How I Judge a Pay-Per-Placement Offer

I would sign a pay-per-placement agreement when the article gives an answer engine a corroborated, machine-readable reason to say our name. I would not sign it because the invoice is cleaner than a retainer. The pricing model decides who carries the risk that nothing publishes. It does not decide what an engine cites after something publishes. I built AuthorityTech from zero to millions without venture capital, and I still would not let our own commercial model grade this decision. The grade is whether the source changes the answer.

The account of the model is Pay Per Placement PR Agencies: How the Model Works Now. That page is the deeper resource. This essay is how I would use it as a founder: the scorecard I would put next to any offer, the ghost-citation test I would require, and the evidence that would make the invoice the deciding factor again.

What the model fixes, and what the citations say it does not

Pay-per-placement moved campaign risk off the founder who hated a retainer with a vague outcome. Baden Bower's April 16, 2026 report says earned editorial placements produced a 31% lead-to-close rate against 12% for paid advertising and 8% for wire distribution, across 512 business owners surveyed, and that earned placements were cited more often by AI systems than paid or wire content (AP News). I would use that as a reason to prefer earned pages over ads and wires when the goal is a lead that closes. I would not use it as a reason to stop reading the page. An engine does not know whether we paid on delivery or on the first of the month. It knows whether the resulting network gives it a clean source.

Three 2026 counts describe that network, and they do not say "buy more placements." Muck Rack's May 2026 Generative Pulse study looked at more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries and found earned media at 84% of AI citations, with paid and advertorial content at 0.3% (Muck Rack). 5W's May 2026 report put 85.5% of AI citations on earned media rather than brand sites, a 325% increase in AI citations from multi-publication distribution versus brand-only publishing, and a 2.8x ChatGPT citation likelihood for brands on four or more third-party platforms (PR Newswire). Fullintel and UConn, at the International Public Relations Research Conference, found 89% of cited links were earned media and 95% were unpaid, with journalistic content alone at 47% of AI citations (Fullintel). Stacker and Scrunch, in March 2026, tracked 87 earned stories across 30 clients and more than 2,600 prompts on eight AI platforms and found a 239% median lift in AI brand citations within 30 days, with third-party publisher sources ahead of owned content (GlobeNewswire).

A thin quote-and-logo page does not join the 84% in any useful way. A page that states a definition, names the entity the same way other sources do, and can be corroborated is the page those studies are describing. I would pay for the second. The first is a cleaner invoice for a clip.

The ghost citation is the failure mode the invoice hides

Superlines' January–February 2026 analysis of 34,234 AI responses across 10 platforms found that 73% of tracked brand presence consisted of ghost citations: the platform links a source about or by a brand and never says the brand's name in the answer (Superlines). The placement can be real. The citation can be real. The buyer can still never see the company. The model uses the thinking and drops the name. I watch this on queries I track. The answer will lift a method and cite a URL, and the founder or the firm disappears in the generated sentence.

I would not hire a firm that cannot talk about that 73%. Closing it takes three things a placement count does not include. The same name, definition, and role have to appear on more than one domain, because engines attribute names they can verify. The page has to contain a line a model can quote with a person attached, because a statistic tied to a named speaker is harder to anonymize than a paraphrase. And there have to be enough independent sources that naming the entity is safer than mining it. The 2.8x figure for four or more platforms is about verification points, not about a thicker media list.

The same Superlines set measured citation rates across 10 platforms and found a 615x spread between the highest and the lowest:

Platform	Citation rate
Grok	27.01%
Perplexity	13.05%
Google AI Mode	9.09%
ChatGPT	0.59%

"Appeared in AI search" is not a result I would accept on an invoice. Appeared where? Grok cites in more than a quarter of the responses in that set. ChatGPT cites in less than 1%. The same article can be obvious on one surface and absent on another. 5W's research puts 42% of B2B decision-makers using LLMs in the initial buying stages (5W). I would ask which of those surfaces the buyers in our category open, and whether the firm's report separates them. A binary mention across a 615x range is a vanity flag.

Superlines also puts brands 6.5 times more likely to be cited through third-party sources, review sites, news, and forums, than through their own domains. That is the case for earned pages over another post on our site. It is also the case for watching whether the third-party page gets our name into the answer or only our ideas into the model's paraphrase. A 2026 paper on pay-per-crawl pricing argues that as AI systems consume sources directly, instead of only sending referral traffic, the economics of access move toward machine consumption (arXiv). Different market, same implication for this contract: a placement matters if the machine can ingest it and reuse it. A PDF nobody fetches is a line item.

Pages updated within two months earn 28% more AI citations than older content, in the Superlines roundup. I would not churn copy to chase that percentage. I would refuse a contract that publishes once and walks away if the category moves. A source that can accept a new figure stays in the pool. A placement agency that ends at the URL has sold the invoice event and left the retrieval event to decay.

What the market is renaming, and the scorecard I would use anyway

Gartner said on May 12, 2025 that marketing budgets had flatlined at 7.7% of company revenue (Gartner). The pressure is to tie the channel to an output. The cheap response is cost per placement. The response I would use is whether the placement became part of a source graph an engine can keep citing.

Trustpoint Xposure's January 14, 2026 announcement reframed PR as machine validation rather than exposure alone (AP News). Ruder Finn's March 9, 2026 launch of rf.Voices made a similar move from the influence side (AP News). I would read the announcements and then ask what happens after the pitch. If nobody aligned the article with a definition on our site, nobody checked whether the answer says our name, and nobody kept the source current, the new vocabulary did not change the work. I have written about what retainers cost. Price is a real conversation. It is the second conversation. The first is the source.

What I ask	A placement-buying answer	The answer I would sign
What am I purchasing?	A guaranteed article	A third-party URL a machine can reuse
What is the KPI?	A count of URLs	Citation on named engines, and whether our name survives in the answer
What makes one URL better?	The logo and the spike	Clarity, corroboration, topical fit
What fails?	A missed delivery	A live URL that never enters the trust layer
What compounds?	A longer press page	The same claim on independent domains, tied to a definition
Ghost rate?	Not in the report	Measured per platform

The frame for that table is Machine Relations. I keep the invoice words from replacing the retrieval words by checking them against the glossary. A commercial model is when we pay. The source system is whether the web teaches a machine to name us six months later.

How I would write the clause for each study

I would not paste 84% into a contract and call it a guarantee. I would paste the decision each study forces, in language a firm can accept or refuse.

Muck Rack's May 2026 Generative Pulse, more than 25 million links, ChatGPT, Claude, and Gemini, 17 industries, earned media at 84% of citations and paid or advertorial content at 0.3%, becomes a clause about form. The page has to be reported, not bought as an advertorial with our adjectives. If the draft reads as a release we wrote, it sits on the 0.3% side of that study no matter what the invoice says. I would reserve the right to reject a draft that fails that reading, without paying for the rejection. Pay-per-placement only works if "placement" means a page an editor would have run. Baden Bower's April 16, 2026 survey, 512 owners, 31% lead-to-close for earned editorial against 12% for paid ads and 8% for wire, is the commercial reason I prefer that form when the goal is a lead. It is a survey of owners, not a crawl of engines. I would cite it as a business outcome and keep Muck Rack as the citation outcome. They are allowed to point the same direction. They are not the same evidence.

5W's May figures become a clause about hosts. Multi-publication distribution associated with a 325% citation increase versus brand-only publishing, and four or more third-party platforms associated with 2.8x the ChatGPT citation likelihood. I would define a platform as a registrable domain we do not own. I would not let five URLs on one syndicator count as five. Fullintel and UConn, 89% of cited links earned, 95% unpaid, journalistic content at 47%, is the clause against a network of unpaid-looking pages that are the same owner. I would ask who owns the domains. Stacker and Scrunch, 87 stories, 30 clients, more than 2,600 prompts, eight platforms, 239% median lift within 30 days, third-party ahead of owned, becomes a measurement window, not a promise. I would re-measure at 30 days on a frozen prompt list. I would not let the firm swap the prompts if the median does not appear. Their median is their clients. Our window is ours.

Superlines is three clauses, and I would not collapse them. Ghost citations, 73% of tracked brand presence across 34,234 responses and 10 platforms in January and February 2026, means the acceptance test is the answer text, not the live URL. I would sample the engines our buyers use and fail the placement if the source is linked and the name is absent, then require a revision of the entity sentence rather than another logo. The 615x citation-rate spread, Grok at 27.01%, Perplexity at 13.05%, Google AI Mode at 9.09%, ChatGPT at 0.59%, means the report lists engines separately. An average is a breach. The 6.5x likelihood of citation via third-party sources versus our own domain means I will not accept a plan that "also updates the blog" as the corroboration. The blog is the record the third party can point at. It is not the second platform. The 28% citation edge for pages updated inside two months means the statement of work names who edits the URL if a primary number changes inside the quarter. I would rather hold the edit myself than pay for a freshness theater of swapped adjectives.

The pay-per-crawl paper on arXiv is not a PR study. I would use it only for the implication it shares: machines consume the source, they do not only send a visit. A PDF that is never fetched fails that implication. Gartner's May 12, 2025 figure, budgets flat at 7.7% of revenue, is why the cheaper contract wins the meeting. I would write the price on the last page. Trustpoint Xposure's January 14, 2026 announcement and Ruder Finn's March 9 rf.Voices launch are evidence that the market is renaming the work. I would read the announcement, then apply the clauses above to the draft. A new name that cannot survive the ghost-citation sample is packaging.

5W's separate note, 42% of B2B decision-makers using LLMs in the initial stages of a purchase, is how I would choose which row of the platform table matters. If our buyers are in that 42%, ChatGPT at 0.59% citation rate and Grok at 27.01% are not interchangeable deliveries. I would name the engines in the statement of work. "AI coverage" would be a defined term or it would be struck.

I would also separate the invoice from the attribution on purpose, because I have a commercial model of my own and it should not grade this. A retainer can satisfy every clause above. A pay-on-delivery contract can satisfy none of them. The founder's name in the answer, the same definition on a second domain, and a re-measure that keeps the prompts fixed are the deliverables. The timing of the payment is a financing choice. I would not let a financing choice be described as an architecture.

What a thin placement looks like against the 73%

I would walk one imaginary draft, built only from the failure modes these studies name, so the firm can see the rejection before they write ours. The draft quotes us in the third person, gives no number, uses a category string we do not use on our site, runs on a single host, and has no link to a definition we control. It will publish. It may even be earned in the narrow sense that an editor approved it. It does not give an engine a name it can verify, so it is eligible to become one of Superlines' ghost citations, a URL in the sources and a paraphrase in the answer. It cannot use the 2.8x, because there is no second platform. It cannot use the 325%, because there is no multi-publication distribution, only a brand-adjacent post. It sits closer to the 0.3% advertorial pattern if the sentences are ours, even when the logo is a newsroom. Thirty days later the Stacker window closes on a prompt list nobody froze, and the report shows a screenshot.

The draft I would accept names the company the way the site does, states one checkable claim with a scope, points at a page we own for the definition, and is designed to be corroborated on a second domain we do not own. The report at 30 days includes the answer text from each engine in the 615x table that we agreed mattered, and it marks whether the name survived. If the name did not survive, the next dollar goes to the entity sentence, not to a second logo of the same shape. That is the whole operating difference between paying for a placement and paying for a source. I would put it in the first paragraph of the order form so the rest of the rate card has something to serve.

What I would keep out of the scorecard

I would keep cost-per-placement off the line that decides whether to continue. Gartner's 7.7% is real pressure, and a founder can do that arithmetic in a spreadsheet without pretending it measures trust. The line that decides continuation is the sample of answer text: name present or name absent, engine by engine, prompts frozen. If ChatGPT's 0.59% citation rate means our buyers' engine rarely cites anyone, I would not fire the firm for a low citation count on ChatGPT alone. I would ask whether the name appears in the minority of answers that do cite, and whether Grok or Perplexity, at 27.01% and 13.05%, show a different pattern on the same article. A firm that is strong where citations are common and invisible where our buyers actually ask has not finished the job. A firm that is named where our buyers ask, even at a low base rate, has.

I would also keep "guaranteed" attached to publication, never to a citation share. Muck Rack did not guarantee 84% to any one brand. 5W did not guarantee 2.8x to the next signup. Stacker's 239% is a median across 30 clients and 87 stories, which means half the stories did worse than the number a sales deck will lead with. I would ask for the distribution, or I would ignore the median and keep our own 30-day read. Superlines' 73% is a base rate of anonymity, not a curse. A placement can beat a base rate. It does not beat it because the contract said pay on delivery. It beats it because the sentence carried a name a second source already confirmed.

The pages I would refuse to count as that second source are our own blog, a syndicated copy of the same article, and a partner page we edited. Fullintel's 95% unpaid and 89% earned are about independence, not about the absence of an invoice somewhere in the company's history. If we paid for the placement, the page can still be earned in the editorial sense. The corroboration has to be a domain that did not take the same order. That is the clause I would not negotiate away to make a quarterly number. Ten placements that fail it are a more expensive failure than a retainer that never promised them, because the failure is in the source graph, where the next answer will find it.

What I would do

Before I signed I would make the firm answer six questions in the proposal, not on a call. Will the article define what the company does in a sentence that matches our site? Will it reinforce the same entities and category language a second source can corroborate? Will it point at a stronger page we own, research or a definition, instead of standing alone? Will anyone re-measure citation on the commercial queries after it publishes? Will the founder's name appear in the generated answer, or is this another ghost? Which platforms, of the set with the 615x spread, do our buyers use?

If the proposal cannot answer, I would treat the money as distribution spend and I would not let it be described as an AI visibility program. I would rather buy fewer articles that strengthen a chain than ten that add logos. Muck Rack's 0.3% for paid and advertorial content is the boundary. A placement that is earned in form and advertorial in substance does not cross it because the contract said "pay on delivery."

Scope: a company that already has a one-sentence definition and at least one page worth corroborating. Pay-per-placement cannot invent that sentence inside someone else's CMS and then complain the engine anonymized it. This is also the wrong instrument when the goal is a relationship with a reporter over a year. That is a retainer, and I would buy it as a retainer, with the same citation questions attached so the labor does not drift back into activity reports.

What would change my mind: a disclosed comparison, same queries and same engines, in which pay-on-delivery placements and retainer placements produced the same citation share and the same ghost rate, and in which the 73% unnamed share had fallen far enough that "we got the URL" reliably meant "the answer says the name." If the invoice structure stopped correlating with any difference in the answer, I would buy on price and convenience. Superlines' 73%, the 615x platform spread, and the 84% earned-media share say the answer still depends on the page and the network, not on the billing event. I will keep judging the page.

FAQ

Does pay-per-placement help AI visibility at all?

It can, when the URL is a durable third-party source an engine will ingest. Muck Rack's May 2026 study of more than 25 million citations put earned media at 84% and paid or advertorial content at 0.3%. 5W's separate May study put a 2.8x ChatGPT likelihood on brands present on four or more third-party platforms, and a 325% citation increase for multi-publication distribution against brand-only publishing. The billing event is not in those findings. The page is.

How is this different from a retainer?

The retainer buys ongoing labor. Pay-per-placement buys a delivered URL. Neither one creates a citation by virtue of the contract. I would attach the same scorecard to both: definition, corroboration, named attribution in the answer, and a re-measure on a fixed prompt list. Gartner's 7.7% budget figure is why the cheaper-looking contract wins meetings. It is not why an engine picks a source.

What is a ghost citation, in the decision?

A linked source that never produces the brand name in the answer text. Superlines found that pattern in 73% of tracked brand presence across 34,234 responses. I would ask the firm how they will know if we become that 73%. Entity consistency, a quotable named claim, and a second domain saying the same thing are the repairs. A screenshot of the article is not.

Which platform number would I refuse to average?

Any average across Grok at 27.01%, Perplexity at 13.05%, Google AI Mode at 9.09%, and ChatGPT at 0.59%. That is the 615x range in the Superlines set. I would report the engines our buyers use. 5W's 42% of B2B decision-makers using LLMs early in a purchase is the reason the choice of engine is the purchase, not a footnote under "AI coverage."

When would freshness change the contract?

Superlines' figure is 28% more AI citations for pages updated within two months than for older content. I would write an update right into the engagement if a new primary number is likely inside a quarter. I would not pay for meaningless edits. A source the firm abandons on publication day is a source I should have kept the rights and the habit to maintain myself.