> ## Content Index > Fetch the complete content index at: https://intelligence.artificialrealitymedia.com/llms.txt > Use this file to discover other available public pages before exploring further. # How Sam Polstein builds a media list, and where AI fails him - URL: https://intelligence.artificialrealitymedia.com/how-sam-polstein-builds-a-media-list/ - Published: 2026-09-02T12:08:00.000Z - Updated: 2026-09-02T12:08:00.000Z - Description: A deep tech comms lead on why AI cannot build a media list: the signals are unpublished, recent coverage disqualifies, and a three-week lag is useless. - Author: Galina Fendikevich - Tags: Public relations, #For AI Buyers, #For AI Companies, #What's not working, #Import 2026-09-03 13:12 AI is bad at building a media list because the signals that decide the list are not published anywhere. Sam Polstein, who leads communications and content for deep tech startups at HAUS, told Artificial Reality that a three-week lag in the data feeding a model makes it useless for this work, and that what a reporter wrote six months ago is largely irrelevant. His verdict on current models: "For whatever reason, I find AI to be pretty bad at this." ## Why does the media list decide the campaign? Because the wrong recipient makes the pitch irrelevant, however good the pitch is. Sam Polstein puts list building ahead of writing in the order of things that matter: > "Building a media list in comms is incredibly important. Who you're reaching out to is — the same way context engineering is crucial to the output of using an LLM — building a media list is the hard work that makes or breaks a PR campaign. Because if you're reaching the wrong people, it doesn't matter how good your pitch is." The failure it produces is one every publicist recognises: > "The amount of times over my career that I've gotten responses from journalists that are like: okay, fine, but I don't cover that. Or: cool, but my editor's not gonna let me talk about this." ## Why can't a model do this from public data? Because the task is prediction, not retrieval, and the determining inputs are unpublished. Sam Polstein expected the opposite going in: > "You would think LLMs \[are\] really good at organizing all the information that exists in the public domain. Reporters have tons of content that they've written ... You would think there would be super powerful tools and I wouldn't be manually — or my team wouldn't be manually — creating media lists. We find that's just not true." What the work actually involves, in his words: > "Building a media list has a lot of tangential thinking involved. It's a little bit more of an art than a science. It's figuring out not what a journalist has covered before — because often if they've already covered it, they'll be like: I already wrote that story, sorry man." > "You're trying to base \[it on\] the vibe of their beat and what their editors are telling them to write, and then helping them with their next story that they've never written about, usually." Editorial direction, what a reporter is bored of, and where they are quietly moving are not in the archive a model has read. ## Why does recent coverage count against a journalist? Because it usually disqualifies rather than qualifies. A reporter who just filed that story will not file it again, which inverts the logic every similarity engine runs on. Sam Polstein's recency rule is unusually strict: > "I don't care what a reporter's written about six months ago, for the most part. It's: what are the last few stories? Because I'm trying to guess what they're writing next." > "If there's a three week long lag in the information that's feeding into [Claude](https://claude.ai/?ref=intelligence.artificialrealitymedia.com), well, that's basically useless." ## Do [Muck Rack](https://muckrack.com/?ref=intelligence.artificialrealitymedia.com) and [Cision](https://www.cision.com/?ref=intelligence.artificialrealitymedia.com) close the gap? Not in this operator's account. Sam Polstein uses the incumbent databases and describes them as CRMs built for exactly this job, carrying data that is still hard to come by and behind: > "There are tools and platforms that we use like Muck Rack and Cision that offer these sorts of things. They are CRMs for exactly what we're describing ... I think though that that sort of information, it's still not that easy to come by, still somewhat out of date." ## What would he actually buy? A short specification comes straight out of the interview. Stated by the operator: - Answer a plain question about a named journalist — "do you think he'll cover this topic?" — rather than returning a ranked list. - Argue back, the way a colleague would. He explicitly wants pushback, not agreement. - Treat very recent coverage as a **disqualifier**, not a match signal. - Track outlet moves and newly started beats. - Reason tangentially — peer companies rather than direct competitors, and thematic adjacency. This is Artificial Reality's framing of what Sam Polstein describes as "more of an art than a science." - Operate on information no more than a few weeks old. - Combine public data with the agency's own accumulated judgment about each reporter. That is Galina Fendikevich's formulation, which Polstein endorsed. ## What should a builder take from this? That the valuable object is the objection, not the suggestion. Every product in this category, incumbent and AI-native, is architected to produce a confident recommendation. Sam Polstein already gets confident recommendations and does not trust them. What he describes wanting is the Slack thread where a colleague pushes back with what that reporter just wrote and where they just moved. Artificial Reality's analysis: a recommendation engine built on models he calls too effusive will validate bad targets, and a pitch to the wrong journalist costs the relationship, not just the send. The demand is also unusually well qualified. Polstein builds his own dashboards in [Claude Code](https://claude.com/product/claude-code?ref=intelligence.artificialrealitymedia.com) and could not build this. The incumbents charge five figures a year and have not solved it. The frontier models are, in his direct testing, bad at it. If you are building here, the question to answer first is not how to rank journalists better. It is whether your system will tell a paying user that they are wrong. ## Where this comes from S1E4: He runs PR for deep tech startups — the full interview with Sam Polstein. Listen or watch: [YouTube](https://www.youtube.com/watch?v=Oo5rT9aS2RQ&ref=intelligence.artificialrealitymedia.com), [Spotify](https://open.spotify.com/episode/3QYcgc5ezkSMvH36WJd7qv?ref=intelligence.artificialrealitymedia.com) or [Apple Podcasts](https://podcasts.apple.com/us/podcast/he-runs-pr-for-deep-tech-startups-if-your-agency-isnt/id6780699025?i=1000773200320&ref=intelligence.artificialrealitymedia.com).