Press release GEO in 2026: writing a release that ChatGPT, Perplexity and Gemini cite

Key takeaways

• 7 percent of people use an AI chatbot for news each week, 15 percent among under-25s, across 48 markets. Reuters Institute for the Study of Journalism (17 June 2025)
• 8 percent of Google visits with an AI summary ended in a click on a result, against 15 percent without one; 18 percent of searches showed a summary. Pew Research Center (22 July 2025)
• Up to 40 percent more visibility in generative engine responses from adding citations, quotations and statistics, in the paper that named the discipline. arXiv 2311.09735 (v3, 28 June 2024)
• Search crawlers are not training crawlers. OAI-SearchBot, PerplexityBot and Claude-SearchBot decide citations; GPTBot, ClaudeBot and Google-Extended only govern training. OpenAI developer documentation (2 September 2026)
• No special markup. Google says a page needs to be indexed and eligible for a snippet to appear in AI Overviews, nothing else. Google Search Central (10 December 2025)
• ChatGPT is used by 47 percent of journalists, Gemini by 22 percent, in Muck Rack's 2026 survey of 897 journalists. Muck Rack (19 March 2026)

What is press release GEO?

Generative Engine Optimization, GEO, is the practice of making a page retrievable and citable by systems that synthesize an answer rather than list links: ChatGPT search, Perplexity, Claude, Gemini and Google AI Overviews. The term comes from a Princeton and IIT Delhi paper presented at KDD 2024, which introduced a benchmark of diverse user queries and found that visibility in generative responses can rise "by up to 40%" when a page adds citations, quotations and statistics (Aggarwal et al., arXiv 2311.09735).

A press release is a natural GEO asset. It carries named entities, a date, numbers, a quotable source and a publisher, which is what an engine needs to attribute a passage. It is also usually published badly for this purpose: no author, no dates in the markup, figures without a link, and a wire copy that outranks the original. This guide fixes the publishing side. For the classic ranking side, see the press release SEO guide.

Why it matters now is measurable. The Reuters Institute's 2025 Digital News Report, based on 48 markets, puts weekly use of AI chatbots for news at 7 percent overall and 15 percent among under-25s, and names ChatGPT and Google Gemini as the examples (Reuters Institute, 17 June 2025). Pew Research Center tracked 900 US adults in March 2025: 18 percent of their Google searches produced an AI summary, and visits with a summary ended in a click on a result 8 percent of the time, against 15 percent without one (Pew, 22 July 2025). Fewer clicks reach the release; the citation has to carry the brand.

How generative engines retrieve and cite

1. Query expansion. Google describes AI Overviews and AI Mode as using "query fan-out" to show "a greater diversity of websites" than a classic result page. One question becomes several searches.
2. Retrieval. Each engine runs its own search crawler. OpenAI documents OAI-SearchBot for "ChatGPT's search features", Perplexity documents PerplexityBot to "surface and link websites", Anthropic documents Claude-SearchBot, and Google uses the normal index: a page must be "indexed and eligible to be shown in Google Search with a snippet".
3. Passage extraction. The engine cites a paragraph, a table row or a list item, not a page. Self-contained passages with a date and a source are the ones that survive this step.
4. Synthesis. Passages from several sources are blended. A brand described the same way across independent pages is easier to name than one described three different ways.
5. Attribution. The answer links back to the passages used. Whether your name appears depends on steps 2 and 3, which is where the 12 patterns below apply.

Which crawlers to allow: the robots.txt table

Most GEO advice skips the only step that is binary. Each operator publishes which user agent feeds its search features and which one feeds training. Blocking the training bot is a policy choice; blocking the search bot removes you from the answers. The role column quotes each operator's own page.

Two details from the same pages. OpenAI says ChatGPT-User is "not used to determine whether content may appear in Search", so allowing it only affects live fetches. Perplexity says Perplexity-User "generally ignores robots.txt rules" because a person asked for the page, which is why the table separates citation eligibility from live answers.

The 12 writing patterns that get cited

These are the patterns shared by the pages that answer engines cited in our September 2026 audit of eight press release queries, checked against the GEO paper's findings on citations, statistics and quotations.

1. Direct answer first. One 40 to 60 word paragraph under the H1 that answers the question in full. No bullet, no link, no preamble.
2. Entity-rich prose. Write "Cision", "PR Newswire", "Accel", not "traditional platforms" or "major investors".
3. Numbered lists with discrete items. One sentence per step. An engine lifts a list item as a fact.
4. Tables with stable columns. The same column names on every comparison page (Plan, Price, Contract, Source) so the extraction is predictable.
5. Explicit comparisons. A section titled "Acme vs Cision" followed by a table serves "X vs Y" questions directly.
6. Dates on every claim. "Pricing as of September 2026." A visible update date under the H1, and dateModified in the markup.
7. Figures with primary-source links. The GEO paper found citations, quotations and statistics to be the additions that moved visibility. Declare them with schema.org citation, defined as "a citation or reference to another creative work" (schema.org).
8. One named author with a page. Google's recommended Article properties include author, author.name and author.url, and Google asks that every author shown on the page also be in the markup (Google, Article structured data). One Person node per page, not two.
9. Clean heading hierarchy. H1, then H2, then H3, no skipped level. H2s phrased as the questions buyers type.
10. Single-topic pages. One page, one entity, one answer. Split merged guides into a hub and spokes.
11. FAQ in visible HTML with matching FAQPage markup. Six to ten questions, 40 to 80 word answers, the first sentence carrying the fact.
12. Descriptive internal anchors. Link the press release distribution guide, the free press release generator and the press release glossary entry with the words that describe them.

The schema.org stack for press release GEO

Google is clear that markup is not a ticket: "There's also no special schema.org structured data that you need to add" to appear in AI Overviews (Google, AI features). What markup does is resolve identity and dates, which is what attribution needs. The stack for a newsroom:

• NewsArticle or Article on every release page, with author, author.url, datePublished and dateModified.
• Organization on the root layout with a populated sameAs array.
• One Person node for the author, referenced by @id from every page rather than redeclared.
• BreadcrumbList on every non-root page.
• FAQPage where the page shows an FAQ, HowTo for procedures.
• citation and mentions arrays to link sources, investors and partners to their own identities.

NewsArticle example for a Series B release

{
  "@context": "https://schema.org",
  "@type": "NewsArticle",
  "@id": "https://example.com/newsroom/acme-series-b#article",
  "headline": "Acme raises 40M USD Series B led by Accel",
  "datePublished": "2026-09-02T09:00:00+02:00",
  "dateModified": "2026-09-02T09:00:00+02:00",
  "inLanguage": "en",
  "mainEntityOfPage": "https://example.com/newsroom/acme-series-b",
  "author": {
    "@type": "Person",
    "@id": "https://example.com#jane-doe",
    "name": "Jane Doe",
    "jobTitle": "Head of Communications, Acme",
    "url": "https://example.com/about/jane-doe",
    "sameAs": ["https://www.linkedin.com/in/janedoe/"]
  },
  "publisher": {
    "@type": "Organization",
    "name": "Acme",
    "url": "https://example.com",
    "sameAs": [
      "https://www.linkedin.com/company/acme",
      "https://www.crunchbase.com/organization/acme",
      "https://www.wikidata.org/wiki/Q123456789"
    ]
  },
  "citation": [
    {"@type": "CreativeWork", "name": "Series B term sheet summary", "url": "https://example.com/newsroom/acme-series-b/terms"}
  ],
  "mentions": [
    {"@type": "Organization", "name": "Accel", "sameAs": "https://www.wikidata.org/wiki/Q2578125"},
    {"@type": "Organization", "name": "Sequoia Capital", "sameAs": "https://www.wikidata.org/wiki/Q1852025"}
  ]
}

Organization with a sameAs graph

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://example.com#organization",
  "name": "Acme",
  "url": "https://example.com",
  "logo": "https://example.com/logo.png",
  "founder": {"@id": "https://example.com#jane-doe"},
  "sameAs": [
    "https://www.linkedin.com/company/acme",
    "https://www.crunchbase.com/organization/acme",
    "https://www.wikidata.org/wiki/Q123456789"
  ]
}

Acme, Jane Doe, example.com and Q123456789 are placeholders and do not exist. The two investors are real: Q2578125 is Accel and Q1852025 is Sequoia Capital, both labels checked through the Wikidata API on 2 September 2026 (Wikidata). Never write a Wikidata identifier from memory; look it up and read the label back. Validate every block with the Rich Results Test, and server-render it: JSON-LD injected by client-side JavaScript is invisible to a crawler that does not execute scripts.

Entity mentions and the sameAs graph

schema.org defines sameAs as the "URL of a reference Web page that unambiguously indicates the item's identity. E.g. the URL of the item's Wikipedia page, Wikidata entry, or official website" (schema.org, sameAs). That is the whole mechanism: an engine that reads your Organization node with a LinkedIn, Crunchbase and Wikidata link can match "Acme" in your release to "Acme" in its knowledge graph, and attach what it already knows. Without it, the engine has a string, and strings are ambiguous.

For the Organization: LinkedIn company page, Crunchbase, the review sites where you have a listing, and a Wikidata item once you qualify for one. For the author: a profile page on your domain and LinkedIn. Inside the release, mentions links investors and partners to their own items, as in the example above, which turns a flat text into a small graph the engine can walk.

How to probe ChatGPT, Perplexity, Claude and Gemini

1. Build a list of 50 to 100 prompts: brand questions, category questions ("best press release tool for a startup"), comparisons, and jobs to be done ("how do I reach French tech journalists").
2. Run each on ChatGPT with search, Perplexity, Claude with web search and Gemini. Save the answer and the citation list.
3. Log the domain, the passage extracted, the date of the source and whether your brand was named. Your citation rate is the share of prompts where you appear.
4. Repeat weekly on the top 10 and monthly on the full list. Compare after each content or markup change.
5. Read your server logs for the user agents in the table above. Crawl frequency on the newsroom is the leading indicator; a bot that never visits cannot cite.

GEO vs SEO vs AEO

• SEO ranks pages on the classic results page. Signals: links, relevance, page experience. Metric: clicks.
• AEO targets featured snippets and voice answers. Signals: question-form headings, short answers. Metric: zero-click impressions.
• GEO targets generated answers. Unit: the passage. Signals: entities, dates, primary sources, one identifiable author, crawler access. Metric: citation rate.

The three compound in one direction only. A page built for GEO ranks and snippets well, because the answer block and the dated figures serve both. A page built only for SEO is rarely citable: the answer is spread over 2,000 words and the author is "the team".

A 30-day GEO plan for a press release newsroom

1. Days 1 to 3. Check robots.txt against the crawler table. Run the baseline probe on 50 prompts. Validate markup on three sample pages. Pick the 10 releases with the most traffic.
2. Days 4 to 10. Rewrite the 10 releases with the answer block, named entities, dated figures with links, and a 6 to 10 question FAQ.
3. Days 11 to 17. Ship the markup: Article with author and dates, one Person node by @id, Organization with sameAs, BreadcrumbList, FAQPage. Server-render every block.
4. Days 18 to 21. Complete the profiles the sameAs array points to. Draft the Wikidata item if you meet the notability rules.
5. Days 22 to 27. Request indexing in Google Search Console. Confirm in the logs that OAI-SearchBot, PerplexityBot and Claude-SearchBot are fetching the newsroom.
6. Days 28 to 30. Rerun the probe. Compare the citation rate. Put the weekly probe in the team calendar.