How I Choose a PR Firm

I choose a PR firm by whether it can show my brand named, with a fact a buyer could check, inside the answers to the queries I need to own. A firm that uses a language model to build lists, draft pitches, and file clips faster has improved its labor. I would pay for that labor only after the firm can show what the coverage does once an engine reads it.

The comparison I would read before a retainer conversation is AI PR Agency vs Traditional PR Firm: Which to Hire in 2026. That page is the deeper comparison. This essay is how I would use it as a founder: which questions I would ask, what evidence I would accept, and what would make me walk away.

Two kinds of work are being sold under one label

One kind of firm uses software to do the old job faster. Media research, first-draft pitches, monitoring, and reporting get quicker. The output is still a placement. The report is still impressions and a share-of-voice number. The reader the firm is writing for is still the journalist, and the success story is still "we got you in."

The other kind of work treats the placement as raw material. The job is to make the brand appear, get cited, and get recommended when someone asks ChatGPT, Perplexity, or Gemini a question the company should own. I coined Machine Relations in 2024 for that job. The order I would hold a firm to is the one on Machine Relations. The terms I use for the outcome, including the difference between a mention and a citation, are in the glossary.

A survey of 858 marketing and PR leaders, reported in March 2026, found that 63.5% say AI-driven search has already influenced their PR strategy. Only 21.8% have defined authority, expertise, and trustworthiness optimization as a strategic priority. That gap is the market I would be hiring into. Most firms know the reader changed. Few have changed the thing they measure.

A placement has to survive the machine that reads it

Earned media is the material engines use. Zeno Group's 2026 media landscape analysis put 85% of generative engine optimization results on earned media sources. The placement still has to be extractable. When a buyer asks which firm to hire, the engine retrieves claims it can lift: named entities, specific facts, and a source it already treats as usable. A famous logo on a vague profile does not give it that.

Concentration is real, and it is easy to misuse. 5W's AI Platform Citation Source Index found that the top 15 domains capture 68% of AI citation share in aggregate across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. The same release describes different habits by engine: ChatGPT concentrates on Wikipedia, Reddit, Forbes, and Business Insider, while Perplexity rewards primary sources and B2B authority sites. Read category by category and the picture changes again. Machine Relations Index research finds the ten most-cited domains in a category hold 7.0 to 24.2 percent of that category's citations, and 49 to 304 distinct domains are needed to reach half. Journalism accounts for 27% of all AI citations overall, and 49% on time-sensitive queries.

I would use those two pictures together. The aggregate list tells me a short set of domains carries a large share of citations across engines. The category picture tells me a pitch list built only from that short set will miss the domains that actually carry my category. A firm that can name only the aggregate top 15, and cannot name the domains that show up for my buyer's question, is optimizing a headline.

Burson's research across 55,000 believability scores, 85 companies, and 7 major AI answer platforms makes the quality point mechanical. Fact-based claims about innovation, products, and workplace culture outperformed subjective qualities like leadership and governance. Burson CEO Corey duBrowa said it plainly: visibility is not the same thing as credibility. I would hire for the second. A clip that makes the brand visible to a human and unbelievable to a machine is an unfinished placement.

Five questions I would ask before I sign

I would take the comparison page into the meeting and ask these five questions against it. Each one maps to a number above.

Do you measure citation, or impressions? Only 13.1% of organizations in the 858-leader benchmark share earned media results with their sales team. If the sample report stops at clips and reach, the firm is measuring the input. I would ask them to show whether the brand appears in ChatGPT, Perplexity, or Google AI Mode for a query list I bring, with the claim the answer used.

Do you write the placement so a machine can lift a fact? Entity clarity, a named number, and a source the reader can open are the difference between a story and a citable claim. Burson's scores say fact-based claims outperform subjective praise. I would want the draft to show the fact before I approve the pitch.

Can you show movement, on a query I name, across more than one date? Share of citation is the figure I would put next to share of voice. One screenshot is a moment. I would want the same queries run again after the placement, on the engines my buyer uses, with the before and after in the same format.

Do you have a plan across domains, or one logo? The 5W index is explicit that engines do not share one source list. I would want owned pages, earned placements, and third-party validation aimed at the surfaces where my buyers ask, including the category domains that sit outside the aggregate top 15. A single Forbes hit is one input. It is a weak architecture.

Are you building a record, or running a campaign? The same benchmark found that 84.1% believe PR will play a larger role in sales enablement within two years, 50% still rely on traditional metrics, and only 21.4% have formal AI policies. Only 6.1% integrate PR into sales enablement. I would rather hire the firm that treats each placement as a claim I can still point to in six months than the firm that celebrates a spike and moves on.

What I am buying	What I expect to see	What would tell me I bought speed
A citable evidence asset	A fact, an entity, and a source URL	A placement with no liftable claim
Citation movement	The same queries, before and after	Impressions and a clip book
Evidence across domains	Category domains plus the obvious logos	One outlet, repeated
A record that compounds	Claims I can still retrieve months later	A campaign with an end date

How I would read the first sample report

I would not start with the logo on the cover. I would start with four lines.

The query, written the way a buyer asks it. The engine. The date. The sentence the engine actually produced, with the URL it cited if it cited one. A report that opens with reach and closes with a clip has skipped the line I would pay for.

Then I would sort the sentences into the three outcomes the survey's gap is really about. A mention means the name appeared. A citation means a source was used. A recommendation means the answer selected the company. The 858 leaders who say AI search has influenced strategy, and the much smaller share who have made authority a priority, are often collapsing those three into one slide. I would keep them apart, because a mention with a subjective compliment and a citation of a fact-based claim are different assets. Burson's scores, across 55,000 of them, already say which kind of claim travels.

I would also check the domain against both pictures of concentration. If every cited URL sits in the aggregate top 15, I would ask where the category domains are. The category math says the ten most-cited domains hold 7.0 to 24.2 percent, and reaching half of the citations can take anywhere from 49 to 304 domains. A report that cannot name a domain outside the famous set is telling me the firm pitched the chart, not the category.

Journalism's 27% of citations, rising to 49% when the query is time-sensitive, tells me when a news placement earns its cost. On a time-sensitive buyer question I would weight a reported fact in a newsroom. On a stable category question I would weight the domain that keeps getting cited in that category, even when it is absent from the aggregate top 15. The firm should be able to say which of those two jobs this placement was bought to do.

The decision

I would sign a firm that can answer the five questions with artifacts: a query list, a source list that includes my category and not only the aggregate top 15, a sample of fact-based claims, and a report that separates mention, citation, and recommendation. I would pass on a firm whose entire proof is that it drafts pitches faster.

Scope: the next retainer, the buyer queries I write down before the kickoff, and the engines named in the work I am buying, at minimum ChatGPT, Perplexity, and Google AI Mode. The comparison page stays the reference for how the two models differ. My file is the query list, the source list, and the before-and-after.

What would change my mind: if a firm I had filed as "faster pitching" showed, on that same query list, that the brand moved from absent to named with a checkable claim after its placements, I would reopen the conversation. If a firm I had filed as citation-capable delivered clips and could not show the before-and-after on the queries in the contract, I would stop at the first renewal point. The 63.5% who say AI search has already changed their strategy, set against the 21.8% who have made authority a priority, is the reason I would write the test into the contract instead of hoping the market sorts itself out.

FAQ

Does the firm need to use AI for me to hire it?

I would hire on the outcome. The 858-leader survey says most organizations already feel AI search in their strategy, and far fewer have made authority a priority. A firm can use software for lists and drafts. I would still require citation movement on the queries I own.

Is a Tier 1 placement enough?

A recognizable placement is useful when it carries a fact an engine can lift. Zeno's figure puts 85% of generative engine optimization results on earned media, and Burson's 55,000 scores say fact-based claims outperform subjective praise. I would judge the placement by the claim inside it and by whether the domain is one engines cite in my category, where the top ten domains hold 7.0 to 24.2 percent and dozens to hundreds of domains make up the rest.

How often would I ask for proof?

I would want a baseline before the work starts and a second read after the placements are live, on the same queries. One appearance is a moment. The 5W index and the category ranges both say engines and categories disagree, so a single engine on a single day is too thin for a renewal.

Would I build this inside the company instead?

I would, if the alternative is a retainer that reports clips. The same benchmark says only 6.1% of organizations integrate PR into sales enablement. Building the query list, the source list, and the before-and-after is a founder job whether or not a firm executes the pitches. I would hire the firm that can operate on that record.