Digital PR for AI visibility: earning the mentions that move answers

Publish the best page in your category and you have still only changed one of the sources an AI engine reads before it answers. Citation analyses from Ahrefs and Semrush show AI engines lean heavily on independent pages, reviews, roundups and reference sites rather than brand-owned domains, and Otterly's 2026 AI Citations Report, built on more than a million data points, confirms the same pattern: reference and community sources dominate, with a long fragmented tail behind them. When an assistant recommends tools or vendors, it is synthesizing what third parties say about you at least as much as what you say about yourself.

That is why earned mentions are frequently the fastest lever in AI visibility. A page the engine already retrieves and trusts for your question carries standing accumulated over years; getting your brand named on that page borrows the standing instantly. Publishing your own page, by contrast, means waiting for indexing and retrieval to discover it. Both levers belong in a program, and the owned-content half is covered in how to write content AI engines actually cite. This playbook is the earned half.

The market gap makes the lever bigger: G2 research found 51 percent of B2B tech brands have zero citations across ChatGPT, Perplexity and Gemini. In most categories, nobody is running citation-led PR yet, and the sources the engines trust have not been claimed. That advantage is temporary by nature.

Classic PR ranks targets by masthead prestige: national papers at the top, trade press in the middle, forums nowhere. The AI citation data inverts much of that ladder. 5W Research's 2026 analysis of ChatGPT citations in the U.S. found Wikipedia at 13.15 percent and Reddit at 11.97 percent, together more than a quarter of all citations, while the Wall Street Journal, the New York Times and Bloomberg do not appear in the top 20 at all. Part of the explanation is access: several major papers license their archives or block AI crawlers outright. The result stands regardless of cause: the sources assistants quote are reference sites, communities and the working middle of the web, and the full breakdown is in what ChatGPT actually cites.

The engines also disagree with each other. Perplexity leans much harder on community content, with Reddit its single largest source at estimates ranging from roughly 17 to 24 percent of citations, and one analysis putting Reddit at 46.7 percent of Perplexity's top-10 share; it also skews toward LinkedIn, NIH and G2. Meanwhile cross-platform analyses find only about 11 percent of domains are cited by both ChatGPT and Perplexity. A placement that moves one engine may do nothing for another, so a serious program builds its target list per engine rather than assuming one win travels everywhere.

For a PR budget, the practical translation is uncomfortable and freeing at once: the tier-one placement that impresses your board may move AI answers less than a niche roundup on a site you had never heard of, if that roundup is what the engine retrieves for your buyers' question. Spend against citations, not against prestige.

The method that replaces the media list is simple to describe. Take the twenty or so questions buyers ask before choosing in your category. Run each through the engines your buyers use, and record two things per question: which brands the answer names, and which sources it cites where citations are exposed. Perplexity makes this easy, averaging roughly 8.2 sources per answer, about 3.4 times ChatGPT's count. After a few runs per question, you hold a ranked list of the exact pages and domains that currently write your category's answers.

That list is your campaign. A domain that appears across many of your questions is a strategic target worth a relationship; a single listicle cited for one high-value question is a tactical target worth one precise pitch. Note which cited pages are updateable, since roundups and reviews get refreshed, and an update to an already-cited page is the shortest path in this entire discipline: the page is already retrieved, already trusted, and your mention rides the next crawl.

Done by hand, the collection takes a day and needs re-running as answers drift. Reachroller runs it continuously: every tracked question stores the raw answer text, so the sources behind each answer sit in a report instead of a spreadsheet you rebuild, and each re-run keeps the target list current as engines change their diets. How the tracking works engine by engine is on how it works.

Two of these tiers carry enough weight and enough rules to need their own playbooks. Reddit's AI citation share grew roughly 73 percent in commercial categories through 2025 and 2026, and it is the one channel where the wrong approach gets you banned by moderators and remembered by the community; the participation model that compounds instead is in a Reddit strategy for brands that AI engines respect. Wikipedia, ChatGPT's single largest source, has the strictest door in earned media and real penalties for pushing on it wrong; the rules are in Wikipedia and AI visibility.

The quieter tiers are often the best return per hour. A category roundup on a mid-sized industry site is written by one identifiable author, refreshed yearly, and cited by engines far out of proportion to its human traffic. A complete, honestly reviewed G2 profile feeds a platform Perplexity demonstrably favors. Neither requires a PR agency; both require the pitch to contain something worth citing.

What you pitch matters as much as where. The Princeton-led GEO study found that statistics, quotations and cited sources are what boost visibility inside generative answers, by up to 40 percent in its measurements. That finding is usually read as on-page advice, and it doubles as a pitch strategy: the publications engines cite are the ones full of exactly that material, so the pitch that gets accepted and then quoted is the one carrying it. Original numbers from your own data, a quotable expert claim with a name attached, a verified fact the roundup's author does not have to check twice.

Original research is the strongest asset in this game. One genuine dataset, even modest, gives a dozen niche publications a reason to mention you and gives engines a statistic that travels with your name attached. When an assistant repeats a number, it tends to repeat the attribution too, which is how a single study seeds mentions across answers for months. The citation studies quoted throughout this post are themselves demonstrations: Otterly's report and 5W's analysis earn their firms mentions in every article that uses their numbers, including this one.

For the pitch itself, precision beats volume. One author, one page the engine cites, one concrete suggested addition with the facts pre-verified and a source for each. Framed that way, the pitch improves a page the author maintains, with material that makes it more citable, and authors say yes to improvements far more readily than to favors. Acceptance rates on that pitch embarrass anything a spray-and-pray press release achieves.

Every shortcut in earned media has an AI-era failure mode. Astroturfing, the fake-account praise campaign, is the worst of them: communities detect it, moderators ban for it, and the detection thread itself becomes content engines can retrieve, permanently attached to your brand name. Undisclosed paid placements fail more slowly but the same way; the sites that sell quiet inclusions are the sites whose citation weight erodes as engines get better at source quality, and disclosure-dodging is a story tech journalists like writing.

Press-release syndication deserves a specific warning because it looks like reach. Blasting a release across wire services creates hundreds of copies of one page on domains engines have every reason to discount, and none of it is an independent voice saying your name. The citation studies show engines favoring pages where someone actually evaluated something. A wire copy evaluated nothing.

The general principle: earned mentions work because they are costly signals, evidence that an independent party chose to name you. Every tactic that fakes the choosing removes the thing that made the mention valuable. Slow is the only speed that compounds here, which is exactly why starting before your competitors matters.

PR has always had a measurement problem, and AI answers finally give it a scoreboard: either the assistant names you for the question you targeted, or it does not. The discipline is in reading that scoreboard honestly. SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brand list, so a single triumphant screenshot after a placement proves nothing. Repeated runs, stored answers and trend lines are the standard, the same protocol as for owned content.

A realistic timeline: when your mention lands on a page the engine already cites for the question, movement can show within days to two weeks of the page's next crawl. A fresh article needs indexing in Google and Bing first, then retrieval pickup, so give it the same one to two weeks you would give your own page. Expect uneven movement across engines, because their source diets barely overlap.

Reachroller gives the program its scoreboard. Tracked questions re-run on schedule, every answer stored as raw text you can open, mentions counted only when your brand name literally appears, and branded questions excluded so the trend is honest. When a placement lands, the recheck shows whether the answer flipped, and the generated fix pages work the owned side of the same questions so both levers report to one number. Starter is $29 per month, with a trial of three days and 50 credits, no card, which covers a full baseline report on your own brand: the before picture every PR campaign should start from.