How I Would Hire an Earned Media Firm in 2026

I would hire the firm that can raise the chance a buyer’s question retrieves a page which names us clearly, and I would pass on the firm that sells a thicker clip report. The screening questions I used when discovery was a human reading a homepage are the wrong scoreboard once the first pass is an answer engine.

The comparison is Best AI PR and Earned Media Agencies for AI Search 2026. That page is the deeper resource. This essay is how I would use it as a founder: the filter I would run on any firm on that list, including my own, before I signed.

The product change that moved the agency question

OpenAI launched ChatGPT search on October 31, 2024 with linked web sources inside the answer (OpenAI). Source inclusion became part of what the person sees, which means a placement is raw material for a retrieval system and a human reader at the same time. I would judge a firm by whether its articles can survive that retrieval: a named company, a category a buyer would actually ask about, and a claim a machine can lift without the surrounding profile.

Ahrefs, across 75,000 brands, found branded web mentions correlating more strongly with presence in Google AI Overviews than classic link metrics in many of the comparisons they published (Ahrefs data via Search Engine Land). I would not turn a correlation into a ranking formula. I would turn it into a hiring rule. If third-party mentions are the signal that moves with machine-facing visibility, then a firm is selling source quality or it is selling activity. A famous domain with a sloppy article can leave the entity unresolved. A narrower publication that states the category, the company, and the proof in extractable sentences can do more for the next answer.

The Princeton-led GEO paper found that adding quotations, statistics, and citations improved visibility in generative engines by roughly 30% to 40% in the settings it tested (Aggarwal et al.). I would brief for that shape without pretending the paper is an engine’s disclosed ranking recipe. The practical reading is narrower: a placement that buries the company in atmosphere gives a retriever less to lift. I want the sentence a later answer can repeat with the attribution still attached.

What I would be buying, by model

Model	What I am actually buying	When I would use it
Retainer PR	Ongoing labor, coordination, and relationships	A newsroom relationship I need kept warm for reasons other than a single query
Placement on a trusted article	A new third-party node with a checkable claim	A buyer question that currently retrieves competitors and not us
Classic SEO	Retrieval on the old results page	Demand that still arrives as a click to a URL I own
GEO or AEO work	Formatting and distribution aimed at answer surfaces	After the claim and the third-party source exist
Machine Relations	Authority, entity, citation, distribution, and measurement as one system	When the firm is still selling the rows above as separate retainers

That last row is the frame on Machine Relations. The words I refuse to let a pitch redefine are in the glossary. A firm can be excellent at one row and wrong for the job in front of me. I would not pay retainer rates for a placement problem, and I would not pay placement rates for a year of counsel. The contract should name the row.

The five questions I would ask before a signature

I would ask which publications the placements land on, and why those publications are the ones a machine already retrieves for the buyer questions we care about. I would ask how they decide a placement sharpens the category rather than adding another mention of the founder. I would ask what, inside the article, a machine can extract: the company, the category, and the proof, in sentences that still make sense out of context. I would ask how a second and third source will corroborate the same claim, on domains that do not share an owner. I would ask what they expect to change in later answers if the work lands, and how they will show me that change without swapping the prompt list halfway through.

If the answers are logo names and a monthly volume, I would treat the offer as old media-list selling in a new deck. I have written separately about why earned-media ROI software misses the decision this hiring call actually is. More impression charts do not answer the five questions. A firm that cannot connect a placement to a later recommendation is selling distribution with a story attached.

I would also look at the article the way a parser would. Can I underline one sentence that states who the company is? Can I underline a number with a scope and a date? Does the same category string appear on our own site, in the pitch, and in the draft? Inconsistent names are how a chain fails the corroboration a retriever needs. I would rather delay a placement than publish a contradiction I then have to clean out of the source graph.

How I would score a sample article before the retainer

I would ask for two live URLs the firm already placed for a company of our stage, and I would mark them the way a retriever would, not the way a clip book does. On each URL I would look for a sentence that still names the company after the paragraph is cut away, a category a buyer could have typed, and one proof with a number, a quotation, or a source. The GEO paper's roughly 30% to 40% visibility change, in the settings it tested, came from quotations, statistics, and citations. A sample that contains none of those is a sample of the work I would be buying. I would not hope the next article, the one about us, will suddenly grow a fact the previous articles did not have.

Ahrefs' look at 75,000 brands is the reason I would then check where those two URLs sit, not only what they say. Branded mentions lined up with AI Overview presence more closely than classic link metrics in many of their comparisons. I would not convert 75,000 into a formula and apply it to two URLs. I would use it as permission to ignore a domain rating speech. If the sample articles live on domains that never appear when I ask our buyer question today, the firm is adding nodes outside the retrieval set. That can still be worth doing for a human reader. It is a different contract from the one that claims to change answers, and I would price it as the human contract.

OpenAI's October 31, 2024 launch is the date I would use to retire the old screening script. Linked sources inside the chat answer mean the placement is visible as a source, not only as a relationship the firm has with a desk. I would ask who at the publication edits, and I would also ask whether the HTML is what a crawler receives. A beautiful piece behind a wall the crawler never crosses can be a human win and a machine zero. I would want that called out in the proposal so I do not discover it in the third month.

The model table is how I would stop scope from sliding. If I need a year of desk relationships, I buy the retainer and I judge it on access and judgment, with citation as a reported side effect rather than a promised one. If I need a node on a domain the current answers already cite, I buy the placement and I write the extract into the statement of work. GEO and AEO formatting are a third purchase, and only after the claim exists. I have watched firms sell all five rows as one retainer and then deliver the labor row. The five questions I would ask in the meeting are there to make that slide impossible: publications and why they matter for the buyer question, category clarity versus another mention, the extractable sentence, the second and third corroborating domains, and the before-and-after on a frozen prompt. A firm that answers with logo volume has told me which row they actually sell.

I would also look at whether the two sample companies are described the same way on their own sites as in the article. A chain fails when the placement invents a category the company does not use. The firm cannot repair that in copy. I would catch it in the sample, before it is our category being invented.

What I would do

I would pick one buyer question that already produces a shortlist, record the domains the current answers cite, and hire only for placements that can join that set with a claim we can repeat elsewhere. I would cap the first engagement at a handful of articles, each with a written extract: entity, category, proof. I would reread the same question after publication, on a set window, and I would keep the prompt stable so a change in the answer is about the source and not about a new wording.

Scope: AI and technical startups whose buyers ask an engine who the serious vendors are, before they ask a salesperson. This is a poor use of money for a company that still cannot say what it does in one extractable sentence. The firm cannot invent that sentence in a newsroom. I would write it first, on a page we own, and then earn the corroboration.

What would change my mind: a category where the answers buyers receive are still overwhelmingly clicks to our own site, and third-party mentions show no relationship to whether we are named. Ahrefs’ finding on 75,000 brands points the other way for AI Overviews, and the October 31, 2024 launch put linked sources inside the chat answer. If a later, disclosed panel in my category showed owned pages carrying the answers and third-party articles adding nothing extractable, I would hire for the owned page and stop paying for clips. I have not seen that panel. Until I do, I hire for the article a machine can reuse.

FAQ

What would I want a placement to contain?

A named entity, the category a buyer would type, and at least one proof a machine can lift: a number, a quotation, or a cited source. The GEO paper’s roughly 30% to 40% visibility lift, in the settings it tested, came from that kind of material. I would reject a profile that says the company is visionary and never states what it does.

How would I tell a retainer from a placement engagement?

A retainer buys continuity of labor. A placement engagement buys a published third-party page. I would write the success condition into the contract in those words. "Top-tier earned media" is not a success condition. A URL on a domain the current answers already cite, containing a claim I can corroborate on a second domain, is.

Does a famous publication automatically help?

No. Prestige and usefulness inside retrieval come apart. The Ahrefs analysis of 75,000 brands ties branded mentions, more than classic links, to presence in AI Overviews. A precise article on a relevant domain can beat a vague article on a famous one. I would ask the firm to show the extract, not the logo.

When would I refuse to hire anyone?

When we cannot yet state the category and the proof on our own site. An agency placement cannot repair a company the machine cannot resolve. I would spend the first money on that sentence, then hire for corroboration on domains a buyer’s question already retrieves.