How to Write a Press Release That AI Engines Actually Cite

Write press releases like structured data: one atomic fact per sentence, named entities, attributed quotes, and stats up top. Distribute through a trusted wire plus your own newsroom, since AI engines cite the pickup domain, not just your site. Add FAQPage and NewsArticle schema, then track citation lift in Claude, ChatGPT, and Perplexity for 30 days.

*By Josh Bernstein | August 17, 2026*

Most press releases are still written for a reporter who no longer exists: the mid-tier trade journalist who used to skim the wire, lift a paragraph, and run it. That reporter has been partly replaced by a retrieval system. ChatGPT alone now serves 900 million weekly active users, according to OpenAI's February 27, 2026 announcement, and a meaningful share of those sessions ask exactly the kind of question a press release is built to answer: who raised money, who shipped what, who partnered with whom, and what the number was.

The problem is that most releases still bury the number in paragraph three, behind a headline built for SEO keyword density instead of extractable fact. AI engines do not skim for tone. They parse for entities, numbers, dates, and quotable sentences they can lift with attribution intact. A press release built the old way gets indexed and ignored. A press release built for extraction gets pulled into ChatGPT, Claude, Gemini, and Perplexity answers for months after the news cycle ends. This piece walks through exactly how to build the second kind, and how to prove it worked.

Why AI Engines Treat Press Releases as Primary Sources

Press releases occupy a strange, useful position in the information supply chain. They are dated, attributed, entity-dense, and published on domains that wire services have spent decades getting indexed and trusted. That combination, timestamp plus named source plus structured facts, is precisely what a retrieval-augmented model looks for when it needs a defensible answer about something that happened recently.

This is also why press releases have become more valuable for AI visibility even as they've become less valuable for traditional media pickup. Journalists rewrite releases into stories; AI engines often summarize the release itself, especially for company-specific facts like funding amounts, executive appointments, or product launches that would otherwise require the model to hedge with "I don't have current information." A well-built release removes the model's excuse to hedge.

It also matters that AI engines are increasingly the front door to information that used to route through a search results page. Ahrefs found that AI Overviews reduced click-through rate for the position-1 organic result by 58% as of December 2025, up sharply from a 34.5% reduction the same research measured in April 2025, a trend that is pushing search steadily toward zero-click behavior. Fewer people click through to your site to read the announcement themselves. More of them get the fact secondhand, from whatever source the model chose to trust. This is the broader shift behind why GEO is displacing traditional SEO as the growth channel that matters: your job with a press release is no longer just "get picked up." It's "be the source the model trusts enough to quote."

The GEO-Ready Press Release Structure

A press release built for AI extraction still follows the classic inverted-pyramid shape journalists have used for a century. What changes is the density of extractable facts inside that shape, and how deliberately each section is written to be lifted as a standalone unit.

The core anatomy looks like this:

1. Headline: names the entity, the action, and ideally a number, in that order. "Acme Raises $40 Million Series B to Expand AI Infrastructure" beats "Acme Announces Major Funding Milestone."
2. Dateline and lede: the first two sentences answer who, what, when, where, and why, with the city and date functioning as a machine-readable recency signal.
3. Atomic facts: three to six standalone sentences, each containing exactly one verifiable claim, a number, or a named entity, written so each sentence still makes sense if quoted alone and out of context.
4. Attributed quotes: one or two quotes from named, titled spokespeople that add interpretation or forward-looking context the facts alone don't cover.
5. Boilerplate: a tight "About [Company]" paragraph that restates the entity's category, founding year, and scale, which doubles as an entity-disambiguation anchor for models that don't already have your company mapped.
6. Contact and schema layer: media contact info plus NewsArticle, Organization, and (where relevant) FAQPage schema so the same facts are machine-readable twice, once in prose and once in structured markup.

*Each layer of a GEO-ready release restates the same core facts in a different format, prose, quote, and markup, so an AI engine can extract them regardless of which layer it parses.*

Structuring Facts, Stats, and Quotes AI Can Extract

The single highest-leverage change most companies can make to a press release is also the simplest: put the number in the first 200 words, state it once cleanly, and never make a reader (or a model) hunt for it across three paragraphs of context.

This isn't just intuition. A 2024 study out of Princeton and Georgia Tech (published at KDD, testing roughly 10,000 queries across nine content optimization tactics) found that adding statistics to a page improved its visibility in generative answers by 41%, adding direct quotations improved it by 28%, and citing authoritative sources produced gains as high as 115% for pages that started ranking lower, around position 5. Pages that already ranked first benefited the least from these changes, which tells you something important: GEO tactics matter most for content that isn't already the obvious default answer, which describes almost every press release on the day it publishes.

Applied to a press release, that research translates into a short set of concrete habits:

• Lead with the number, not the narrative. "Acme reduced onboarding time by 63%" belongs in sentence one or two, not buried after two paragraphs of scene-setting.
• Name every entity explicitly. Say "Acme Corporation," not "the company," on second and third reference too. Models resolve pronouns imperfectly across long documents; repetition costs you nothing in a release.
• Attribute every claim to a source. "According to Acme's Q3 customer data" or "per an independent Forrester analysis" gives the model a citation chain it can repeat, rather than a bare assertion it has to either drop or hedge.
• Write quotes that stand alone. A quote that only makes sense with the paragraph before it won't get lifted cleanly. Write it as a complete thought a model could drop into an answer with zero surrounding context.
• Use plain numerals, not spelled-out numbers or vague ranges. "40%" extracts more reliably than "forty percent" or "nearly half."

None of this requires bloating the release. If anything, it argues for cutting adjectives and adding specificity. A release that says a product is "revolutionary" gives an AI engine nothing to cite. A release that says the product cut processing time from 12 minutes to 90 seconds gives it a fact worth repeating in a comparison answer six months later. If your release is built around proprietary research or benchmark data rather than a company announcement, the same stat-first, quote-ready structure is what lets a single original data study earn dozens of downstream citations well after the initial news cycle ends.

Wire Distribution vs. Owned Newsroom: Where Citations Actually Come From

Distribution strategy for GEO looks different from distribution strategy for traditional PR, mostly because the destination that matters most isn't a journalist's inbox, it's whichever domain ends up hosting the version of your release that an AI engine's retrieval layer decides to trust. That's rarely a single channel. It's usually a combination of a wire service placement, your own newsroom page, and whatever earned or community pickup happens in the days after.

The strategic implication is that no single channel is sufficient. Wire distribution buys you a fast, structurally clean, high-authority version of the release that AI crawlers can find quickly. Your owned newsroom buys you permanence and control over the canonical facts once the wire cycle passes. Earned trade pickup and community discussion, especially on Reddit and LinkedIn, are what actually shape how models describe your brand months later, because those are the sources engines lean on hardest for anything resembling an opinion, comparison, or recommendation rather than a bare fact. Building a durable version of that owned home is its own discipline; the same structural logic that makes a content hub work as a citable source applies directly to a newsroom page.

Earning Pickups on Domains AI Trusts

There is no single dominant source AI engines pull from, and brands that plan around one platform get blindsided when the mix shifts. An analysis of 30 million sources across ChatGPT, Google AI Mode, Gemini, Perplexity, and AI Overviews found Reddit ranks as the single most-cited domain in AI-generated answers overall, followed by YouTube and LinkedIn. ChatGPT leans toward Wikipedia, Reddit, and editorial sites like Forbes, while Perplexity emphasizes Reddit, LinkedIn, and G2, especially for B2B queries. Put together, community and earned media routinely outweigh brand-owned domains in the citation mix, which is a very different picture than the SEO-era assumption that owning the top organic result was enough.

That changes what "earning a pickup" means for a press release. It's no longer just about which reporters cover the news. It's about whether the release generates discussion on the platforms models actually pull from:

• Seed the conversation, don't just broadcast it. After a release goes out, a founder or spokesperson answering questions in a relevant subreddit or LinkedIn comment thread creates exactly the kind of attributed, discussion-format content that Reddit and LinkedIn citations are built from.
• Pitch trade press that AI engines already treat as category authorities, not just outlets with the biggest general audience. A mid-size vertical publication that shows up repeatedly in AI answers for your category is worth more than a general-interest outlet that doesn't.
• Make the release itself quote-friendly for third parties, so bloggers, analysts, and community members who reference it can copy an accurate stat instead of paraphrasing it into something the model can no longer verify against your original.
• Track which domains actually pick up the story, not just how many outlets ran it, and prioritize follow-up outreach toward domains with a track record of AI citation, not just circulation.

This is also where entity clarity compounds. A press release that consistently uses your full legal or trademarked name, ties it to a stable set of facts (founding year, headquarters, category), and gets referenced the same way across wire, trade press, and community posts makes it easier for a model to resolve "Acme" as one confident entity rather than three ambiguous fragments. That's the same underlying mechanic behind building a stronger brand entity footprint more broadly, and it compounds every time you publish.

Dateline and Entity Signals That Anchor the Release

Two structural elements do more anchoring work than most PR teams realize: the dateline and the schema layer underneath it.

The dateline (city, state or country, and date) is a recency signal a model can parse without ambiguity. It tells the retrieval layer this fact is timestamped, not evergreen marketing copy, which matters enormously for any query where recency changes the correct answer, like "who just raised funding in the AI infrastructure space" or "what did Acme announce this quarter." Releases that drop or bury the dateline lose that recency anchor entirely.

Schema markup does the same job in a machine-readable format, and it's worth being precise about what it can and can't do in 2026. Google removed HowTo rich results from search back in 2023, and it fully removed FAQ rich results from Google Search as of May 7, 2026. That means neither schema type produces the visual expandable snippet in a Google results page anymore, and you shouldn't expect a SERP-position lift from adding them. But FAQPage and HowTo remain fully valid schema.org types, and they still do real work helping AI engines parse and extract structured question-and-answer content or stepwise facts from a page, independent of whether Google renders a rich result for it. For a press release specifically, pairing NewsArticle and Organization schema with the visible dateline, and adding FAQPage markup for a short "quick facts" block, gives an AI crawler two clean, structured paths to the same facts instead of one.

Named entities matter just as much outside the schema layer. Every person quoted should include a full name and title on first reference. Every product should be named consistently with how it's named on your website and in prior releases, since inconsistent naming across documents is one of the most common reasons models describe a product incorrectly or attribute a feature to the wrong entity. This is the same discipline behind strengthening a knowledge-graph presence more broadly: consistency across every surface, not cleverness on any single one, is what earns a model's trust in a fact.

Measuring Citation Lift From a Release

A press release that gets picked up by ten outlets and cited by zero AI engines has still done half its job. The other half, tracking whether the release actually changed what ChatGPT, Claude, Gemini, and Perplexity say about you, requires treating the release as an experiment with a before-and-after, not a one-time broadcast.

The practical approach looks like this:

1. Establish a baseline before you publish. Run the exact prompts a prospect or journalist might ask (about your funding, your product category, your latest milestone) across the major engines and log whether and how you're currently mentioned.
2. Publish and distribute on a known date, so you have a clean before/after boundary, and note the wire and owned-newsroom URLs separately since they may get cited independently.
3. Re-run the same prompts on a schedule, at day 3, day 14, and day 30, since wire pickups tend to surface in citations faster than earned trade coverage or community discussion does.
4. Log which specific domain got cited, not just whether the fact appeared, so you can see whether it was the wire placement, your newsroom, or a piece of earned coverage that the model actually pulled from.
5. Compare citation accuracy, not just presence. A model that cites your funding round but gets the amount wrong is a different problem than a model that doesn't mention it at all, and it usually points back to inconsistent numbers across your own published sources.

*Citation lift from a press release rarely peaks on day one; wire pickup drives the first spike, but trade and community pickup usually drive the citations that persist.*

This kind of tracking used to require manually prompting each model by hand, which doesn't scale past a handful of releases a year. It's also why a dedicated GEO monitoring layer has become a normal line item rather than a nice-to-have: Sitecore's $225 million acquisition of Scrunch in June 2026 and Profound's $96 million Series C earlier that year (at a reported $1 billion valuation) both signal that tracking brand visibility inside AI answers is now treated as core marketing infrastructure, not an experiment. Whatever tool you use, the report you want out of it is the same one described in a solid GEO reporting stack: citation presence by engine, citation accuracy, and which specific domain earned the mention, tracked release over release so you can see which structural changes actually move the number.

A Pre-Publish Checklist

Before a release goes out the door, it's worth running it against a short list rather than trusting instinct. Teams that skip this step tend to catch the gaps only after the release has already gone out uncited.

• Is the single most important number in the first 200 words, stated in numerals?
• Does every sentence in the body still make sense if read in isolation, out of order?
• Is at least one quote fully self-contained, with no pronoun or reference that depends on the prior sentence?
• Does the boilerplate restate category, founding year, and scale, consistently with the website and prior releases?
• Is the dateline present and correctly formatted?
• Is NewsArticle, Organization, and (where relevant) FAQPage schema attached to the owned newsroom version?
• Is there a plan to distribute beyond the wire, into trade press pitches and relevant community or LinkedIn conversation?
• Is there a baseline AI-engine check logged before publish, so lift is measurable afterward?

Releases that pass all eight are rare, and that's precisely the gap worth closing. Most competitors are still optimizing the headline for a journalist who skims once and moves on. A release built for extraction keeps working for a model that gets asked the same question, in some form, for months.

Frequently Asked Questions

Do wire services like PR Newswire and Business Wire still matter for AI citations?

Yes. Wire services remain one of the fastest ways to get a release onto a high-authority, consistently structured domain that AI crawlers already index frequently, which makes them useful for early citation. They shouldn't be the only channel, though, since owned newsroom pages and earned or community pickup tend to drive the citations that persist past the first few weeks.

How soon after a press release goes out do AI engines start citing it?

Wire pickups can surface in AI answers within days to two weeks, but citations sourced from trade press, your own newsroom, or community discussion typically take one to four weeks to show up consistently. Running baseline and follow-up prompt checks at day 3, day 14, and day 30 gives a realistic picture of when lift actually happens.

Should I still use FAQ schema on a press release page?

Yes, though not for the reason it used to matter. Google fully removed FAQ rich results from Search as of May 7, 2026, so it no longer produces the visual snippet in Google's results page. FAQPage remains a valid schema.org type, and it still helps AI engines parse and extract structured question-and-answer content from the page.

What is the ideal length for a GEO-optimized press release?

There's no fixed word count, but density matters more than length. A tightly written 400 to 600 word release with atomic, standalone facts extracts better than a padded 1,000 word version with the same number of actual claims buried in narrative language.

Can a press release outrank an established competitor's page in AI answers?

It's possible, and the research suggests it's actually easier for content that isn't already the default answer. The Princeton and Georgia Tech GEO study found that citing authoritative sources produced gains as high as 115% for pages starting around position 5, while pages already ranking first saw the smallest benefit from the same tactics, meaning a well-built release has real room to gain ground precisely because it isn't the incumbent.

How do I measure whether a press release actually earned AI citations?

Log a baseline set of prompts related to the announcement before you publish, then re-run the same prompts across ChatGPT, Claude, Gemini, and Perplexity at set intervals afterward, tracking which specific domain got cited and whether the cited facts were accurate. Purpose-built visibility tools can automate this across many releases and prompts at once rather than requiring manual checks per model.

Does Reddit or LinkedIn discussion after a release help AI citation?

Yes, often more than people expect. Reddit is the most consistently cited domain across major AI engines, and LinkedIn ranks among the most-cited sources for B2B and brand-reputation queries specifically, so genuine discussion on either platform after a release can shape how models describe your brand well after the original wire pickup fades.

What's the biggest mistake brands make with press releases for AI search?

Burying the core fact in narrative language instead of stating it plainly and early. Vague, adjective-heavy claims like "industry-leading" or "revolutionary" give an AI engine nothing concrete to extract or quote, while a specific number, named entity, or attributed statement gives it something worth repeating.