OpenRush

Fast, opinionated technical / content health audit across a sampled set of a domain's pages, tuned for organic-search signal over noise. Discovers URLs from the sitemap/robots (or a homepage link-crawl fallback), then samples one representative per URL template BY VALUE: money/content pages first, while enumerated low-value sections (e.g. a jobs board or event listings) get a single spot-check rather than eating the budget. Fetches the sample concurrently, then GATES OUT non-indexable pages — anything intentionally `noindex` (on a low-value page) or canonicalized to another URL is excluded from scoring instead of being reported as a problem (a noindexed job listing is working as intended). Returns: grouped `issue_clusters` (issue + severity + affected page count + which page roles + a fix), a ranked `recommendations` list weighted by the *value* of the pages affected, an `onpage_score`, and cross-page duplicate-title/description checks. Full per-page rows — including excluded pages and why — are linked via `resources` (fetch with export_dataset). Runs in seconds. - `max_pages` (default 10, max 20): how many sampled pages to audit. Key fields: `page_class_summary` (sampled pages by role: core/content/ utility/ephemeral), `pages_excluded` + `excluded_summary` (non-indexable pages skipped, by reason), and `worst_page` (the highest-value page with the most issues — the best drill-down target). Each cluster/recommendation carries `page_classes` so you can tell whether a finding hits money pages or just utility ones. Scope: this reads each page's raw HTML — it does NOT execute JavaScript or measure Core Web Vitals. `render_warning: true` means ≥1 audited page looks client-rendered (its raw HTML is near-empty), so its content checks are low-confidence. Also read `coverage` (audited vs discovered), `discovery_source` (a `homepage_fallback` means no sitemap was found), and `fetch_failures` (pages that didn't respond in time). To confirm which findings hit pages that actually earn search traffic, follow the `inspect_domain` / `inspect_page` next_actions. Only audit a domain the user owns or operates. If ownership isn't already clear from the conversation, confirm with the user before calling this tool.

audit_site

THE backlink gap / link intersect, done server-side in ONE call: referring domains that link to several of a client's competitors, ranked by how many competitors link them, spam- and authority-filtered, with the ones the client already earned flagged. This replaces the hand-run recipe (pull each competitor's referring domains, union thousands of rows, dedupe against the client, enrich for true authority, filter, sort) — all of which used to run in the agent's context. Here it is one call and a compact, report-ready result. How it works: ranks the competitors by true authority, fans out referring-domain pulls to the strongest of them (the top-N recover essentially the full intersect), unions and counts the overlap, attaches TRUE domain authority + spam via the bulk endpoint (the row-level rank on a referring row is page-level and must not be trusted), drops spam ≥ `max_spam` and authority < `min_authority`, flags each row `earned` (client already links from it) or a fresh gap, and sorts by (competitor_count desc, authority desc). Fan-out and per-competitor depth are handled for you — there are no knobs for them. Read `data.summary` for the funnel (union → candidates → after-filter → gap/earned) and per-competitor coverage, and `coverage.scope_note` for what was and wasn't covered. `data.gap` holds the top `limit` rows; `resources` holds the FULL ranked list — fetch it with export_dataset, don't re-run per-competitor pulls. A short list is a good result: if the client already owns the strong domains, that shows up as earned=true, not as junk padding. First get a real competitor set from discover_competitors (seed_keywords mode for a specific vertical); passing domain-similarity giants produces a meaningless gap.

compare_backlink_gap

Keyword gap analysis: terms one or more competitors (1-5 domains) rank for that `domain` does not, plus terms where the target is losing ground. Use after discover_competitors to turn a competitor set into a concrete content/keyword target list. Results group into intent clusters so the agent can prioritize. Both target and competitors must be reasonably indexed for the gap set to be meaningful — see `coverage`.

compare_keyword_coverage

START HERE. List the live OpenRush surface: enabled data domains, the tools in each, the fact types every tool can emit, which sources back them (live SERP, search index, backlinks, owned Search Console/Analytics/Ads), and which tools need an owned-data connection. Use this to discover what's available before assuming a tool exists. Returns the standard OFE envelope; the tool list is in `data.tools`.

describe_capabilities

Which domains AI cites on a category — the map of where to earn a mention, and who AI cites instead of you. Provide `topic` (a real category keyword) or `domain`, and if both are given `topic` wins. Prefer `topic` whenever you know the brand's real category: pass the category as `topic` (and the brand as `domain` if you also want its standing). Fall back to `domain` alone only when the category is unknown — the tool then derives it from the domain's top ranking keywords, which is reliable for an established site but can pick a tangential term for a thin or new site that barely ranks (e.g. a website-migration brand whose top keyword is "migrate ai", an AI/cloud term). When that happens the result is labeled `data.category_confidence: "derived_weak"` and carries a next-action asking you to supply the real category as `topic`; re-run with an explicit `topic` for an accurate map. Coverage is Google AI Overview only (the deep, independently validated surface); this does NOT cover ChatGPT, Claude, Gemini, Perplexity, or CoPilot. The metric is "this domain cited as a source", not "brand discussed". Results are domain-anchored, never brand-name matched. A scattered or tagline-like input returns a low-signal `data_freshness` note instead of junk rows, and "AI search volume" is never surfaced. Use this even when a brand has zero AI presence: it shows where its category is cited so it knows where to earn placement. `next_actions` point to inspect_ai_visibility to see where a domain stands among these sources.

discover_ai_citations

Find organic competitors. Provide exactly one of `domain` or `seed_keywords`. - domain mode: competitors by keyword overlap with the target's indexed set. Broad, but unreliable for new/small sites because the search index lags the live SERP — the response flags this with a `data_freshness` fact. - seed_keywords mode (1-20 queries): aggregates who consistently ranks across those queries by hitting the live SERP database directly. Needs the SEEDS indexed in that market, not the target — returns empty, with a reason, when they are not. Prefer it when domain mode came back thin; read `coverage`.

discover_competitors

Fetch the full rows behind a dataset `uri` that another tool returned in its `resources` block (results too large to inline). Pass the uri verbatim; do not construct one yourself. Datasets are retained for a limited window and may be dropped sooner under load, so a valid uri can come back empty. That is not an error: when it happens the response has `data.available = false` with a `note`, and `next_actions` contains the tool to re-run to regenerate the dataset. Follow that action, then export the fresh uri from the new `resources` block.

export_dataset

Measured, normalized ad performance (spend, impressions, clicks, CTR, CPC, CPM, conversions, conversion value, cost/conversion, conversion rate, ROAS) for one of the user's OWN connected ad accounts — Google Ads today, more ad platforms later — straight from the platform, not a modeled estimate. THE single tool for every ad platform: ask once, and OpenRush handles which platforms are connected. `platform` and `ad_account` are optional: omit both and the single connected platform + activated account are auto-selected. Several connected platforms return `platform_required`; several activated accounts return `account_required`; nothing connected/activated returns `connection_required` (set up in the dashboard). One platform per call — `next_actions` point to any other connected platform and to deeper slices. - `report`: overview (default: account totals + top campaigns + daily trend + Google network split) | campaign | ad_group | asset_group | ad | date | device | network | keyword | search_term (the last three Google-only). Use asset_group for Performance Max campaigns — they have no ad_group/ad rows; their creative unit is the asset group. Every other type (Search, Shopping, Demand Gen, Video, Display) has ad groups and ads. next_actions route to the right drilldown per campaign type, and to both on a mixed account. - `group_by`: the lower-level equivalent of `report` (one dimension) - `period`: last_7_days | last_28_days (default) | last_3_months | last_6_months | last_12_months - `compare_to`: previous_period (default) | previous_year | none Every monetary value carries its currency and is NEVER converted; every conversion/ROAS number carries its attribution basis. Windows are in the account's own time zone; the most recent ~1-2 days are preliminary. Ad clicks are PAID clicks — not GA4 sessions and not Search Console clicks. Read-only: no campaign, budget, or bid changes.

get_ad_performance

Measured Google Search performance (clicks, impressions, CTR, impression-weighted position) for one of the user's OWN websites, straight from their connected Search Console — not a modeled estimate. `website_id` and `domain` are optional, interchangeable selectors: omit both and the single connected website is auto-selected. If the account has several connected websites the call returns a `website_required` error listing safe options in `data.options` — retry with one `website_id`. If nothing is connected yet it returns `connection_required`/`binding_required` (set up in the dashboard). - `period`: last_7_days | last_28_days (default) | last_3_months | last_6_months | last_12_months | last_16_months - `compare_to`: previous_period (default) | previous_year | none - `group_by`: any of page | query | date | device | country (omit for the site total; page + query is allowed but bounded) - `search_type`: web (default) | image | video | news | discover The most recent ~3 days are labeled preliminary; GSC reports search performance, which is not the same as total website traffic.

get_search_performance

Measured website analytics (sessions, users, engaged sessions, engagement rate, average engagement time, page views, bounce rate) for one of the user's OWN websites, straight from their connected Google Analytics (GA4) property — not a modeled estimate. `website_id` and `domain` are optional, interchangeable selectors: omit both and the single connected website is auto-selected. If the account has several connected websites the call returns a `website_required` error listing safe options in `data.options` — retry with one `website_id`. If nothing is connected yet it returns `connection_required`/`binding_required` (set up in the dashboard). - `report`: overview (default, a rich summary: site totals + top channels + top landing pages + a daily trend) | channel | source_medium | campaign | landing_page | country | device | date - `group_by`: the lower-level equivalent of `report` (one dimension) - `period`: last_7_days | last_28_days (default) | last_3_months | last_6_months | last_12_months - `compare_to`: previous_period (default) | previous_year | none Windows are computed in the property's own reporting timezone. The most recent ~1-2 days are labeled preliminary. Sessions are visits, not users, and are not the same as Search Console clicks. This is the universal traffic/attribution/engagement tool: it deliberately reports NO conversions, revenue, or custom events.

get_website_analytics