What I Require Before a Placement Counts

I would count a placement when an answer engine can extract the company name and a bounded claim from a page it can crawl, and I would leave every other placement in a clips file until that test is met. The reader I am writing for, on the queries that now shape a short list, is an inference engine deciding what it will attribute. A lede written only for a human can still be a good story. It is not yet a citation.

The fuller account of coverage built for that reader is PR for Machine Readers. That page is the deeper resource. This essay is how I would use it as a founder: the three properties I would require in the copy, the order of outlets and owned pages, and the result that would make me count a placement on prestige alone.

How I would brief the retrieval, without pretending I have the ranker

When someone asks an answer engine which companies belong on a short list, the engine retrieves. It scores candidate documents for relevance, authority, and freshness, passes a high-scoring set to the model, and the model chooses what to surface. Perplexity's citation design, as practitioners have documented it, weighs relevance, domain signals, freshness, and extractability: can the model pull a claim off the page and attribute it?

Extractability is the criterion a traditional clips report ignores. Human-readable coverage favors a narrative, a buried quote, and color. A model doing retrieval looks for a declarative sentence, a named entity, a specific figure, and a structure it can lift. A quote in paragraph seven of a long feature can be the line a person remembers. It can also fail the extraction pass. I would ask for the company name and the claim in the same sentence, early, and I would treat a buried quote as a human asset that still needs a machine-readable twin.

Research from April 2026 on citation selection versus citation absorption shows those are different events (arXiv). A page can be cited as a source and still contribute little or nothing to the answer. That gap is where a program leaks while the clip report stays green. I would add an absorption check to the report: on the buyer questions we track, did our sentence survive into the answer, or did only the URL survive?

The three properties I would require

Named-entity clarity. The brand, the person, and the claim should sit close enough to resolve without a puzzle. A sentence that names the company and states the fact is extractable. A sentence that says "the company" and trusts the previous paragraph is a sentence I would rewrite before I called the placement done. I would not fill the blank with a superlative. I would fill it with a fact we can defend, because a specific false claim is worse than a vague true one. The requirement is specificity we can stand behind, not specificity for its own sound.

Claim specificity. "A leading company in the space" survives a human skim and fails extraction. A bounded claim, with a scope and a time, is the kind of sentence a model can lift and a reader can check. I would rather ship one such sentence than a page of atmosphere. I would not invent the number to satisfy the shape. If we do not have a figure, the sentence should say what we do and for whom, in the category's own words, and stop.

Crawlable context. The page has to be indexable and reachable, without a paywall or a rendering path that leaves the text out of the fetch. A precise claim on a page the engine cannot read is a precise claim in a drawer. I would ask, before the exclusive, whether the archive is open. Prestige that cannot be fetched does not enter the candidate set, however much a person values the logo.

Where a release fits, and where it does not

I would not brief the team that press releases are useless for answer engines. Writing that treats the release as a citable document points at newsroom-published releases as roughly 18% of ChatGPT citations (the post's own approximate figure, with no primary study named), with original editorial the larger share across platforms. The question is whether the release contains a clean claim. Most are written as an inverted pyramid for a journalist: a quote, then boilerplate. The model is indifferent to the boilerplate. The first paragraphs are where the name, the claim, and the context have to compress. "We are transforming the category" is not a claim I would expect to survive. A sentence that names the company, names the category, and states a checkable fact is the sentence I would lead with, if the fact exists.

I would keep the release in the program as distribution of a fact. I would not let the release be the only place the fact exists. One great placement is weaker than several consistent ones, because triangulation is how extraction confidence rises. Distribution and the pitch are the same system. A placement that links to a vague page on our site loses the attribution on the hop. A placement that links to an answer-first page, with the same name and the same claim, gives the engine a second copy it can resolve.

How I would place this next to the older disciplines

I coined Machine Relations in 2024 so I would have a name for the success condition these engines actually apply. The older rows are still real. They optimize for a different finish line.

A team can be excellent at the digital-PR row while the buyer is being assembled in the last row. That is a mismatch of finish lines. I would not insult the placement. I would add the test the placement has not yet passed. The metric I would put next to the clip count is the brand's share of qualifying citations across the queries we track. Divide by answers instead of by citations and you have a citation rate, a different number from the same sample. If that share is flat, the program is producing visibility a person can see and a model does not use. I would say that in the quarterly review, with the denominator on the page.

The sequence I would run is short.

I would target publications whose archives can be fetched. A gated or technically blocked site can still be worth a human reader. It is a weak candidate for the retrieval set.

I would ask that the brand, the relevant person, and the core claim appear in the body of the first paragraphs, not only inside a quote. Quotes are for humans. The body sentence is for the extract.

I would build corroboration on purpose. Four consistent sources beat one spectacular source that no other page repeats.

I would make the owned page the placement links to as extractable as the placement. They are one system. A vague destination wastes the citation.

I would measure absorption, not only the placement. An editor's yes says a person wanted the story. A sentence that shows up in a tracked answer says the model used it. That is the distinction I would build the plan around.

The discipline those five steps belong to is Machine Relations. The metric I would not let a clips report rename is share of citation.

The decision

I would add the extraction test to every placement we count this quarter. Name and claim in one early sentence, on a page that can be fetched, corroborated by at least one other independent source before I call it a pattern, and checked against the buyer prompts for absorption rather than for the clip. I would keep writing for the human editor. I would stop letting the human editor's yes be the only close.

Scope: earned coverage whose job is to be cited in Perplexity, ChatGPT, Gemini, or Claude, using the selection-versus-absorption distinction in the April 2026 paper. It does not govern a profile whose only job is a human reputation moment, and it does not say a paywalled feature has no value for people. It says that feature is a different line on the report.

What would change my mind: a disclosed measurement, on a corpus large enough to compare with the absorption study, showing that narrative color and buried quotes are absorbed as often as early, bounded claims. I would also count prestige alone if our tracked prompts started citing paywalled or script-blocked pages at the same rate as open archives. I would not count it because a logo looked good in a board appendix.

FAQ

What do I mean by a placement that counts for answer engines?

A page an engine can fetch, where the company name and a bounded, true claim sit close enough to lift, and where that claim shows up in the answers to the buyer questions we track. Selection as a cited source and absorption into the answer are different events, which is the distinction drawn in the April 2026 citation paper. A clip that never enters the answer stays in the human report.

Do press releases still belong in the program?

Yes, when the first paragraphs contain the name and a checkable claim, and when the release is not the only copy of that claim. Newsroom-published releases are a real input to ChatGPT citations in one blog's approximate estimate, and original editorial remains the larger one across platforms. Boilerplate and a transforming-the-industry sentence are the parts I would not expect a model to keep.

How is this different from SEO, GEO, or a classic PR placement?

SEO closes on a results-page position. GEO closes on a citation and tends to emphasize format. AEO closes on being the extracted answer. Classic PR closes on the placement itself. I use Machine Relations for the finish line that includes all of that and still asks whether the entity resolved and the citation showed up across engines. A placement can succeed at the classic finish line and fail this one.

What would I put in the quarterly number?

Share of qualifying citations across the tracked queries, with the query set and the engine list written down. The same sample divided by answers is a citation rate, and I would not let the two share a name. If the share is flat while the clip count rises, the program is producing human-facing coverage the retrieval pass is not using. That is the sentence I would rather hear in the review than a record placement total.