Core terms

AI Presence Management

The practice of benchmarking, monitoring, and improving how AI models represent your company when buyers ask evaluation questions. Covers factual accuracy, competitive positioning, recommendation rates, source influence, and sentiment across AI platforms. Full guide.

AI Presence Score

A 0-100 composite metric quantifying how well AI models represent your company, written at scoring time for every evaluated response. Built from seven structured signals: recommendation, feature comparisons, claim accuracy, sentiment, source influence, coverage depth, and confidence weighting. The decomposition into components is what makes it actionable. Methodology.

AEO (Answer Engine Optimization)

The discipline focused on getting your content cited by AI answer engines. Measures mention frequency, share of voice, citation count, and sentiment. AEO tells you whether you appear in AI responses. AI Presence Management tells you whether what AI says is accurate. AEO vs APM.

GEO (Generative Engine Optimization)

The practice of structuring content so AI models can extract and cite it effectively. Focuses on the input side: page structure, schema markup, answer placement, content format. Originated from a Princeton study (ACM KDD 2024) that quantified which content attributes improve AI citation rates.

Zero-click search

A search where the user gets the answer without clicking through to any website. 58.5% of Google searches are zero-click. AI chat interactions are nearly always zero-click by design: the buyer reads the answer and never visits a vendor site.

Measurement loop

The core operating cycle of AI Presence Management: benchmark how AI represents you, diagnose which sources and gaps produced each answer, fix the sources, then re-measure and compare against the previous run. A persistent prompt library makes the loop repeatable: the same prompts run each time, so deltas in mentions, citations, and scores are attributable to the changes you shipped rather than to a shifting question set.

MCP server

A server exposing a product’s data and actions as tools that AI assistants can call directly via the Model Context Protocol. KnitKnot’s MCP server (mcp.knitknot.ai) exposes roughly 40 tools, so customers can connect Claude, ChatGPT, or any MCP client to their workspace and run benchmarks, query score trends, pull competitive overviews, and manage the prompt library from inside the assistant itself. Your AI presence data becomes something your AI agents can monitor and act on.

Scoring and metric terms

Coverage

How prominently a company appears in a single AI response. A categorical rating assigned per evaluation: primary, substantial, peripheral, incidental, or absent. Coverage is a per-response label, not a rate; the rate built from it is the visibility rate.

Visibility rate

The share of scored responses where coverage is not absent, expressed as a fraction. Computed organic-only: prompts that name the company directly are excluded, so the rate measures whether AI brings you up unprompted in category and comparison questions.

Sentiment

How favorably the AI’s response treats the company, scored 0-100 per response and aggregated on the same scale. Distinct from accuracy: a response can be warm and wrong, or cold and correct.

Head-to-head outcome

The result of a competitive AI evaluation: we_win (the AI recommended you), competitor_wins (it recommended them), tie (it declined to choose), or not_compared (no direct comparison was made). Outcomes are derived deterministically from the judge’s structured signals, not from a single holistic “who won” judgment. Our published benchmark corpus covers 11,600 head-to-head evaluations across 136 competitors.

Win rate (W-L-T)

Wins divided by decided outcomes, where decided outcomes include ties in the denominator. Wins, losses, and ties are tallied separately; not_compared evaluations are excluded entirely. A stricter variant, the decisive win rate (wins divided by wins plus losses, ties excluded), appears in our engine divergence analysis, where the aggregate across engines was 70.7%.

Claim accuracy

The fraction of the AI’s factual claims about your company that are correct. Misrepresentations are severity-weighted: a critical error about core positioning counts five times more than a minor tone issue. Scoring methodology.

Confidence weighting

An adjustment to claim accuracy based on how confidently the AI stated a false claim. False claims stated with certainty carry a 1.3x penalty. Hedged claims carry a 0.9x discount. Each response’s hedging language is classified as certain, tentative, or uncertain. Confident lies.

Misrepresentation

A factual claim an AI model makes about your company that contradicts the verified record. In KnitKnot reports, each misrepresentation carries a proof receipt: the exact knowledge-base source that contradicts the AI’s claim, shown alongside the claim itself. The receipt turns “the AI is wrong” from an assertion into an auditable finding.

Persona

The buyer role an AI response is implicitly written for, inferred per response and canonicalized to six standard roles. Persona matters because the same comparison question gets a different answer for a compliance lead than for a startup founder, and benchmark coverage should span both.

Source and citation terms

Citation

A URL or source that an AI model references in its response. Not all models display citations visibly. Perplexity and Claude show sources. ChatGPT sometimes does. The number of citations matters less than which citations shaped the recommendation (source gravity).

Source balance

The ratio of your cited sources to competitor cited sources in an AI response. A source balance of 0.2 means the AI built its answer primarily from competitor content. Source influence.

Source gravity

A measure of how much influence a cited source had over the substance of an AI response. High-gravity sources shaped the recommendation and framing. Low-gravity sources appeared as footnotes. Source ownership model.

Owned sources

The classification of every cited domain as yours, a competitor’s, or third-party. KnitKnot tracks ownership per workspace with subdomain matching, and re-classifies past citations when ownership changes, so source balance and source-leak analysis stay consistent over time.

E-E-A-T

Experience, Expertise, Authoritativeness, Trustworthiness. Google’s quality framework, increasingly used by AI models to determine source credibility. Named authors, visible dates, inline citations, and brand mentions are the key signals.

Engine-specific terms

Bing index

The search index ChatGPT uses for web-connected queries. Submitting your sitemap to Bing Webmaster Tools directly affects ChatGPT’s access to your content.

Brave Search

The search engine Claude uses for web lookups, publicly documented by Anthropic. Different index from Bing and Google, which means the same content may rank differently per engine. How we approximate the Claude engine.

AI Overviews

Google’s AI-generated answers that appear at the top of search results, synthesized from crawled pages with citation links. Visibility in AI Overviews depends on extractable, directly-answering content rather than traditional ranking alone.

AI Mode

Google’s dedicated AI chat interface (limited rollout). Nearly all AI Mode queries are answered within the AI response without the user visiting any website, making it an almost entirely zero-click surface.

Benchmark terms

Adversarial benchmarking

Testing AI models with the hardest questions buyers actually ask, not just brand queries. “What are the drawbacks of [your product]?” “Why would I choose [competitor]?” These prompts surface errors that standard monitoring misses. Prompt methodology.

Prompt library

The persistent set of benchmark prompts a company is evaluated against. Prompts are generated keyword-first from real Google search data (with search volume attached per prompt), layered across features and buyer personas, and persist across runs: each benchmark run snapshots the library, so results are comparable run over run. A brand and each of its product lines maintain separate libraries.

Spot test

A single benchmark prompt run on demand, outside a full benchmark run. Useful for quickly checking how an engine answers one specific question after a content change, before committing to a full re-measure.

Grounding tier

The evidence level for why a benchmark prompt was generated. Direct: maps to a real Google query with measurable search volume. Category: no direct query but category-level searches exist. Synthesized: AI-generated to test an uncovered edge.

Cross-model disagreement

The rate at which different AI models produce different competitive outcomes for the same prompt. In our data: 48.6%. Engine divergence.

Absence rate

The fraction of brand perception evaluations where the AI doesn’t mention the company at all. Overall: 26.6%. Per engine: Perplexity 38.8%, ChatGPT 26.3%, Claude 23.9%, Gemini 17.7%.