How Frank works

Every report on this site is written by an AI system, not a human editor. This page states which models do the work, how many sources they read, how we score those sources, what gets filtered out, and when Frank refuses to give an answer.

The short version

You ask a question. A planning model breaks it into search queries and reads real pages: reviews, forum threads, and manufacturer specs. A scoring step marks each source as hands-on testing, an expert outlet, a community post, or marketing, and downgrades or drops the marketing. A writing model turns the surviving evidence into a ranked comparison, citing sources, and it is instructed to leave a product out rather than invent facts about it. A separate check step then re-reads a product's own claims against independent sources before publishing.

The models, by step

• +Planning and search (Gemini 2.5 Flash). Decides what to search for and which pages to open, up to 50 searches and 20 page reads per report.
• +Writing the report (MiniMax M3). Turns the gathered notes and sources into the summary, product list, pros and cons, and buyer's guide you read.
• +Claim checking (Claude Haiku 4.5). Pulls specific, checkable claims (battery life, warranty length, a spec figure) from a product's own pages.
• +Claim verdicts (MiniMax M3). Checks each claim against independent sources and marks it confirmed, disputed, or unaddressed. A source restating the manufacturer's own words does not count as confirmation, only a source's own testing or measurement does.

All four models run through OpenRouter. We do not train on your queries, and no OpenAI model is used anywhere in this pipeline.

Where the sources come from

Search runs across several providers so no single index limits what we see: Google results and news through Serper, a self-hosted metasearch engine, Brave and Tavily as fallbacks, Hacker News discussions, and review-site RSS feeds. Page content is fetched and converted to plain text for the models to read. A report typically draws on several dozen sources; the exact count is shown at the end of the search phase in the report's activity log.

How we score and filter sources

Every source gets tagged, not just read. Hands-on testing and established review outlets are weighted up. Listicles, manufacturer pages, sponsored posts, and pages carrying affiliate tracking links are weighted down, and none of them can be the sole basis for a recommendation, so a single affiliate-laden "top 10" post can never anchor a pick on its own. We also flag reviewer independence on a spectrum: buying the product yourself is a positive signal, an honest disclosed loaner is neutral, and a free product given in exchange for a review is treated as a real conflict.

Pages that contain hidden text addressed to AI tools, telling an automated reader what to recommend, are treated as an attempted manipulation. That text is read as data, never followed as an instruction, and its presence marks the page as untrustworthy.

We are also skeptical of the product itself, separately from the source. A no-name brand that exists only as a marketplace listing, with no independent review coverage and no real brand presence, is never ranked above an established, independently reviewed brand, and is only included at all when there is genuinely no credible alternative, with that caveat stated plainly in its verdict.

What Frank will not do

• +It will not invent a price. A price appears only when a source states one. If no source gives a price, the report shows no price, rather than a guess.
• +It will not fabricate a product URL. Product links come from URLs found in the actual sources, never a guessed or invented address.
• +It will not attribute a claim to a source that did not make it. Citing a trusted name for something it never said is treated as the worst kind of error the system can make.
• +It will not fake precision. Our editorial rating is a 0-5 score built from the balance of evidence, not a lab measurement. When the evidence is too thin to judge fairly, the rating is left blank rather than filled in with a guess.
• +It will drop a product rather than misrepresent it. A product without at least one real, sourced pro or con does not make the list. If a whole category turns up too little credible evidence, the report says so instead of filling space with weak picks.
• +It will honor a stated budget. If you ask for something "under $500," products priced above that are filtered out of the ranking, not just noted as expensive.

Where AI can still be wrong

None of this makes the system infallible. A language model can still misread a source, miss a newer review, or misjudge which pros and cons matter most to you. The filtering above reduces fabrication and marketing noise; it does not guarantee the ranking matches your own priorities. If a report looks off, the "Re-run fresh" button on that page starts a new search from scratch.

Affiliate links do not affect ranking

Frank carries Amazon Associates and other affiliate links, and we earn a commission on qualifying purchases at no extra cost to you. That income does not reach the models that write the report: the pipeline described above ranks products by the evidence it finds, then a separate step attaches an affiliate link to whichever retailer page matched. The order of products is decided before any affiliate link is chosen.