not medical advice. Always verify findings with the primary studies (every report links them) and consult a healthcare provider for personal decisions, especially if you take prescription medication or have a medical condition.
How accurate are the safety sections?
Drug interactions and contraindications draw on both retrieved studies and established pharmacological knowledge, with every non-study statement labeled by source ("Based on established pharmacological principles:", "Based on regulatory records:", and so on). You always know whether a safety statement came from a study, a pharmacology reference, or a regulatory body.
How is my data handled?
We store your account email and the demographics needed to personalize reports (age, sex, goal). We don't sell your data, share it with advertisers, or use it to train models. See the privacy policy for full details.
Why pay $5/month when I can use ChatGPT for free?
Two different tools. ChatGPT is a great generalist; Veraflux is a specialist for evidence-based supplement research with structural guarantees you can't get from a chat thread. Even with web search and careful prompting, no prompt verifies that every citation maps to a study in a selected evidence pool, runs a parallel safety pipeline with its own retrieval profile (case reports, regulatory signals), or reserves selection slots for dose-response studies, null findings, and demographic-matched evidence so they aren't crowded out by popular results. One supplement that doesn't actually work pays for the year. The free tier gives you one full Veraflux report on signup, so you can compare directly before paying.
Built to augment, not replace
Veraflux is an experimental research tool designed to make clinical literature more accessible. It is not a substitute for professional medical judgment. Always verify findings with the primary sources and consult a healthcare provider for personal decisions.