Core concepts

A handful of nouns carry the entire product. Learn these and every screen reads clearly.

The entity model

MentionFlow's data nests top to bottom:

Workspace → Brand → Monitor → Prompt → Run → (Mentions, Citations, Shopping placements)

• Workspace — the account. It holds your plan, billing, members, and referral code, and it owns one or more brands. Quotas (prompts, seats, engine slots) are pooled at this level.
• Brand (also called a project) — the thing you track: a name plus a primary domain. All metrics are computed per brand.
• Monitor — a sampling configuration: one brand across one market/locale, a chosen set of engines, an optional persona, and a set of prompts. A brand always keeps at least one monitor.
• Prompt — a natural-language question you want to watch, e.g. "best CRM for startups". Prompts belong to a brand and can be organised into topic folders and tagged.
• Run — the atomic unit of data: one prompt × one engine × one region × one UTC day. A run stores the verbatim answer text, the raw payload, the model version, and the web searches the engine ran. Runs are idempotent — the same prompt/engine/region/day is only ever collected once.

What a run produces

Each completed run is extracted into:

• Mentions — every brand or product named in the answer, with its order (1-based position of first mention) and a sentiment (positive / neutral / negative, or null when unscored). Sentiment is never guessed: a run whose language model pass failed is marked "unscored", not "neutral".
• Citations — the source links the engine attached, each classified (owned, competitor, editorial, UGC, reference, social, other).
• Shopping placements — for shopping prompts, each product shown in an AI shopping carousel, with its slot position.

Receipt

A receipt is a completed run shown in full: the answer as the engine wrote it, the ranked list of brand mentions, the citations, and any ad or shopping surface. Every metric in MentionFlow traces back to a receipt, so you can always audit a number down to the exact answer that produced it. See Answers and receipts.

Fan-out

A fan-out is a real web-search query an AI assistant ran while answering your prompt, lifted verbatim from the stored payload — never simulated. Fan-outs reveal the searches that actually shaped an answer, and each can be promoted into a tracked prompt in one click. See Fan-outs.

Carousel

In shopping answers, a carousel is one product shelf the engine rendered; a placement is one product in one slot of that carousel. Shopping metrics are computed over carousels, not raw placements. See Shopping overview.

Engines and engine classes

MentionFlow samples ten answer engines: ChatGPT, Perplexity, Gemini, Google AI Overviews (AIO), Google AI Mode, Claude, Copilot, DeepSeek, Grok, and Mistral. They fall into two classes:

• Search-grounded — engines that run live web searches (ChatGPT, Perplexity, Gemini, AIO, AI Mode, Claude, Copilot, Grok). These carry citations.
• Knowledge-only — engines answering from model memory with no web search (DeepSeek, Mistral). Their citations are honestly empty.

This distinction matters for the numbers: headline blended metrics (Visibility, Share of Voice on the Overview) include only search-grounded engines, so the numbers keep one consistent meaning. Some surfaces deliberately widen or narrow that set — the Data honesty page names every place they differ.

Windows

Metrics are computed over a rolling window of days. Two rules make windows trustworthy:

• A window anchors at your newest collected day, not at the wall-clock "today". A collection gap therefore never silently stretches a window.
• Missing days are filled with null, never zero, so a chart never draws a fabricated dip through a day you did not collect.

The dashboard default window is 7 days; you can switch to 14 or 28, or a custom range up to 56 days.

Confidence

A metric computed from fewer than five runs is flagged low confidence. Volume estimates are always shown as bands with a ≈, never as unit-precise counts, because AI-search demand cannot be measured to the unit. See Data honesty.