> ## Documentation Index
> Fetch the complete documentation index at: /docs/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Assistants connect to Parse over MCP at [https://agent.parse.gl/mcp](https://agent.parse.gl/mcp) (Streamable HTTP). With no sign-in it answers from Parse's public view (the numbers parse.gl's public pages show, up to 5 rows a list), starting with `check_brand`. Signed in with Parse, or sent a workspace key (`parse_mcp_…`), it works in the person's workspace, starting with `get_brief`. Scripts call the same tools over REST at `https://agent.parse.gl/v1/tools/<tool>`.
>
> ```bash
> npx parse-gl init
> ```
>
> To install just the skills: `npx skills add Parsegl/skills`.
> Every tool is also a POST to https://agent.parse.gl/v1/tools/<tool>; its OpenAPI description is https://agent.parse.gl/openapi.json.

# Reading the numbers

> What each measure the agent service returns counts, how to read a range and a significant change, what window and engines a number covers, and why the app can differ.

Every number is a count or a share of AI answers to the questions Parse asked. It is not a share of buyers, traffic or sales. The app's own numbers and the public pages' are explained on [Methodology](https://parse.gl/metrics); this page is for the numbers `query_metrics` and the other tools return.

## Measures

| Measure | What it is |
| - | - |
| `answers` | Answers in the cell: the denominator of every share. |
| `share_named` | Mentioned in, the number Parse's public pages, rankings and the app lead with: share of answers naming the brand. |
| `share_recommended` | Recommended, a diagnostic breakdown: the share of answers recommending the brand (recommended, offered as an alternative, or listed with a price or spec as an option). Answers naming the brand whose judgment is still pending count in no share, so its own `answers` can be fewer than the cell's. |
| `answers_naming` | Answers naming the brand, its family counted as one. Read it next to answers. |
| `share_first` | Share naming it first. |
| `share_spread` | How much the share depends on the wording of the question: the lowest and highest share over questions with enough answers. |
| `avg_position` | Average position among the brands an answer names, when the brand is named. |
| `tone_mix`, `treatment_mix` | How answers that name the brand speak of it. |
| `citations`, `pages_cited`, `sites_cited` | Citations, and the distinct pages and sites cited. |
| `comparisons_won`, `comparisons_lost` | Head-to-head comparisons the brand comes out ahead in, or behind in. |
| `concentration` | The share of answers naming the top 1, 3 and 5 brands, and how many brands make up half of all mentions. It has no range. |
| `search_reach` | Of the questions with a retrieval list from a search API, the share where a surface retrieved a page on one of the brand's own sites. |
| `memory_gap` | The share of answers naming the brand with search on, against search off, on the questions both answered. |
| `listing_share` | For answers that list local businesses, the share whose list includes the brand, and its mean rank there. See [Agencies and local businesses](/docs/agencies.md#places). |

[`resolve_ids`](/docs/mcp/tools/resolve_ids) with `kinds: ["measure"]` serves each measure's definition and the evidence behind it, and the facets to filter and group by.

## Ranges and "not significant"

* Every share comes with `range95`, the margin either side of the value. The answers to one question count as one cluster, so ten answers to one question are not ten pieces of evidence.
* Under 20 answers a share is weak. The free level withholds it and returns `below_floor`.
* `compare: "prior"` compares a window with the one of equal length before it, on the questions asked in both (`matchedQuestions`).
* `significant: true` needs three things: at least 20 answers in each window, at least 5 matched questions, and a change outside `changeRange95`. Report only significant changes; the reply's `context.significance` says the rule.
* `composition` is how much of the change comes from which questions were asked, not from how the answers changed.
* `explain: "change"`, with a subject and `compare: "prior"`, splits a change into same-question change and question mix, and lists engine model changes by date.

## Windows

* The default is the last 28 days.
* `window.last` takes `7d`, `28d`, `90d`, `365d` or `all`; or give `from` and `to` as YYYY-MM-DD, in UTC.
* `by: ["week"]` or `by: ["month"]` gives a trend on matched questions.
* The last 90 days, ending on the snapshot's last day, are free. A window that reaches before them reads past answers: a public answer from then costs \$0.001 the first time the workspace counts it, and is the workspace's after; its own answers are never charged. The charge is in `meta.costCredits` ([Spending](/docs/credits.md#paid-calls)).

| history mode | What a call older than 90 days does |
| - | - |
| `"auto"` | Default. Reads the deepest even sample worth at most 25% of the active monthly budget (capped at \$5) and reuses samples already bought. |
| `"all"` | Counts every matching past answer within your monthly budget (at \$0.001 each once for answers not yet owned). |
| `"free"` | Buys nothing. Reads only the last 90 days and returns `context.history.rungs` estimating fuller counts. |

* `context.sample` tells you when a past read was sampled (for example, "1 in 4"). Use answer-level shares and their 95% ranges as reported. Counts describe the sample. Question-level measures such as `search_reach` and `share_spread` favor questions asked more often, because those questions are more likely to appear in the sample.
* `context.history` reports `freeFrom`, `pastAnswers`, `newToWorkspace`, `chargedUsd`, `moreToCountAllUsd`, and `budgetLeftUsd`; `next` carries the ready-to-run `history: "all"` call.
* If a workspace cannot buy past answers (no credit, spending paused, or logged out), the call reads the last 90 days plus any past answers already owned and notes approximate omitted past rows under `limits: [{ what: "history_unpaid" }]`.
* Assistants get no single weekly "mentioned in" figure: it moves when the questions do.

## What is counted

* ChatGPT and Google AI Mode, unless `filter.engine` or `filter.surfaceClass` says otherwise, or `by` names `engine` or `surface_class`. `context.surfaces` names what was counted.
* A `subject` with no filter reads the brand's own markets (up to three), and `context.scope` names them. `filter: {}` reads every answer.
* `level: "product"` measures one product apart from its company.
* `filter.place` is the place a question is about; `filter.country` is where it was asked from. See [Agencies and local businesses](/docs/agencies.md#places).

`filter.scope` says whose answers:

| Value | Reads |
| - | - |
| `all` | The default: public answers plus the workspace's own. |
| `own` | Only the answers the workspace bought. |
| `tracked` | Its own answers, its workspace's saved brands, and those about the questions, topics and brands it watches with track. |
| `public` | Public answers only. |

## Levels

Every data reply says in `meta.level` whether it read in full or at the free level, and in `meta.levelReason` why.

* **The free level**, with no key or no credits: samples of up to 5 rows with no paging, shares only over 20 or more answers, and excerpts in place of full texts.
* **The workspace's own brands** are read in full with or without credits.
* **Deep paging and exports** need a first card top-up: a search that names no brand, market, topic, question, page or site pages no further than its first page, and [`export`](/docs/mcp/tools/export) writes nothing, until then.

## Missing is not zero

`{ "missing": "<reason>" }` (for example `below_floor`, `not_collected` or `withheld`) means Parse has no value there. Say why, or run the call that fills it; `next` names it when there is one.

## When Parse's data changed

`context.dataChanges` lists the days a backfill moved counts without any change in the AI answers themselves, with what was returned. A count over those months can differ from a reading before the date. Dated entries are on the [Changelog](/docs/changelog.md).

## Why the app can differ

Mentioned in uses the same extraction rule across surfaces. The answer sets differ. The app's weekly tables count mentions the agent service does not, and the Overview reads answer packs built from answers with usable extraction. The agent service reads collected answers in the requested scope and window. The Overview reads the last 28 days and the public rank counts answers since May 24, 2026. The app prints a change when it passes a pooled test at 85% confidence. The agent service reports 95% ranges on matched questions, so fewer changes read as significant.

`context.measures.differsFromDashboard` says this in every reply. When a person's app figure and the assistant's disagree, neither is wrong: they count differently. Tell the person which one you are quoting.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.