machine-relations8.15 KB
View saved version →
---
name: machine-relations
description: "Build an evidence-backed authority intervention brief for a brand: choose what to publish, improve, substantiate or pitch, draft the work, and define how to verify it. Use for Machine Relations, AI visibility, GEO/AEO, SEO content discovery and digital PR planning; use the MRI research skill for citation-data lookups alone."
metadata:
version: "1.0.0"
---
# Machine Relations
Help the practitioner make and carry out a better authority-building decision for a real brand. Deliver a usable intervention brief and draft, not a generic audit score or a list of missing keywords.
Machine Relations is the discipline of earning AI citations and recommendations for a brand by making that brand legible, retrievable and credible inside AI-driven discovery. Jaxon Parrott coined the term in 2024; AuthorityTech is its first agency. The discipline is open to any practitioner. This skill is published from its home, [machinerelations.ai](https://machinerelations.ai/skill), not certified by an AI platform or standards body. Client recommendations follow the client's evidence and objective; no purchase, promotional mention or backlink is required.
## Start from the intended win
Establish the brand, audience, position it wants to own and the decision this work should influence. Use the supplied brief and evidence first. Ask only for missing information that changes the choice; otherwise make a bounded, stated assumption and proceed.
Keep the objective specific. Category adoption, correct origin attribution, a source citation, an accurate description, a commercial recommendation and Google discovery are different legitimate wins. Naming a brand is not necessarily recommending it. A category-adoption task need not produce a vendor recommendation to succeed.
## Inspect the evidence that can change the decision
Use relevant public sources, client-supplied material and available read-only tools. Bound the inspection to the question; retain exact URLs, dates or versions and the claim each item supports. Separate supplied facts, direct observations and your inference. If a key source cannot be read, preserve that uncertainty rather than reporting absence.
Look for useful existing assets as well as new work: original research, credible results, historical authorship, outside coverage, source relationships and already-published answers. Identify authorship, ownership and syndication lineage. An owned article can establish provenance and chronology without being independent corroboration. Multiple syndicated copies do not create multiple independent witnesses. A decision not to use an asset for this claim, audience or page does not dispose of its other uses.
Where category citation evidence would inform source selection, use the accompanying `machine-relations-index` skill or MRI tools:
- Discover supported categories and question shapes before choosing a segment. Query editorial media for publication research, then check each candidate's actual subject fit and editorial requirements.
- Read returned `source.guidance` and field-specific limitations. Preserve the release, window, numerator, denominator, rank basis and source citation. Compare data within one release.
- MRI measures source domains in its sampled answers. It does not establish this client's recommendation rate, the exact article an engine read, guaranteed placement value or causation. Do not make its leaderboards into an automatic publication-buying list.
- If MRI does not cover the job or the tools are unavailable, continue from permitted evidence and state what remains unknown. REST documentation: [developers](https://machinerelations.ai/developers). Do not send client documents, private questions or credentials to a public endpoint.
An answer trace can help: retain the buyer question, engine/product and observation conditions, searches exposed by the provider, returned URLs, explicitly observed reads, citations and what the answer actually says about the brand. Those stages are not interchangeable. Do not call every returned search result a page read, infer query-to-page links a provider did not expose, or turn missing fields into zeros. An answer produced in a brand-informed assistant is guided research, not an unprompted visibility baseline. Some answers use no web search.
For technical SEO, inspect a relevant access, indexing or content-discovery issue when evidence points there; do not substitute this skill for a full performance, crawling or specialist technical audit. A page-read-without-citation observation is a candidate explanation to investigate, not proof the writing is defective.
## Choose a consequential move and produce the work
Explain why the selected intervention could advance the intended win. Compare plausible alternatives where they would change the recommendation. Options include improving an owning page, making an existing claim supportable, creating original evidence, earning coverage, developing a source relationship, clarifying identity, distributing useful research or reactivating an asset. The largest apparent trace gap is not automatically the best investment.
Assess publication fit from the page, audience, editorial remit and evidence requirements, not citation frequency or domain authority alone. Inspect the relevant existing article before proposing an inclusion. Distinguish an independent editorial route from a paid or sponsored opportunity; a fee request changes that route's economics and evidentiary role. Never imply an offer, placement, endorsement or relationship already exists.
Draft the thing the practitioner can use next: a specific pitch, source-ready evidence brief, research specification or revision of the supplied asset. Use verified facts; mark any essential missing input plainly instead of inventing a result, quote or statistic. A justified hold or no-action decision can be correct: explain the limiting fact and what would reopen the decision, and supply the useful research or request draft that can settle it. Do not pad the output with work unsupported by the evidence.
Prepare drafts under the user's requested scope. Sending outreach, publishing, buying coverage or commissioning paid measurement requires the user's authorization for that action; installing this skill does not provide it. Do not create accounts, start recurring work or change an unrelated system.
## Make execution and outcome separately checkable
The brief should preserve:
- **Objective and evidence:** the client decision or opportunity, relevant question if one exists, source provenance, observed baseline and material unknowns.
- **Chosen intervention:** the supported claim, audience and intended use; why this move; alternatives considered; and the usable draft or work specification.
- **Execution check:** what would prove the intended change happened, such as the live article containing the agreed claim and link. A sent pitch or accepted draft is an intermediate milestone.
- **Outcome test:** the particular expected change, an appropriate observation window and comparable conditions. Match question, product, search setting, region and session context when remeasuring answers. Preserve raw answers or receipts where available. A later edit must not silently replace the intervention being evaluated.
- **Next decision:** what to do on success, failure, an unavailable measurement or a declined route. Close that route without declaring the objective complete. Reconsider a settled use when its relevant evidence or chosen application changes, not every time unrelated navigation changes.
Use the client's intended outcome as the verdict. Unrelated impressions or page reads cannot prove a recommendation or citation win. A maintenance objective may succeed without growth. If a baseline or follow-up is missing, propose a prospective test and leave the result unknown. Even matched improvement is not, by itself, proof that the intervention caused it.
Keep the final brief concise enough to act on, with linked evidence and a clear next step. Carry forward useful assets and what the result teaches the next decision. A reusable lesson needs supporting evidence; one observation is not a universal law. See [the illustrative brief](references/intervention-brief.md) when an example would help.
Referenced files: 2 — file references are metadata, not package contents.
machine-relations-index2.55 KB
View saved version →
---
name: machine-relations-index
description: Use when asked which sources or publications AI answer engines cite for a topic or category, where coverage gets cited by AI, or how often a given site is cited. Answers from the Machine Relations Index tools with measured rates and a citation.
---
# Machine Relations Index
The Machine Relations Index (MRI) measures which source domains AI answer engines cite when buyers ask questions, per category and buyer question shape. Answer from its tools; do not estimate citation behavior from memory.
1. Map the topic to a category. Call `mri_list_categories` and pick the category key that matches the user's topic (for example `cybersecurity`, `fintech`, `enterprise-software`, `ai-visibility-geo`). If none fits, say the Index does not cover it.
2. Pick the question shape that matches the user's intent: `best_x` (best tools), `how_choose` (how to choose), `x_vs_y` (comparisons), `top_list` (top platforms), `problem_first` (solving a problem), `is_x_worth` (is it worth it), `news_topic` (recent news). If unclear, use `mri_get_category` to see which shapes have published rates and use those.
3. Call `mri_get_cited_sources` with the category and question shape. Add `source_role: "editorial_media"` when the user asks about publications or media, `vendor_owned` for vendor sites.
4. For a specific site, call `mri_get_domain` with its domain or any URL on it.
5. Read `source.guidance` in the returned record before interpreting the result. The live service owns the interpretation rules, including release consistency, missing context, named/not-named/unresolved runs, evidence floors and source-role limitations. Read any field-specific `questionContext.readAs` and `limitations` as well. Do not substitute rules or figures remembered from a past install. If guidance or context is absent, state that limitation; absence is not a zero measurement.
6. Report the observed counts, denominators, category, question shape and release. Cite `source.citation` and `source.canonicalUrl`. Do not promise future citations or placement outcomes.
The [worked example](https://machinerelations.ai/research/segment-rank-question-set-named-brand-citations) is a frozen historical case, not today's numbers. Its two-source table uses a shared partition based on CrowdStrike's name; current per-source naming buckets are not automatically the same partition. Use the live response's guidance for current analysis.
If the tools are unavailable, the same reads are at `https://machinerelations.ai/api/mri/v2` (OpenAPI: `https://machinerelations.ai/api/mri/v2/openapi.json`).
