How I Brief Earned Media for a Machine Reader

I would brief a placement for the machine that retrieves it, and I would keep a separate brief for the human who reads it. The human brief can care about tone and the relationship that got the piece assigned. The machine brief cares whether a claim can be lifted, attributed, and repeated. If I only write the first brief, the article can look credible to a person and never enter the answer a buyer asked for.

The method is AI Platforms Like ChatGPT and Perplexity Decide What to Cite. That page is the deeper resource. This essay is how I would use it as a founder: the four properties I would require in a draft, the questions I would ask before I paid for reach, and the measurement that would let me go back to briefing for people alone.

What a retriever does with a finished article

Research on structure-aware reading, including the DeepRead work on arXiv, finds that retrieval favors claims that are clearly attributed, specific to an entity, backed by a checkable figure, and extractable without the surrounding story (arXiv). A paragraph that a human editor would call elegant, and that buries the company inside the narrative, is easy for that retriever to skip. I would not call this a style preference. I would call it the difference between a page that can be cited and a page that can only be admired.

Google's documentation describes structured data as a standardized format for information about a page and for classifying the page content: explicit clues, rather than a parser left to infer the meaning (Google). I would not pretend a block of JSON is the whole job. I would treat it as the same instinct the prose should follow. "A leading firm" fails. A sentence that names the company, what it does, and who stands behind it passes, because the engine can attach the claim to one node. The citation-absorption research makes the second demand: engines prefer claims that still mean something after they leave the original paragraph (arXiv). A number with a scope survives. A vibe does not. I have written about building for both selection and reuse. The brief is where that distinction becomes an edit.

Pew measured how rarely the person gets a second chance to do the reading the model skipped. Across 68,879 Google queries from 900 tracked U.S. adults, users who saw an AI summary clicked a link inside it on 1% of visits, and clicked any result on 8% of visits, against 15% when no summary appeared (Pew). I would brief as if that 1% is the click I might get, and as if the summary is the reading I will definitely get. The article has to work as a source for the summary.

The four properties I would not let a draft leave without

Entity clarity comes first. The draft names one company, one category, and the person if the person is part of the claim. A knowledge graph cannot resolve "the team" into a citation I can recognize later.

Extractable claims come next. I would mark the sentences that still stand alone: a subject, a verb, an object, a bound. If the only good line depends on the anecdote above it, I would rewrite the line. The absorption paper is the reason. A claim that dies outside its paragraph cannot show up in an answer except as a vague trace.

Source stacking is the third. One article is one signal. If our site, a research domain, and two independent outlets state the same entity and the same claim, the retriever has a pattern. I would not call a single prestigious URL a layer. I would call it a start, and I would write the second corroboration into the plan before the first piece runs.

Topical proximity is the fourth. Engines cluster sources around the query, not around the founder's sense of which magazine is impressive. A placement in a general business outlet, with no proximity to the question a buyer asks, can be absent from that cluster no matter the outlet's authority. I would ask which sources the engines already cite for the query, and I would aim the pitch there. Reach, impressions, and social lift remain legitimate things to buy for an investor or a human brand. They are a different purchase from citation infrastructure. I would not let one statement of work pretend to be both without saying so.

The questions I would ask before I approved a list of targets are concrete. Which outlets does the engine already cite for this query category? Which entities has it already associated with the category? How many sources already carry a consistent, extractable claim about us? Does the strongest domain in that set attribute the claim to a named entity, or only to a sector? I can answer the first two by asking the engines and saving the citations. The gap between that list and our current coverage is the brief. Everything else is taste.

What I would do

I would write the brief as four lines at the top of the draft: the entity sentence, the claim sentence, the second domain that will corroborate it, and the query whose current citations justify this outlet. If a writer cannot satisfy the four lines, the piece can still run for human reasons and I would label it that way in the report. I would not label it as work on the answer.

The buyer has already started the loop Pew measured. They meet the summary before they meet the article. Traditional coverage can still earn trust with a person who opens it. A program that never builds the four properties is absent at the step that now happens first. I have written about how PR shows up in AI search. The operating version is this brief. A placement without it is spend I cannot pull back out of an answer that already shipped.

I would run the program in the order on Machine Relations: the claim we can stand behind, then the third-party page, then a check that the answer used it. When a report calls a profile "machine readable" because it ran on a known site, I send the phrase to the glossary and ask which of the four lines it satisfies.

Scope: categories where buyers ask an engine whom to hire or what to use, before they open a vendor's site. A company that still needs human credibility with a specific reporter should buy that credibility and account for it honestly. I would not stuff entity sentences into a sensitive reported feature if they make the piece worse for the human editor and the editor is the entire point. I would publish the extractable version on a page we control and let the reported feature do the human job.

What would change my mind: a repeat of the Pew design, or a close cousin with the queries and the sample disclosed, in which people who saw an AI summary clicked through at rates near the 15% no-summary baseline, and in which extractable structure showed no relationship to which URLs the summary cited. If the human read the article again, and the machine ignored the shape of the claim, I would brief for the editor and stop spending editorial capital on sentences built to be lifted. The current Pew figures, 1% and 8% against 15%, and the extraction findings in the two arXiv papers, point the other way. I will keep the four lines at the top of the draft.

FAQ

What would I put in the brief and leave out of the prose?

The entity, the bounded claim, the corroborating domain, and the query. Tone, anecdote, and the reporter's angle can stay in the prose if they do not bury those four. DeepRead's result is that structural clarity is what retrieval prioritizes. I would edit for that clarity and let a human editor edit for the reading experience, without letting either pass cancel the other.

Do I stop ordinary PR?

No. Ordinary PR still matters for a person who reads the piece and for credibility with people who are not asking an engine. I would stop pretending it automatically feeds the answer. Pew's 900 adults and 68,879 queries are the reason: the summary is clicked inside on 1% of those visits. The brief for the summary is additional work, not a replacement for the reporter.

Which engines is this for?

Perplexity, ChatGPT, Gemini, and Claude all draw on indexed sources. The exact mix moves. Outlet authority, a consistent entity, and repeated claims across sources are the stable parts I would brief for. I would re-check the citation list for the query rather than freeze a 2024 target list and call it a machine strategy.

How would I audit the current program?

I would take the one query that fills the pipeline, ask the engines, and write down the cited domains. Then I would write down where we actually have coverage. The domains in the first list and not the second are the deficit. A citation layer is the pattern across those domains: the same entity, the same claim, extractable on each. One placement is not a layer, even when the placement is the one we framed for the office.

What would I do with a draft that reads well and fails the four lines?

I would publish it if the human reason is real, and I would not count it toward the machine brief. I would add the extractable sentence on a URL we own, in the language the other sources use, and I would point the next pitch at a domain the query already cites. Google's structured-data documentation is the standard I would imitate in prose: explicit clues about what the page is, not a parser left to guess.