Someone pushed back on me recently. They said: AI absolutely cites press releases. I ask it to research a product and it pulls directly from the vendor's newsroom.
They were right. And the conventional wisdom circulating in marketing circles, that press releases are dead to AI, deserves a harder look. When someone asks an AI assistant what a company announced, what a product does, or when a feature became available, a well-structured press release on the vendor's own domain is often the most authoritative source available. It is dated. It is attributed. It comes from the source. Those are exactly the signals that AI retrieval systems are designed to trust.
So why do so many marketers believe their press releases are being ignored? Because in many cases they are. But the cause is not the algorithm. The cause is an infrastructure problem that the communications team usually does not own and the technical team usually does not know matters.
What the data actually shows
The pessimistic case rests on real numbers. Analyses of millions of citations across ChatGPT, Gemini, and Perplexity show syndicated press releases, the versions distributed through wire services and picked up on aggregator sites like Yahoo Finance, accounting for a fraction of a percent of AI citations. Earned editorial content dominates, typically north of 80 percent of all news-related citations in large-scale studies. That is a genuine gap, and it would be dishonest to wave it away.
But the data tells a more specific story than "AI ignores press releases." It tells you that syndicated copies of press releases, stripped of context and hosted on third-party domains, perform poorly. Press releases hosted on the company's own primary domain, with proper crawl access, perform measurably better. One analysis of ChatGPT citation patterns found own-domain newsroom releases cited at rates approaching 18 percent in certain product research query categories. That is not nothing. The question is whether your press release infrastructure qualifies for that category, or the syndicated-copy category.
The robots.txt mistake from 2023 that is still costing you
In 2023, a wave of companies and publishers added broad blocks on AI crawlers to their robots.txt files. The concern was legitimate: large language model companies were scraping the web to build training datasets, and content owners wanted control. The execution was blunt. A blanket disallow rule aimed at GPTBot, OpenAI's training crawler, often caught every AI-related bot, including the retrieval agents that answer user research questions in real time.
Those are not the same thing. A crawler pulling your content to train a model is a business and legal question. A crawler pulling your content because a potential customer just asked "what does this product do" is a demand generation question. Many companies have been blocking the second type while trying to stop the first, and nobody on the communications team was in the room when it happened.
The fix requires distinguishing between bot types. Training crawlers include GPTBot and Google-Extended when used for model training. Retrieval agents, the ones serving live user queries, operate under different user-agent strings. A clean robots.txt that explicitly allows major search crawlers and legitimate retrieval agents, while optionally restricting training-specific bots, is the right configuration for a company that wants its announcements found. Test any changes in Google Search Console before assuming they work.
Where your press release should live and how it should be written
Wire distribution alone is not a visibility strategy. Treat PR Newswire and BusinessWire as one channel for journalist outreach, not the primary home of your announcement. The canonical version belongs on your own domain, under a consistent URL path like /news/ or /press-releases/, rendered in clean static HTML. JavaScript-heavy newsrooms are a real problem because many AI retrieval crawlers do not execute JavaScript. Your content may exist and still be invisible.
Maintain a full archive. AI systems used for product and competitive research pull historical context, not just the latest announcement. A company that deletes old press releases to keep its newsroom tidy is erasing the record that AI systems use to understand what the company does and has done.
Structure matters more than length. Lead with the news in the first paragraph: who, what, when, where, why. Use short paragraphs and clear subheads. Embed verifiable facts, specific numbers, and named quotes. These are the elements that AI systems extract and cite. Hype-heavy intros ("Company X, a leading innovator in next-generation solutions, today announced...") delay the factual content and reduce parseability. AP-style factual prose outperforms marketing prose for AI citation purposes, which is also true for journalists, so the incentives align.
Add structured data. Implementing Schema.org markup, specifically the NewsArticle or Announcement type, via JSON-LD helps AI systems parse key facts, dates, entities, and relationships more reliably. This is a one-time implementation on your newsroom template that pays ongoing dividends for every release you publish.
Technical access is necessary but not sufficient
Fixing your infrastructure gets you into the game. Winning requires building on top of it. AI citation patterns show that editorial coverage from journalists and analysts who link back to your primary release creates compounding authority. The release is the foundation. Third-party validation is the multiplier. A technically accessible, well-written release that earns no editorial pickup will still underperform compared to the same release that generates coverage on three credible outlets that link back to it.
This means the press release is not the end of the workflow. It is the anchor for an amplification strategy. Pair each major announcement with supporting owned content, a blog post, a landing page, a detailed FAQ, that expands on the announcement and builds topical depth on your domain. Owned content and press releases pointing in the same direction, on the same domain, create the kind of consistent entity signal that AI systems use to build their understanding of what your company does and how authoritative it is on a given topic.
Measure differently
Traditional PR metrics, media impressions, pickup volume, share of voice in trade outlets, do not tell you whether your announcements are appearing in AI-generated answers to product research questions. That requires a different kind of monitoring. Query the AI tools your customers use with the real questions they would ask about your product category. See what gets cited. See whose press releases appear and whose do not. The companies showing up consistently are not necessarily doing better PR. They may simply have better infrastructure.
The press release has not lost its authority as a primary source. For product research queries, it may have more structural value than ever, because it provides exactly what AI retrieval systems need: a dated, attributed, factual statement from the official source. The companies that fix their access, structure their content for extraction, and back it with earned coverage will show up in those answers. The companies that do not will keep concluding that AI ignores press releases, when the real issue is that AI cannot find them.
When did your communications team last sit in the same room as your web or SEO team and audit your newsroom's crawl settings together? If the answer is never, you have probably been blocking demand generation without knowing it. The audit takes an afternoon. The cost of not doing it compounds every quarter.