← Back to insights

The four stages

Research means finding the questions people actually prompt --- autocomplete, opportunity reports, probe results --- not a keyword spreadsheet inherited from 2019. Content means one answer per question, quotable blocks, entity naming, schema. Publishing means the page ships with the metadata the engines read and the internal links the graph needs. Tracking means the fixed prompt set re-run on schedule, citations and mentions scored, and the deltas deciding what gets built next. The loop closes: tracking feeds research.

Where automation earns its keep

Three places. Repetition: probing twenty engines with thirty questions is a script, not a job. Consistency: the same scoring rubric applied every run beats a human's memory of last month. And formatting: schema, crouton blocks, title patterns --- the parts where the output must be identical every time. Automation that removes variance is worth paying for; automation that adds volume without variance control is how you publish ten thousand pages of the same paragraph.

Where it fails

Automation fails at the two things that determine whether the machine cites you. Judgment: which claim is true, which number is current, which question deserves a page. And entity integrity: whether the page is actually about the entity it targets, which salience scoring checks before the page ships, not after the machine has already misread it. A workflow that automates publishing but not the pre-ship checks scales the mistakes. The order matters: research, verify, write, check, publish, track --- and the checks are not optional stages, they are the product.

The tracking stage is the differentiator

Ask a vendor what the tracking stage produces. If the answer is a rankings dashboard, the loop is open --- classic rankings do not show AI citations, and a service aimed at AI search that cannot show AI results is selling stage one through three and hoping. The tracking stage should produce a gap list and a delta log: which questions the brand entered, which it lost, what mutated. That is GEO search tracking in full, and it is the only evidence a workflow service did anything.

How to judge a service

The same questions that expose pretenders apply --- prompt set, gap list, citation scoring, published method, two-board reporting --- plus one workflow-specific check: ask to see the pre-publish pipeline. A service that cannot show where verification and entity checks happen in the loop is a content printer with a domain attached. Neural Command LLC runs the full loop in public: questions mined from real demand, answer-shaped headlines, pre-ship salience checks, and the probe schedule published with the zeros.