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

# Deep reads

> Four read-only tools that return the stored answers, citation days and page versions your metrics were counted from.

Every number Attensira shows is a fold over stored readings. The deep reads hand back the rows themselves, so a number can be checked rather than trusted. All four are free, read-only, and scoped to the workspace the credential names. They are available as agent tools and as [MCP tools](/mcp/tools); there is no REST endpoint for them yet.

## answer\_history

Every sampled answer to one tracked prompt, verbatim, newest first, each with its platform, surface (consumer or api), slot day, whether it named you, its citations in order, and the reading id. `prompt` is required; `platform` and `days` (1 to 365, default 28) narrow it.

* An answer longer than **8 KB** is cut on a character boundary and carries `truncated: true`, with `text_bytes` giving the full stored length.
* A page holds at most **50 readings** (default 10), each with all of its samples. Pass `next_cursor` back as `cursor` for the next page; `has_more` says whether there is one.
* A failed sample is returned with `failed: true` and the customer-facing reason; it has no text and never counts toward n.

## answer\_slice

Every answer matching a predicate: `platforms`, `country`, `surface`, `competitor_named` (a tracked rival's domain, named or linked), `cited_domain` (among the citations), `brand_named` (true or false), and `from`/`to` days or `days`. Terms AND together and select the **sample**, not the reading: "consumer answers that named Jasper and cited g2.com" returns exactly those samples. At least one term is required; for one prompt's answers use `answer_history`. Same 8 KB cut and 50-reading page as `answer_history`. A day range is widened by one UTC day on each side because slot days straddle UTC dates; the exact `day` label is on every row.

## citation\_timeline

One domain's, or one page's, citations per slot day, split by platform, from the same rollup as the sources board, so the two never disagree. Every day of the window is present: a day with zero citations is a real zero. `mean_position` is null when nothing that day carried a position (never 0). `prompts` is how many distinct prompts' answers cited it that day. `days` 1 to 365, default 28.

## page\_history

One of your own pages' snapshot lineage, oldest first: when each version was captured, its title, description, word count, status, HTTP status, JSON-LD presence and types, canonical and robots meta, and `changes`, the fields that moved since the version before. `changed_bits` is how far the body drifted.

* `json_ld` is `present`, `absent`, or `not_captured`. **Not captured is not absent**: versions taken before we recorded metadata cannot say the page lacked JSON-LD, only that we did not look. Metadata is only diffed between two versions that both captured it, so starting to capture never reads as the page changing.
* `known` says whether the page is in your inventory; a known page with no versions has never been read, which is different from "unchanged".
* The URL must be on the workspace's own domain; other sites are refused.

## Over REST

Not yet. All four are [MCP-only for now](/api/endpoints#mcp-only-for-now) — call them through the [MCP server](/mcp/overview), which takes the same `atn_live_` key, with `project` when the credential covers several workspaces. The payloads are the contract, so an endpoint that lands later will answer the way the tool does today.