Who owns AI search is the first fight in most organizations, and the fight happens because the honest answer is inconvenient: the work spans SEO, content, PR, and paid at once. Retrieval visibility lives in SEO and content. Memory visibility lives in PR, links, and community. Paid placement is its own budget line. No existing department contains all of it, so give it one accountable owner, map the contributors, and run it as a program with a cadence. This chapter is the program: the production pipeline, the ownership models, the governance layer, and the standing procedures.
The production pipeline, brief to monitor
Content for this channel goes through seven stages, and each stage exists because a specific failure appears without it.
Brief. Every piece starts from the prompt inventory, per Prompt and Topic Research: which prompts should this page win, in which mode, memory or retrieval, against which incumbent. A brief that cannot name its prompts produces content with no measurable job.
Draft. Written to be lifted and credited, per Writing for Retrieval and Citation. Claims go in complete, extractable sentences with the brand and category vocabulary inside them.
Evidence check. Every factual claim gets a source or gets cut. This is the stage that protects you years out, because published claims are future training data, and a wrong claim on your own domain can come back out of a model's mouth long after you fix the page.
Chunk review. Read the piece the way retrieval will: as isolated passages. Each section has to make sense, and name its subject, without the surrounding page. Content Architecture: Building for the Chunk gives the tests.
Markup. Structured data goes on per Structured Data and Machine-Readable Facts. Every fact in the markup is pulled from the brand-facts repository below, never typed fresh.
Publish and monitor. Publication is the midpoint, not the finish. The page enters the tracking system from Building a Visibility Tracking System, and the brief's target prompts are the metrics it gets judged on.
Using models to make content models will judge
Ban nothing; gate everything. A blanket ban on AI-assisted drafting is unenforceable and gives up real speed, and a blanket allowance produces the scaled, unoriginal output that platform policies already target: Google's spam policies define scaled content abuse as many pages generated primarily to manipulate rankings rather than help users, explicitly regardless of how the content was produced.Google's spam policies, which cover scaled content abuse, site reputation abuse, and cloaking, and apply however content is generated, by hand or by model.
The workable policy is a quality gate at two stages. Models may draft; the evidence check stays human, because models fabricate sources under exactly the fluency that makes their drafts attractive. And every piece carries a named editor of record who signs the claims. If no one will put their name on a page, the page does not ship.
Ownership at 1, 10, and 100
Team of one. You are the owner and every contributor, so the only strategy is sequencing: run The First 90 Days in order, and resist running its phases in parallel. The failure mode at this size is a half-built tracking system next to half-optimized pages, with neither producing a decision.
Team of ten. The owner is usually the senior SEO, retitled or otherwise, because the measurement machinery here descends from SEO practice. PR contributes coverage on brief, content contributes to the pipeline, paid stays coordinated on the prompts where organic presence is weak. The retraining priorities for an SEO team moving into this work, roughly in order: sampling-based measurement instead of rank checking, entity and structured-data work, and reading citation data as competitive intelligence, per Competitive Analysis.
Team of one hundred. Central platform, distributed production. A small core team owns the tracking system, the prompt inventory, the governance layer, and the experiment log; business units produce content through the pipeline with the core team as the gate. What has to stay central is the fact layer, because that is where distributed organizations rot first.
The fact layer: one truth, everywhere machines read
Models assemble your brand from every copy of your facts they can reach: your site, your schema, Wikidata, directories, profiles, press boilerplate. When the copies disagree, the machine's picture of you blurs, and Entities: Becoming a Thing the Machine Knows shows why consistency is the entity signal. Two artifacts keep the copies identical.
The brand-facts repository holds the canonical version of every fact machines encounter: legal name, founding date, locations, pricing, product names, executive names, category descriptions. One file, one owner, versioned, and every downstream copy traces to it. The approved-claims library holds the sentences you are willing to see a model repeat: each claim paired with its evidence and its review date. Marketing wants to say it; the library decides whether it is sourced. In regulated industries the library is also the compliance boundary, the list of statements legal has cleared for machines to absorb and repeat.
The standing procedures
Misinformation response. When tracking shows a model stating something false about you, the SOP is: document the answers and their citations, correct the source pages the citations point to, refresh the canonical facts everywhere they are published, and use the platforms' feedback channels for the specific false answers. Manipulation, Spam, and Risk covers the adversarial version, where the false information was planted.
Quarterly refresh. Stale facts are the most common self-inflicted wound, and Freshness, Cutoffs, and Temporal Accuracy explains the two staleness clocks. Each quarter, reverify the repository, sweep the highest-value pages for outdated claims, and update dates only when substance actually changed.
Experiment log. One change, one hypothesis, a before-and-after measurement window, written down. Answers are probabilistic, so undocumented changes plus noisy metrics equals learning nothing; the log is what turns a year of activity into a year of evidence about what works in your category.