How This Site Works
The Tennis Index finds and organizes video and written reviews for tennis racquets, then builds structured summaries from what reviewers actually said. The goal is straightforward: surface genuine consensus, report real disagreement, and provide an unbiased composite review. This page explains how.
How Sources Get Here
Every racquet page should show you the most relevant reviews for that model — not every video that mentions it or every article that references it. The system tries to strike a balance: comprehensive enough that valuable coverage isn’t missed, selective enough that the page isn’t cluttered with tangential content. Video and written sources arrive through automated discovery and go through classification and matching before they appear.
Continuous monitoring
New videos are discovered through targeted YouTube searches run daily across every racquet model in the catalog. Written editorial sources are discovered through RSS feeds, sitemaps, and index page polling from a registry of independent tennis publications. Searches are structured around brands, model lines, and generations, not channels. Coverage isn’t limited to a preset list of creators.
Content classification
Every discovered source is classified by format and subject matter before it can be matched to a racquet. For videos, the system distinguishes racquet reviews and playtests from instruction videos, string reviews, vlogs, podcasts, top lists, and other content types. For written sources, editorial reviews are separated from product listings, spec pages, and promotional content. Only sources focused on firsthand racquet evaluation proceed to model matching.
Channel trust
There is no curated whitelist. Any channel can appear on the site. Channels earn trusted status after multiple independently verified matches across different racquet models, which gets their new content processed faster. The accuracy standard is the same regardless.
Model matching
Each video is matched to a specific racquet model and generation through a multi-step pipeline that reads the title, description, and full transcript. Generation resolution prioritizes explicit year references and transcript evidence, falling back to publish-date inference when needed. Temporal validation rejects impossible matches. The full decision trail is stored.
Edge case review
Most videos are matched and approved automatically. When the matching pipeline can’t reach a confident result (unusual titles, ambiguous model references, atypical content formats), a human reviews the match before it goes live. The system handles the volume; a person handles the exceptions.
If you’re a creator whose work appears on this site, we have a page for you.
How Summaries Are Built
Once a racquet has enough review coverage from multiple sources, the site can generate a summary. Sources include video transcripts and written editorials. Summaries are not written by hand and are not editorial opinions. They are synthesized from structured analysis of what reviewers said, with guardrails at every step to keep the output grounded in real consensus.
The pipeline separates signal from noise: extract specific claims, verify their basis, count cross-source agreement, then generate prose. The same process runs for every model, with no manual intervention.
Extract
Each source — whether a video transcript or written review — is processed in full and broken into discrete claims about playing characteristics: power, spin, feel, control, stability. Every claim gets tagged with its evidential basis (firsthand hitting experience, spec-derived inference, or secondhand reporting). Claims about previous generations, competitors, or ambiguous targets are identified and set aside.
Filter
Low-confidence claims, unverified secondhand reports, and spec-derived inferences about subjective characteristics get filtered before they can influence the summary. Claims about feel are held to a stricter standard: only firsthand hitting experience qualifies.
Group
Multiple videos from the same channel count as a single source, so no one voice can inflate the consensus. When a creator has both a video review and a written review of the same racquet, those also count as one source. Claims are deduplicated across sources and grouped by how many distinct reviewers made the same observation. The number of reviewers backing each claim is the primary signal that drives the summary.
Synthesize
A large language model generates the summary under strict constraints: it can only reference claims present in the extracted evidence, must attribute observations by source count, and cannot editorialize, recommend, or use superlatives. Summaries with fewer sources are kept deliberately short; models with richer coverage earn longer, more detailed treatment. Only the characteristics reviewers actually engaged with are covered: where they reached agreement, split into camps, or showed a notable spread of impressions. An aspect reviewers did not meaningfully address is left out rather than filled in, so a thinly-reviewed racquet yields a shorter summary by design. Majority views are stated with confidence. When a meaningful share of sources contradict the majority on a given characteristic, the summary reports both positions. The voice is wire-service neutral: factual, third-person, precise.
Source Categories
Summaries draw from video reviews and written editorials. All sources contribute equally to claim counts once their claims pass extraction — there is no weighting by source type. When a creator or publication has both a video and a written review of the same racquet, they count as one source, not two.
How We Guard Against Bad Data
Aggregating reviews is only useful if the inputs are trustworthy. The pipeline applies several layers of filtering designed to catch promotional bias, generation confusion, and inflated consensus before they reach the summary.
Content screening
A classifier identifies the format and subject matter of each video before any analysis begins. Non-review content is filtered out, and only videos focused on firsthand racquet evaluation proceed to extraction.
Promotional bias filtering
Tour professionals discussing their own sponsor’s racquet have their claims excluded from extraction. The videos still appear on the site, but the structural conflict of interest means those claims cannot contribute to summaries.
Generation verification
Explicit year references, transcript evidence, and publish-date context are all used to verify that each claim refers to the correct racquet generation. If the generation can’t be confidently determined, the claim is excluded rather than guessed.
Independence counting
Multiple reviews from the same channel or author count as a single source. No one voice can inflate the consensus.
Minority view preservation
When enough sources disagree with the majority on a characteristic, the summary reports both sides rather than defaulting to the majority.
Feel claim strictness
Subjective characteristics like feel, comfort, and feedback are only accepted from firsthand sources. Spec-derived inferences about how a racquet should feel are excluded.
The Consensus Indicator
Every summary carries an indicator reflecting the strength of agreement across sources. It’s computed from the claim data, not editorially assigned.
Reviewers broadly aligned on the same playing characteristics. The summary reflects a clear, well-supported picture.
Reviewers agreed on core characteristics but differed on some aspects. The summary says where they align and where they diverge.
Reviewers reported meaningfully different experiences. The summary reflects the range of opinions rather than forcing agreement.
Reviews are still arriving. These summaries are based on limited coverage and update as new sources are added.
The Minimum Bar
Models without enough coverage don’t get a summary. The minimum requirements exist so that every summary on the site reflects real consensus, not isolated opinion. Sources that only relay specifications or secondhand reports don’t qualify.
Source Detail
The source detail panel beneath each summary shows the full evidence chain: every extracted claim, how it was classified, which sources supported it, and what was filtered out. Claims about previous generations or competitors are shown separately, so you can see what was excluded and why.
This methodology is under active development. The version tracks the summary pipeline. As we refine how reviews are found, filtered, and summarized, we record what changed.
Changelog