Benchmark how ChatGPT, Claude, Perplexity, and Gemini find, describe, and recommend your company. Inspect organic visibility, head-to-head outcomes, claim accuracy, sentiment, citations, and the captured answers behind the metrics.
KnitKnot separates two questions that conventional mention tracking blends together. Organic discovery measures whether an engine surfaces your company when the prompt does not name it. Competitive evaluation measures whether the engine recommends you, a competitor, or neither when buyers compare vendors.
Each benchmark records AI Presence Score, categorical coverage, organic visibility, head-to-head wins, losses and ties, claim accuracy, sentiment, source ownership, the captured response, and the model version returned by the engine.
Prompts persist across runs and each benchmark snapshots the active library. Keeping that library stable makes periods comparable; editing, adding, archiving, or regenerating prompts changes the measuring set. Brands and individual products can carry separate libraries, competitors, and report slices.
Every aggregate can drill into its scored evaluations, captured answer, claim ledger, and citations. Reports can be pinned to a measurement period and published to a read-only URL.
KnitKnot executes controlled benchmark requests against supported engines. It does not promise that a captured result is identical to every buyer's personalized consumer session; account history, geography, model changes, and interface differences can affect outputs.
KnitKnot supports ChatGPT, Claude, Perplexity, and Gemini. You can run all four or choose a subset, and each captured response keeps its engine and model-version record.
Visibility is the share of organic evaluations where your company appears at all. The AI Presence Score is a broader 0-100 composite that combines category-specific signals such as exposure, competitive outcomes, claim accuracy, sentiment, and source balance.
For head-to-head prompts, KnitKnot classifies each decided evaluation as a win, loss, or tie. Win rate is wins divided by wins, losses, and ties; responses that make no comparison are excluded, and ties remain in the denominator.
Yes, when the active prompt library stays stable. Every run snapshots the current library, so deliberate edits, additions, or archives should be treated as changes to the measuring set.
Yes. A brand and each product can have separate prompts, competitors, and result slices, while product recommendations still roll up correctly to the parent brand.
KnitKnot executes benchmark prompts against supported engines and stores the returned responses. Personalization, account history, geography, model updates, and engine interfaces can produce different answers, so the benchmark is a controlled measurement rather than a promise of identical consumer output.
Raw mirror of this content: https://knitknot.ai/product/ai-benchmarks.md. Site-wide summary: /llms.txt · full content: /llms-full.txt
Run a stable library of buyer questions across ChatGPT, Claude, Perplexity, and Gemini. Measure organic discovery and competitive selection separately, then inspect the captured answer behind every result.
Weekly benchmark
Jul 6-12 / 120 prompts
A vendor can appear in every category answer and still lose every comparison. KnitKnot keeps discovery and evaluation separate so the metric tells you which problem you actually have.
Organic discovery
Open category and use-case questions measure coverage, rank, and visibility without naming your company. Competitive prompts are excluded from this visibility rate because a forced mention is not organic presence.
Competitive evaluation
Head-to-head questions measure explicit recommendations and feature verdicts. Each decided answer becomes a win, loss, or tie; responses that make no comparison remain outside the record.
Freeze the active prompt set for the brand or product.
Run each question against the selected supported engines.
Store the returned answer, citations, engine, and model version.
Extract claims and derive evaluation and feature outcomes.
Build metrics and reports from the same canonical rows.
Each metric answers a narrower question. Keeping their denominators explicit prevents a favorable headline from hiding a competitive or factual weakness.
| Metric | What it measures | Important boundary |
|---|---|---|
| Coverage | How central you are to one answer: primary through absent. | Categorical per response, not a percentage. |
| Visibility | Share of organic scored answers where coverage is not absent. | Prompts that name you are excluded. |
| Win rate | Wins divided by wins, losses, and ties in decided head-to-head answers. | Ties stay in the denominator; not-compared answers do not. |
| Accuracy | The factual record across claims the validator could check. | Unverifiable claims remain visible but are excluded. |
| AI Presence Score | A 0-100 category-specific composite of exposure or outcomes, accuracy, sentiment, and sources. | It is a composite, not an average of the other headline metrics. |
Each engine is executed and scored independently. Open the exact response to see which model found you, which competitor it selected, what it cited, and which claims changed the result.
Captured benchmark response and cited-source record
Anthropic response with live web-search sources
Search-grounded answer and citation composition
Captured answer with engine-specific source signals
Scores, reports, and drill-downs read the same persisted evaluation facts and feature verdicts. There is no second reporting formula that can silently reinterpret the result.
Captured response. Read the returned answer with its engine, timestamp, and model-version record.
Claim and source evidence. Inspect anchored claims, factual verdicts, severity, citations, and attribution when the answer exposes a reliable signal.
Period-pinned reports. Publish a read-only benchmark overview or head-to-head brief tied to a measurement period.
KnitKnot stores the responses returned by its supported execution paths. Personalization, location, account history, model updates, and consumer-interface changes can make an individual buyer's answer differ. Use the benchmark as a repeatable market measurement and read movement across periods, engines, and evidence.
Start with a benchmark across ChatGPT, Claude, Perplexity, and Gemini, then read the evidence behind the result.
