Saylent

Sources & opportunities

Where the answers come from, who they name, and which off-site pages are worth winning. Counts with receipts, built from the run's own data.

A run already stores every answer, its verdict, its citations and the cited pages it read. Sources & opportunities reads that data three ways:

  • Citation share: which domains and pages the answers cite, per engine and answer surface
  • Share of voice: how many answers name you, and how many name each rival
  • Off-site opportunities: domains that are neither yours nor a rival's, cited in answers that name a rival and not you

It makes no model call and no fetch of its own, so it costs nothing. Every number is a count, and every count lists the exact answers behind it, by question and engine, so you can open each one.

What counts as an answer here

  • An answer is counted once per question and engine: the answer the report shows, never the extra sampled draws. That is why every receipt opens.
  • Only answered rows count. A surface question that was not shown is not an answer.
  • Names you means the judge's verdict says your brand is present. Recommends means its mention type is "recommended". A rival is named when the verdict lists it among the brands the answer prefers or recommends.
  • Names a rival and not you means the answer was judged, your brand is absent, and at least one rival is named.
  • Branded questions ("is your brand any good") are left out of the opportunities. They are about you, not about the category.

Citation share

For every cited domain:

FieldWhat it is
CitationsCitation rows on this domain, across the counted answers
ShareIts citations divided by every citation in the run
Per engineCitations by each engine and surface
AnswersDistinct answers citing it
Name youAnswers citing it that name you
Name a rival, not youAnswers citing it that name a rival and not you
Top pagesIts most cited pages (up to five), and for each page we read, whether you and which rivals are on it
TypeIts domain type, and the rule that set it (below)

The host is the real source: www. is stripped, Internet Archive snapshots count for the site they copy, and a Gemini grounding redirect link counts for the site its citation names. The run also gets a per-type total (citations and domains per type) and, per engine, its answers, citations, domains and top five domains.

Share of voice

Counts, not percentages: the answers naming you, the answers naming each rival (up to twelve rivals), and how many answers recommend you, each split by engine and each with its receipts. A rival has no "recommended" count, because the verdict does not separate "recommends" from "prefers" for other brands.

This counts mentions in the answers we drew. It is not market share.

Off-site opportunities

An off-site opportunity is a domain that:

  1. is not your site and not a rival's site (by its domain type)
  2. is cited by at least one answer that names a rival and not you
  3. on a question that is not branded
  4. where you were not found on any page we read on it

They are ranked by evidence: the number of distinct such answers that cite the domain. Ties go to the domain cited by more engines, then the one with more citations, then by host name. There are no weights and no score.

Each opportunity carries:

  • its type, and the rule that set it
  • the engines and questions behind it
  • the rivals named in those answers, most often first
  • what the pages we read on it show: rivals are there and you are not, you are not there, or no page on it was read
  • the top cited page on it, the one to pitch first
  • one plain sentence saying why it is listed, the same sentence in every view
  • receipts: the question, the engine, the cited URLs on that domain, the rivals the answer named with the reason it gave, and (in saylent sources and the MCP tool) a short quote of the answer

A run keeps up to twelve opportunities, with up to six receipts each.

Domain types

TypeLabel
own_siteYour site
rival_siteRival's site
review_siteReview site
forumForum / community
docsDocs / help
newsNews
marketplaceMarketplace
referenceReference
videoVideo
socialSocial
editorialBlog / editorial
otherOther

A short list of rules decides each type. The first rule that matches wins, and every view shows the rule that fired.

#RuleTypeShown as
1Your override for this host or a parent domainthe type you set"your override"
2Your domain, or a subdomain of itown_site"your domain"
3A known review site (g2.com, capterra.com, trustpilot.com, ...)review_site"known review site"
4A known community site (reddit.com, stackoverflow.com, ...)forum"known community site"
5A known reference site (wikipedia.org, wikidata.org, britannica.com)reference"known reference site"
6A known video site (youtube.com, vimeo.com, ...)video"known video site"
7A known marketplace or app store (apps.apple.com, play.google.com, any amazon.*, ...)marketplace"known marketplace"
8A known social network (linkedin.com, x.com, ...)social"known social network"
9The host is named after a rival: a configured competitor, or a rival the answers named, spelled out wholerival_site"named after" the rival
10The host contains a configured competitor's name, or a rival's domain redirects thererival_sitethe name it contains, or "a ... domain redirects here"
11A subdomain pattern: docs., help., support., api. and similar; community., forum.; news., press.; blog., resources.; marketplace., apps., store.docs, forum, news, editorial, marketplace"docs/help subdomain" and so on
12What its cited pages look like: when at least half its citations point to one kind of page (best-of listicles or comparisons, docs, news, review listings, community threads, reference entries, videos)editorial, docs, news, review_site, forum, reference, videofor example "3 of 4 citations point to best-of listicles"
13Nothing matchedother"no rule matched"

The host lists live in packages/engine/src/classify.ts and packages/engine/src/sources.ts.

Correct a type

When a rule gets a domain wrong, override it. The override always wins.

In saylent.config.json:

{
  "sources": {
    "domainTypes": {
      "softwarereviews.example": "review_site",
      "blog.rival.example": "rival_site"
    }
  }
}

A parent domain covers its subdomains, and the most specific entry wins. The value must be one of the twelve types above. saylent audit and saylent verify apply it, and the run stores the overrides it was computed with (sources.overrides in run.json), so a re-render reproduces it.

In the app, the brand's run options page has a Source types editor. It changes how every report for that brand reads its stored answers, never what a run asks, so it never starts a new baseline.

Where it shows up

WhereWhat you get
report.html and the app's full report viewSection 04, Sources & opportunities. Every count opens the answers behind it. Site checks, fixes and method are now sections 05, 06 and 07
report.mdA ## Sources & opportunities section
saylent audit finish summaryA Sources line and an Off-site line
saylent sourcesThe full view in the terminal, or JSON (below)
The MCP serverThe sources tool, and sources_summary in the audit result (The MCP server)
The app's brand pageSources over time, across the last six finished runs (below)

The finish summary prints two lines, counts only:

  Sources   <n> citations · <n> domains · most cited: <host> (<n>, <type>)
  Off-site  <n> opportunities · first: <host>, in <n> answers that name a rival and not you

saylent sources

# everything for one run, as text
npx saylent sources ./example.com/2026-11-02/run.json

# only the opportunities that Google AI Overviews cited
npx saylent sources ./example.com/2026-11-02/run.json --view opportunities --engine google_ai_overview

# one domain: its pages and every answer citing it
npx saylent sources ./example.com/2026-11-02/run.json --domain g2.com

# the same JSON the MCP sources tool returns
npx saylent sources ./example.com/2026-11-02/run.json --json
FlagWhat it does
--view <name>summary, opportunities, domains, voice or all (default: all)
--engine <id>Only rows cited by this engine or surface
--type <type>Only domains of this type (one of the twelve above)
--domain <host>One domain: its pages and every answer citing it
--limit <n>Rows per list (default 10, max 50)
--jsonPrint the same JSON the MCP sources tool returns

It reads the bundle and nothing else: no network, no model call, $0. A folder works in place of the run.json path.

History

Pass two or more bundles of the same brand and saylent sources prints the history instead:

npx saylent sources ./example.com/2026-10-02/run.json ./example.com/2026-11-02/run.json

Per run: the citations and the number of off-site opportunities. Per domain: its citations in each run, and how many runs it was an off-site opportunity in. Per brand: the answers naming you and the top rivals in each run. Bundles of different brands are refused. With --json it prints the history as JSON.

The app's brand page shows the same thing as Sources over time, over the last six finished runs: citations per domain per run, "opportunity in N of M runs", and named counts per run. The receipts live in each run's report.

In run.json

run.json has an optional sources block with schema_version: 1:

KeyWhat it holds
totalsanswers, answers with citations, citations, domains, and answers that name a rival and not you (non-branded)
enginesPer engine: answers, citations, domains, top domains
domainsEvery cited domain, most cited first, with the fields in Citation share
byTypeCitations and domains per type
voiceYou and the rivals: named, recommended (you only), per engine, receipts
opportunitiesRanked, with evidence, engines, rivals, page check, reason, top page and receipts
overridesThe type overrides it was computed with

Older bundles have no sources block and still read unchanged: the report, saylent sources and the MCP tool compute it from the bundle's own rows with the same function. A block with a schema_version this build does not know is recomputed the same way. The MCP tool says which happened (recomputed).

What it does not claim

  • A citation shows where an answer pointed. It does not show that the page caused the answer.
  • An opportunity is a place rivals are cited and you are not. Being there is no promise that the answer changes.
  • Share of voice counts the answers we drew, not the market.
  • Domain types come from short rules, not a model. Each one shows its rule, and you can override it.

On this page