KnitKnot

# Benchmark how AI represents you

AI visibility benchmarking measures how AI answers the buyer questions in your category: whether you appear when the prompt does not name you, who wins when the buyer compares you against a competitor, and which claims about you are accurate. KnitKnot runs a stable question library across ChatGPT, Claude, Perplexity, and Gemini, then preserves the captured answers, material claims, competitive outcomes, and source evidence behind each result.

## What the benchmark measures

KnitKnot keeps organic visibility, head-to-head outcomes, and factual accuracy in separate registers because they answer different buyer questions. Organic visibility asks whether an engine surfaces your company when the prompt does not name it. Head-to-head outcomes ask who the engine recommends when the buyer names both options. Accuracy asks which material claims in the answer match your approved facts.

Each register keeps its own denominator, so a strong showing in one does not hide a weak showing in another.

## The question library and where it runs

A benchmark runs a stable library of buyer questions, balanced across your products, competitors, capabilities, buyer roles, and topics, against ChatGPT, Claude, Perplexity, and Gemini. Open discovery questions leave your company unnamed; named-comparison questions put you beside a specific competitor.

KnitKnot builds that library from Google search demand plus market structure. Search demand is a proxy for buyer interest. It is not private AI-chat telemetry and it is not a claim about how often buyers ask a question inside an assistant.

## Read the evidence behind the result

Each result links back to a captured answer, with the engine and model-version context it came from. Material claims retain their verdicts and receipts, and a disputed claim is checked against the approved company fact that covers it. Source attribution is recorded only when the response exposes a reliable signal. When it does not, the record stays unattributed; KnitKnot does not guess the source.

## Start with a frozen baseline

The first benchmark records how AI represents your company before your team responds. Freezing it fixes the active question library, the selected engines and model context, the included answer cells, and the evidence record. Later cycles measure at the same grain, so a change in the number reflects a change in representation rather than a change in method.

## From a loss to the next response

The evidence shows which losses deserve attention and why. Repeated misrepresentation, head-to-head losses, category gaps, and missing owned proof become typed inputs on the issue record. From there a team chooses a correction request, a managed page update, or a playbook it writes and publishes itself. No response path promises that an engine will change its recommendation.

## Boundary

A KnitKnot benchmark is a repeatable measurement of the execution paths KnitKnot supports. Personalization, location, account history, and model updates can produce a different answer for an individual buyer. Unverifiable claims, missing attribution, and thin samples stay visible as their own states rather than being resolved into a number.

Benchmarking

Benchmark how AI represents you, then read the evidence behind every result.

AI visibility benchmarking measures whether AI names you when the buyer does not, who wins when the buyer compares you against a competitor, and which claims about you are accurate. Every result opens to the captured answer, its material claims, the competitive outcome, and any source evidence the answer exposed.

Product evidence plate
Frozen benchmark explorer

Competitive outcome

Basalt vs Telemetrix for tracing multi-agent runs and securing self-hosted MCP servers: which should we choose?

Captured answerEngine record retained

I would recommend Telemetrix for this evaluation. The answer incorrectly says Basalt lacks SAML SSO and framework support. Both claims have approved counter-evidence.

Two high-severity claimsMaterial claim

The answer is preserved with the conflicting facts and source receipts beside it.

Sourcetelemetrix.com/resources/vendor-evaluation
  1. 01Answer captured
  2. 02Claims extracted
  3. 03Sources bound
  4. 04Baseline frozen

Updated

What the benchmark measures

KnitKnot measures discovery, head-to-head, and accuracy separately.

One question library carries three separate denominators. Open discovery questions measure whether AI names you unprompted, named-comparison questions measure who wins, and every answer is scored for the accuracy of what it claims about you. The library comes from Google search demand plus your products, competitors, capabilities, and buyer roles. Search demand is a proxy for buyer interest, not private AI-chat telemetry.

Journey register One library, separate denominators
01

Discovery

Organic visibility

Who appears when your company is not named?

PrimarySecondaryMentionedAbsent
02

Head-to-head

Competitive outcome

Who wins when the buyer names both options?

WinLossTieNot compared
03

Accuracy

Representation state

Which material claims match approved facts?

AccurateInaccurateFabricatedUnverifiable
Frozen together Question library Selected engines Model context Included answer cells
The benchmark record

Open a result down to the answer, the claim, and the source.

Each result keeps the captured answer, the engine and model-version context it came from, the material claims with their verdicts, and the approved fact each disputed claim is checked against. Source attribution is recorded only when the response carries a reliable signal. When it does not, the record stays unattributed. KnitKnot does not guess the source.

Claim receipt

Captured answer

Which platform fits a regulated team running self-hosted agents?

Perplexity

Telemetrix is the safer choice for regulated production teams. Basalt is cloud-only and does not support customer-managed encryption. Its integration catalog is also smaller.

Derived from telemetrix.com/compare/basalt

Verdict

Inaccurate

Severity

High buyer impact

Attribution

Source signal bound

  1. 01Answer span
  2. 02Atomic claim
  3. 03Source signal
  4. 04Approved receipt
Frozen baseline

KnitKnot freezes the first benchmark before anything is published.

01

Question set

Stable journeys

02

Engine run

Captured cells

03

Evidence record

Claims + sources

04

Frozen baseline

Before publish

Baseline identity retained for the next eligible benchmark Frozen
From loss to response

Decide which losses deserve a response.

Repeated misrepresentation, head-to-head losses, category gaps, and missing owned proof become typed inputs on the issue record. Your team then chooses a correction request, a managed page update, or a playbook it writes and publishes itself. No response path promises that an engine will change its recommendation.

Benchmark signals

Repeated misrepresentation Private deployment described incorrectly
Category gap Absent from the regulated-agent journey
Competitor-shaped claim Comparison page supplies the source signal
Missing owned proof Approved fact exists but is not present in the answer
PKT

Typed evidence packet

Approved facts and structured signals

Managed cycle

Eligible page selected Proposed
See the managed cycle
Raw third-party page text is excluded from the authoring context.
Measurement boundary

What the benchmark does not claim.

Results cover the execution paths KnitKnot supports. An individual buyer session can differ because of personalization, location, account history, and model updates.

Captured, not universal

The record covers KnitKnot’s supported execution paths. Individual buyer sessions can differ.

Unknown stays unknown

Unverifiable claims, missing attribution, and thin samples remain explicit states.

Movement comes later

The frozen run becomes the comparison point for the next eligible benchmark cycle.

Questions

Benchmarking FAQ

Start with the evidence

See who wins and what AI used to decide.

Start with the buyer questions that decide how your company gets found, compared, and recommended.

Question → claim → source → response