Glossary

The practice of making your brand's facts findable, parseable, and trustworthy enough that AI models will surface them accurately when users ask. Different from SEO in that the unit of optimization is the fact, not the page. See AEO 101 for the full treatment.

Google's synthesized answer at the top of search results, generated by Gemini. Appears on roughly 47% of commercial queries as of April 2026. Functionally the largest answer engine in production today, and the reason organic traffic to information sites has dropped 18-42% in the same period.

In misquoted's measurement model: the fraction of model-asserted facts about your brand that match the ground truth (typically sourced from your primary domain). Distinct from consensus, which measures whether models agree — they can agree on something wrong.

A source URL that an AI model attributes a claim to. Perplexity cites by default; ChatGPT and Gemini cite selectively. Citation rate (how often your URL is among the citations) is one of the four core AEO metrics.

A discrete factual assertion made by a model about a brand — e.g. "founded in 2018," "pricing starts at $49/mo." The atomic unit of accuracy measurement. A single answer may contain 1-12 claims.

misquoted's headline metric. The fraction of brand questions where three independent models give substantively equivalent answers. Scored 0-100. Above 70 is "Verified," 40-69 is "Mixed," below 40 is "Critical."

When a model invents a plausible-sounding fact rather than admit uncertainty. Distinct from hallucination in that confabulations tend to be structurally coherent — wrong dates, invented customers, fabricated press quotes. The harder failure to detect.

The gradual divergence between a brand's ground truth and what AI models say about it. Caused by model retraining, ecosystem reorganization, or the brand itself changing. Detected by re-scanning; Monitor Pro catches drift within a week.

Google's pre-AI synthesized answer box, shown above the ranked results. Predecessor to AI Overview. Optimizing for featured snippets is roughly 60% of the work toward optimizing for AI Overview.

Synonym for AEO, slightly more common in academic literature. We use AEO because "answer engine" is the more precise product category. Same idea.

The practice of constraining a model's answer to a specific set of retrieved sources. Perplexity is heavily grounded; ChatGPT is selectively grounded; old-school LLMs without retrieval are ungrounded. Grounded models are more accurate but slower.

A model output that is detached from input or training data — typically nonsensical, structurally weird, or factually impossible. Less common in modern models than confabulation, which is the subtler and more dangerous failure mode.

A structured-data format for embedding machine-readable facts in HTML. The dominant schema delivery format. Embed in a <script type="application/ld+json"> tag in the document head. The single highest-leverage AEO ship for most sites.

Google's right-rail entity card showing brand essentials — logo, founding date, leadership, key links. Sourced from Wikidata, Google Business Profile, and schema markup. A high-trust signal for AEO.

A statistical measure of inter-rater agreement, adapted by misquoted to measure cross-model agreement. Adjusts for chance agreement; a kappa above 0.6 indicates substantial agreement. Underlies our Consensus Score.

A 2024 proposal for a markdown file at /llms.txt that summarizes a site for AI crawlers. Loosely analogous to robots.txt — but where robots.txt restricts, llms.txt instructs. Increasingly adopted by AI-aware tooling.

Large Language Model. The underlying technology for ChatGPT, Gemini, Claude, and others. We track three (ChatGPT, Gemini, Perplexity) as the minimum diverse set that captures 94% of consensus signal.

A company shipping an LLM-backed product (OpenAI, Google, Anthropic, Perplexity). Distinct from the model itself (GPT-4o, Gemini 1.5 Pro). misquoted scans three providers as standard; Monitor Pro includes a watchlist of three more.

The input given to an LLM. In misquoted's context, the standardized brand questions used across all three models — pricing, products, founding, competitors, support, reputation. Held constant for comparability.

Per-model accuracy rating — how often a given provider gets your facts right. Calculated independently of the consensus score. Useful for diagnosing which model is misrepresenting you.

The standardized list of brand questions used in a scan. Free scans use 6 questions; $49 reports use 30; $199 reports use 50-80 via dynamic generation. Held constant across models within a scan.

Retrieval-Augmented Generation. The architecture where a model retrieves relevant documents before generating an answer. The mechanism behind grounded LLMs. All three of our tracked providers use RAG in some form.

How often a model declines to answer a question about your brand — typically with "I don't know" or "I can't find that information." Perplexity refuses at 18% on pricing questions. High refusal rate is a signal that your facts aren't on the web in machine-legible form.

The shared vocabulary for structured data on the web, maintained by Google, Microsoft, Yahoo, and Yandex. Defines types like Organization, Product, Article. Delivered in HTML primarily via JSON-LD.

Google's earlier name for what became AI Overview. Term has fallen out of use as of 2025; we list it for completeness.

How damaging a given inaccuracy is, weighted by query frequency and conversion impact. Pricing wrongness is high severity; founding-date wrongness is medium. Available in $49+ tiers.

The atomic unit of LLM input/output. Roughly equal to 0.75 of an English word. Models are priced per million tokens; scan costs are dominated by token counts. Useful only as a cost-modeling concept for end users.

misquoted's three-tier classification for sources: Authoritative (the brand's own domain), Confirmed (third-party press, Wikidata), Disputed (forum posts, secondary blogs). Used to weight evidence in the accuracy scoring engine.