How BS Inspector Works

This page explains, in plain English, how BS Inspector produces its product analyses. We publish it because we think readers, product owners, and brand operators all deserve to know how a verdict about a product is reached, what the verdict actually means, and what the verdict does not mean.

BS Inspector is operated by Sawary Trading LLC, a Wyoming Limited Liability Company in the United States (“we”, “us”, “our”). The Service includes our website at bsinspector.com and our browser extension distributed via the Chrome Web Store, Microsoft Edge Add-ons, and Firefox Add-ons.

1. The short version

When you click the BS Inspector badge on a product page, our extension sends the public URL of that page, the product name and category, and a limited excerpt of public text from the page to our servers. Our servers run a large-language-model (LLM) analysis that:

• searches publicly available sources on the open web for independent reviews and discussions of the product;
• weights those sources by an internal credibility framework (described below);
• extracts and cross-checks specific factual claims that appear in the product listing or in marketing copy;
• produces a numeric BS score (0–100), a confidence score (0–100), a categorical verdict, a written summary, lists of strengths and weaknesses, and any red flags worth highlighting;
• declines to draw a strong conclusion when the information available is too thin, too new, or too generic to support one.

The output of that analysis is our opinion, formed by an automated process and grounded in publicly available evidence. It is published as informed editorial commentary, not as a statement of established fact about any product or seller. The remainder of this page describes that process in detail.

2. Our editorial position

We publish BS Inspector analyses as labeled subjective opinion based on a disclosed methodology and on publicly available evidence at the time of analysis. Reasonable, well-informed readers reviewing the same evidence could reach different conclusions, and we expect them to.

Where our analyses include statements that read as factual (for example, “multiple reviewers report battery degradation after six months”), those statements describe what we observed in publicly available sources at the time of analysis. They are not claims of objective scientific truth about the product, and we do not represent them as such.

Where our analyses include opinion words such as SUSPICIOUS, AVOID, CAUTION, or TRUSTWORTHY, those are conclusory shorthand for whether the publicly available evidence we found, weighted by our credibility framework, points more toward consumer caution or consumer confidence. They are not legal judgments, medical advice, regulatory determinations, or accusations of fraud.

We accept that any methodology, however carefully designed, can produce wrong or incomplete results. We commit to a public correction process (Section 15) and to updating analyses when we are shown evidence we missed.

3. The analysis engine

3.1 The model

Every BS Inspector analysis is produced by a large-language-model (LLM) provided by Google LLC. As of this writing, our primary model is Google’s Gemini 3 Flash Preview, with Gemini 2.5 Flash as an automatic fallback when the primary model is unavailable. We will update this page when we change models.

The model is augmented with Google Search grounding and additional Google Custom Search calls so that the analysis is informed by sources outside the product page itself. We do not use third-party data brokers, undisclosed paid databases, or scraped private content.

3.2 Why an LLM and not a rules engine

Product BS—misleading marketing claims, paid-influencer-driven hype, review manipulation, vague or implausible specifications—does not have a finite set of patterns that a rules engine can enumerate. It varies by category (a “BS” cosmetic ingredient claim looks very different from a “BS” battery-life claim), it varies by language and region, and it evolves over time. A modern LLM augmented with web search is the best general tool we currently have for reading the same kinds of evidence a careful human reviewer would read and summarizing what that evidence suggests. It is also imperfect: see Section 13.

3.3 What we tell the model to do

Our system instructions to the model direct it, in summary, to:

• read the product page and category context;
• search the web for independent reviews, forum threads, expert assessments, and aggregated sentiment;
• extract individual marketing claims (e.g., “30-hour battery life”, “clinically proven”, “dermatologist-tested”) and check whether independent sources support, contradict, or fail to corroborate each one;
• weight evidence by source credibility (Section 6);
• score the product across category-specific dimensions (Section 11);
• assign a BS score and a verdict only when there is enough evidence to support a meaningful conclusion, and otherwise flag the analysis as insufficient data;
• use opinion-framing language (“reviews suggest”, “multiple sources report”) rather than bare factual assertions about the product or its maker;
• refuse to analyze products in our prohibited categories (Section 12).

4. What we analyze

The data sent from the browser extension to our servers for each analysis is:

• The public URL of the product page you scanned.
• The product name and, where available, description and price.
• The product category (one of the categories listed in Section 11), inferred from the URL and page or selected by you.
• An excerpt of public page text, capped at approximately 8,000 characters of HTML text content. This excerpt comes from the same page anyone visiting the URL would see; it does not include private content, account-only content, paid content, or pages behind a login.
• Your country and preferred language, used to localize the analysis (e.g., to weight regional review sources and to translate the output).

What we do not send to the model:

• Your browsing history.
• Pages you have not explicitly scanned.
• Your account email, name, payment information, or any other personally identifying detail. The analysis is keyed to the product URL, not to you.

Our Privacy Policy describes the full data flow.

5. Where our information comes from

For each analysis the model is grounded with publicly available content from sources that broadly fall into these groups:

• Independent forum and community discussion (e.g., Reddit threads, specialist forums, enthusiast communities) where users describe their actual experience with a product without commercial sponsorship.
• Independent consumer-protection and review publications (e.g., Consumer Reports, established review sites and publications with editorial separation from advertisers).
• Specialist editorial coverage (e.g., expert review sites within a product category).
• Mainstream review aggregators (e.g., the review sections of major retailers and price-comparison engines).
• Manufacturer and seller marketing (the product page itself, official brand pages, official spec sheets).
• Regional and local sources matching the user’s country (e.g., regional review publications, regional forums).

We do not pay any of these sources. We do not have commercial agreements with review publications, retailers, or manufacturers that influence which information surfaces in an analysis. We do not accept payment from product owners to alter, suppress, or favor an analysis (see Section 16 for our policy on this).

6. Source credibility framework

Not all sources carry the same weight. A user’s account of their own experience over six months on a specialist forum is not the same kind of evidence as a manufacturer’s own marketing copy. Our system applies an internal credibility weight to each source so that better-grounded evidence has more influence on the analysis. The current weights are, approximately:

• Independent expert and consumer-protection publications (e.g., Consumer Reports, established testing labs): weight ~0.95.
• Specialist communities and forums (e.g., relevant subreddits with established norms, enthusiast forums): weight ~0.95.
• Independent specialist review sites (e.g., category-specific editorial sites with disclosed review methods): weight ~0.85–0.90.
• Generalist tech and consumer review sites: weight ~0.75–0.85.
• Major-retailer review aggregations (e.g., Amazon, Best Buy review sections): weight ~0.55–0.60. These are useful but are subject to incentivized reviews, manipulation, and platform-side moderation choices.
• Manufacturer and brand-owned content (the product page itself, brand sites): weight ~0.25–0.30. Marketing content is informative about what is being claimed but is not independent evidence of whether the claims are true.

These weights are heuristics, not laws of nature. They do not penalize a retailer or manufacturer for being a retailer or manufacturer; they reflect the well-established observation, in consumer-protection literature, that commercial incentives reduce the independence of self-published claims. We review the weighting framework periodically and will update it on this page when it materially changes.

7. How a verdict is produced

For each scan, the pipeline runs roughly as follows:

1. Detection. The extension recognizes a supported product page (e.g., on Amazon, Walmart, Target, Best Buy, eBay, Etsy, AliExpress, Trendyol, Hepsiburada, Noon, Udemy, Coursera, and many others) and offers to run an analysis when you click its badge.
2. Category classification. The product is classified into one of the categories in Section 11 (or marked “general”), which determines the dimensions on which it will be evaluated.
3. Prohibited-category check. If the product falls into a category we will not analyze (Section 12), the request is rejected before any analysis runs and no charge is made.
4. Cache check. If we have a recent analysis (less than 7 days old) for the same product URL and category, we return that cached analysis. This keeps results stable across users in a short window. See Section 13.
5. Source gathering. The model issues several category-aware web searches to gather independent coverage of the product (between roughly 8 and 14 search queries per analysis, depending on category and locale).
6. Claim extraction and verification. Marketing claims are extracted from the product page; each is checked against the gathered sources and labeled as supported, contradicted, or uncorroborated.
7. Dimension scoring. The product is scored on the dimensions associated with its category (e.g., for electronics: build_quality, battery, display, thermals, upgradability, value, reliability; for skincare: ingredients, effectiveness, side_effects, irritation, longevity, value; etc.).
8. Aggregation. Dimension scores, source-weighted evidence, and red-flag signals are aggregated into a single BS score (0–100), a confidence score (0–100), and one of four verdicts (Section 8).
9. Sufficient-data check. If the model determined that the available evidence base was too thin to support a confident verdict, the analysis is flagged with insufficient_data: true and the confidence score is capped (Section 9).
10. Translation. If your preferred language is not English, the analysis is translated for display.
11. Delivery. The analysis is returned to your extension and displayed in the side panel.

8. What the four verdicts mean

Every analysis ends with one of four labels. These labels are shorthand for our opinion based on the evidence we found, weighted by the framework above. They are not regulatory ratings, scientific determinations, or legal conclusions, and they are not allegations of wrongdoing by any seller, manufacturer, or brand.

• TRUSTWORTHY. Independent sources broadly corroborate the marketing claims; few or no significant red flags surfaced; complaints are minor or rare relative to the volume of evidence. Reading: most readers are likely to consider this a low-risk purchase, in our opinion.
• CAUTION. Marketing claims are partially corroborated, mixed, or context-dependent; some red flags exist but are not severe; significant trade-offs are reported by independent sources. Reading: we think the product warrants extra reading before purchase.
• SUSPICIOUS. Multiple marketing claims are contradicted or uncorroborated by independent sources; recurring complaints exist across sources; the source ecosystem looks unusually thin or skewed for a product of this kind. Reading: in our opinion, the publicly available evidence does not support the strength of the marketing claims.
• AVOID. Independent sources broadly contradict marketing claims, or recurring serious complaints (e.g., safety concerns, persistent malfunctions, undelivered orders) outweigh corroborating evidence. Reading: in our opinion, the publicly available evidence weighs heavily against this purchase decision.

A verdict can change between scans. Evidence accumulates over time, products receive updates, the source ecosystem moves, and our methodology and model evolve. A product labeled SUSPICIOUS today might be labeled CAUTION next month if the evidence shifts. We do not preserve historical verdicts as if they were permanent factual records.

9. Confidence scores and the “insufficient data” flag

Alongside the BS score and verdict, every analysis returns a confidence score from 0 to 100. The confidence score reflects how much independent evidence the model was able to ground its analysis in.

• Insufficient data. When fewer than approximately five independent sources are available to ground the analysis, we flag the result with insufficient_data: true and cap the confidence score at approximately 35. In that state, the verdict and BS score should be read as provisional and weak.
• New product. When a product appears to be too new for an independent review ecosystem to have formed around it, we additionally flag the analysis with is_new_product: true. We surface this to readers as an explicit recommendation to wait for more reviews before relying on the verdict.
• Generic product. When the product page describes a generic, unbranded item (no clear brand to ground in independent reviews), the analysis switches to a value-for-money frame and the verdict reflects category benchmarks rather than brand-specific evidence.

We surface these flags directly in the analysis. We strongly recommend readers treat low-confidence or insufficient-data analyses as a starting point for their own further reading rather than a basis for a confident judgment.

10. The categories we cover

Each product is classified into one of these categories, and the analysis scores it on a category-specific set of dimensions:

• Electronics — build_quality, battery, display, audio, thermals, performance, reliability, upgradability, value.
• Fashion — fabric, fit, sizing, durability, craftsmanship, value, color_accuracy.
• Food & supplements — ingredients, taste, freshness, palatability, side_effects, value, authenticity.
• Furniture — materials, build_quality, assembly, dimensions, durability, value.
• Digital products & courses — content_quality, instructor_credibility, completion_rate, format_quality, update_frequency, refund_policy, value.
• Beauty & personal care — ingredients, effectiveness, irritation, side_effects, skin_comfort, longevity, value.
• Sports & fitness — build, comfort, durability, performance, fit, value.
• Automotive — fitment, reliability, build, durability, value.
• Baby & kids — safety, age_appropriateness, materials, durability, value.
• Health & medical — ingredients, effectiveness, accuracy, side_effects, value.
• Pets — ingredients, palatability, growth_results, side_effects, value.
• Books & media — content_quality, accuracy, engagement, format_quality, value.
• Office & stationery — build_quality, ergonomics, durability, value.
• Garden & outdoor — weather_resistance, build, durability, ease_of_use, value.
• Tools — build_quality, power, precision, sharpness, durability, value.
• Jewelry & watches — materials, craftsmanship, authenticity, durability, value, resale.
• Kitchen & home appliances — build_quality, performance, noise_level, ease_of_cleaning, durability, value.
• Home & bath — materials, water_pressure, leak_resistance, installation, durability, value.
• General / other — quality, usability, durability, value, reliability.

11. What we do not analyze

We refuse to run an analysis on the following categories. These refusals happen before any analysis is generated and before any credit is consumed:

• Tobacco and nicotine products (cigarettes, cigars, vaping, hookah, chew, etc.).
• Alcoholic beverages and distilled spirits.
• Cannabis, cannabis derivatives (THC, CBD oil, hemp extract), kratom, psilocybin, and similar restricted substances.
• Adult/sex-industry products.
• Firearms, ammunition, and weapons.

We may add to or modify this list at any time. The Service is not intended to evaluate restricted products, regulated medical devices, or items where category-specific regulatory expertise is required. Where you encounter such a product, please consult appropriate regulatory or professional sources.

12. Cache and refresh policy

Each analysis is cached for up to 7 days from the time it was generated, keyed by the product URL and the inferred category. Within that window, repeat scans of the same product return the same analysis. After 7 days, the next scan triggers a fresh analysis, and the cache is replaced.

Cached analyses are not permanent records. If a product changes (a new formulation, a recall, a corrected listing), the cache may briefly show stale information until the next scan triggers a refresh. If you believe a cached analysis is materially out-of-date or incorrect, you can request a refresh or a correction through the process described in Section 15.

13. Known limitations

BS Inspector is a useful tool. It is also imperfect. We list its known limitations openly because we think readers and product owners both deserve to know them.

• LLM error. Large language models can misread evidence, mix up entities, hallucinate sources, or extrapolate beyond what their inputs justify. We test for this and design the system to reduce its frequency, but we cannot eliminate it.
• Source ecosystem bias. The set of sources our system can find is biased toward English-language, high-traffic, well-indexed parts of the open web. Products primarily reviewed in less-indexed languages, in private communities, or in offline channels are systematically under-served.
• Time-bound information. Each analysis reflects evidence available at the time it was generated. Reviews accumulate, products are revised, recalls are issued, brands change hands, and our model itself improves. A verdict from one moment may not reflect a fairer reading at another.
• Manipulated evidence. Reviews, forum posts, and aggregated ratings can be manipulated, bought, brigaded, or astroturfed. Our credibility framework reduces but does not eliminate the influence of manipulated sources.
• New products. Genuinely new products may have no independent ecosystem yet. The system flags this, but a thin evidence base can still skew a verdict.
• Generic products. Unbranded or commodity items have no brand-specific evidence to ground in. The verdict in those cases reflects category benchmarks, not the specific item in front of you.
• Page parsing. The system reads a limited excerpt of the product page (about 8,000 characters of text). Pages with unusual structures, content gated behind interactions, or essential information rendered as images may be evaluated on incomplete inputs.
• Translation. Non-English analyses are produced in English and machine-translated. Translation can lose precision, especially around technical terms and qualifying language.
• Category misclassification. A product placed in the wrong category will be evaluated on the wrong dimensions.
• Model and prompt change. Both the model and our system instructions are updated periodically. A scan today may produce a different verdict than the same URL produced six months ago, even if nothing about the product changed.

None of these limitations are unique to BS Inspector; they apply to any AI-assisted product-evaluation system. We surface them so that readers can weight verdicts appropriately and so that product owners can understand the ways in which an analysis might be wrong about their product.

14. How we minimize errors

To reduce avoidable errors we apply the following safeguards:

• Multi-source grounding. A confident verdict requires evidence from multiple independent sources, not a single review or a single page.
• Credibility weighting. Self-published commercial content cannot, on its own, drive a verdict; it is contextualized by independent evidence (Section 6).
• Insufficient-data flag. Below a minimum source threshold, the system declares the analysis provisional rather than guessing with high confidence (Section 9).
• Opinion framing. The system is instructed to use opinion-framing language (“reviews suggest”, “multiple sources report”) rather than bare assertions about the product or its maker.
• Prohibited-category refusal. Categories where category-specific expertise is required for a responsible analysis are refused outright (Section 11).
• Caching with expiry. Verdicts are not permanent: caches expire after 7 days and re-run from a fresh evidence base.
• Public correction process. Product owners and brands can request a correction through the process described in Section 15.

15. Right of reply for product owners and brands

We take the possibility that an analysis may be incorrect seriously. If you are a product owner, manufacturer, brand operator, authorized representative, or any other party with a substantive correction to offer, you can submit a correction request:

bsinspector.com/correction-request

A submission should include:

• The exact product URL the analysis is about.
• The specific factual statements in the analysis you believe are inaccurate.
• The independent evidence supporting your position (links to test reports, regulatory documents, reviews, corrected product information, etc.).
• Your name, organization, and a contact email.
• If you are an authorized representative, the basis of that authorization.

We commit to acknowledge correction requests within 5 business days and to complete an initial review within 15 business days of a complete submission. Outcomes can include: refreshing the cache so that a new analysis is generated against an updated evidence base; updating our system to better recognize the evidence we missed; adding a notice to the analysis; or, when our review confirms our prior analysis, declining to alter it with reasons explained to you.

We do not accept payment to alter, suppress, or favor an analysis, and we do not negotiate verdicts as a commercial matter. We will, however, correct verifiable inaccuracies regardless of who points them out.

Submission of a correction request creates a record in our systems used solely to manage the request and to maintain an auditable history of our editorial decisions. See our Privacy Policy for retention details.

16. Editorial independence and conflicts of interest

BS Inspector earns revenue from end-user subscriptions and credit purchases. We do not earn revenue from product owners, manufacturers, retailers, brand agencies, public-relations firms, or marketing intermediaries in exchange for favorable analyses, suppressed analyses, or any influence over the verdict an analysis produces.

We may participate in standard affiliate programs (e.g., the Amazon Associates program) where outbound clicks to a retailer attribute commission to us. Where we do, the existence of an affiliate relationship is unrelated to and does not influence the analysis, the verdict, or the editorial language used to describe a product. We will disclose material affiliate relationships on this page when they are added.

17. Legal posture

BS Inspector publishes editorial commentary about consumer products. This section describes our position under several relevant legal frameworks. It is not legal advice to any reader.

17.1 United States — opinion, defamation, and trade libel

Under U.S. law, statements of opinion based on disclosed, true, and non-defamatory facts are protected speech. Under Milkovich v. Lorain Journal Co., 497 U.S. 1 (1990), and its progeny, courts evaluate whether a reasonable reader would understand a statement as one of fact or one of opinion, considering the statement’s context, the surrounding language, and the broader publication in which it appears. We publish our analyses with their methodology, evidence base, confidence score, known limitations, and right-of-reply mechanism in plain view, precisely so that readers can recognize them as informed editorial opinion rather than as factual claims about a product or a seller. Where an analysis includes a specific factual statement (e.g., a description of what a named source reported), that statement is presented as a description of public evidence, with the source available for inspection.

We do not assert that any seller, manufacturer, or brand has committed fraud, violated a law, or engaged in misconduct. The verdicts SUSPICIOUS and AVOID are not allegations of unlawful conduct.

For trade-libel and business-disparagement claims, we note that the verdicts and analysis text are conclusions drawn by a disclosed methodology from publicly available evidence, accompanied by a confidence score and a public correction process. We expect such claims to be addressed primarily through our correction process (Section 15).

17.2 European Union — GDPR and the Digital Services Act

For EU and EEA users, we treat each analysis as containing both editorial opinion and a record of automated processing. With respect to Article 22 GDPR (automated decision-making with legal or similarly significant effects), we note that BS Inspector verdicts do not produce legal effects on any reader or product owner; they are editorial information for consumer decision-making. Even so, we publish the logic involved (this page), surface the confidence score in the analysis itself, and provide a meaningful human-in-the-loop pathway through the correction request process (Section 15).

With respect to the Digital Services Act (Regulation (EU) 2022/2065), the correction-request process functions as our notice-and-action mechanism for third parties who consider an analysis inaccurate or unlawful. We aim to apply it consistently and to keep records of the requests we receive and the decisions we make.

17.3 United Kingdom — Defamation Act 2013

For UK users we note the operator-of-website defense under section 5 of the Defamation Act 2013 and our cooperation with the corresponding notice procedure. Our correction-request process accepts and processes complaints in the form contemplated by the regulations.

17.4 Canada — defamation and fair comment

Under Canadian law, the defenses of fair comment and responsible communication on matters of public interest (WIC Radio Ltd. v. Simpson, 2008 SCC 40; Grant v. Torstar Corp., 2009 SCC 61) apply to expressions of opinion on matters of public concern that are based on disclosed facts and that any person could honestly express. Consumer-product evaluation, conducted on publicly available evidence with disclosed methodology, falls within this framework.

17.5 Australia — honest opinion defense

Under Australian uniform defamation law, the honest-opinion defense protects statements of opinion on matters of public interest that are based on proper material. Our methodology, source framework, and disclosed evidence base form the proper material on which the opinions in our analyses are based.

17.6 Other jurisdictions

Comparable doctrines protecting informed editorial opinion on matters of public concern exist in many other jurisdictions, including (without limitation) Germany (Article 5 of the Basic Law and the case law of the Federal Court of Justice on Werturteil and Tatsachenbehauptung), France (freedom of expression under Article 11 of the Declaration of the Rights of Man and Article 10 ECHR), Brazil (Article 5 of the Federal Constitution), Turkey (Article 26 of the Turkish Constitution and Article 10 ECHR), and the Gulf Cooperation Council jurisdictions in which we operate. Where local consumer-protection or media regulations require additional disclosures, we comply with them.

18. Disclaimers

BS Inspector analyses are not:

• professional advice (legal, medical, financial, nutritional, regulatory, or otherwise);
• warranties or guarantees of any product’s performance, safety, or fitness for any purpose;
• regulatory determinations, certifications, or compliance attestations;
• statements of objective scientific fact;
• allegations of fraud, illegality, or misconduct against any party.

The analyses are intended as one informational input alongside your own research, professional advice where appropriate, and authoritative regulatory or scientific sources. To the maximum extent permitted by applicable law, the Service is provided “as is” and “as available”, and our liability for the use of the analyses is governed by our Terms of Service.

19. Updates to this page

We treat this page as a living document. When we change models, change our source-credibility framework in a material way, change our prompt design, add or remove a category, or change our cache or correction policies, we update this page and revise the “Last updated” date at the top. For material changes that would meaningfully affect how a reasonable reader understands a verdict, we will preserve a brief change log here.

20. Contact

Contact: hello@bsinspector.io
Operator: Sawary Trading LLC, a Wyoming Limited Liability Company (USA)