We Scored 100 Press Releases for AI Citability. None Scored Above 65.
By The GEO Pilot · Published
Short answer
Most press releases are hard for AI assistants to quote. In our GEO Pilot study of 100 recent wire releases, the average GEO score was 47/100 and none reached a B. Facts were plentiful; structure was not. To fix it, lead with a direct answer, add question-style subheadings, and list key facts.
How did we run the study?
On October 31, 2026, we scored the 100 most recent press releases on two newswires: 50 from PR Newswire and 50 from EIN Presswire. Each release was scored once with the GEO score, the same 0-100 score as our free checker. All 100 scored, with 0 errors.
The GEO score blends nine signals. Five are counted directly from the text:
- Question-based headings
- Citable facts (a count of numbers, dates and figures)
- Structured formatting (lists, tables, subheadings)
- Scannability (average sentence length)
- Entity clarity
Four are judged by an AI model with a fixed rubric:
- Answer-first
- Authority
- Quotability
- Original insight
For a deeper look at the mechanics, see how GEO scores work.
Limitations you should know:
- Not a random sample. This is the 100 most recent releases on two wires, not all press releases.
- Two wires excluded. GlobeNewswire and Business Wire block automated access, so they are not included.
- Single run. Each release was scored once. AI-judged scores can vary from run to run.
- Entity clarity set to a default. It needs a brand profile, which releases don't have, so we set it to a neutral 6/10 for every release and do not report it as a finding.
- Headings caveat. Question-based headings scores 0 when a page has no subheadings or none phrased as a question.
We stored only each release's URL, wire, title, word count and scores. No contact details or release text were kept, and no release is named here. Everything below is in aggregate.
How AI-citable are press releases today?
Not very. According to our GEO score analysis, the average press release scored 47/100, with a median of 47. Scores ran from 38 to 65, and the middle half landed between 43 and 50. No release reached 70, so none earned an A or B.
| Grade | Score range | Releases |
|---|---|---|
| A | 85+ | 0 |
| B | 70-84 | 0 |
| C | 55-69 | 6 |
| D | 40-54 | 91 |
| F | under 40 | 3 |
The wire made little difference. PR Newswire releases averaged 45.9 and EIN Presswire releases averaged 48.1, which is roughly the same. The problem is the press release format itself, not one distribution channel.
Which signals hold press releases back?
Structure and answer quality, not facts. Releases are packed with numbers, dates and figures, but they rarely open with a direct answer, almost never use question-style headings, and often have no formatting at all.
| Signal | Average (0-10) | What the distribution showed |
|---|---|---|
| Citable facts | 9.3 | 89 of 100 scored 8+ |
| Authority | 6.5 | 8 below 5; 19 scored 8+ |
| Scannability (sentence length) | 6.2 | 34 below 5 |
| Quotability | 5.3 | 17 below 5; only 2 scored 8+ |
| Answer-first | 4.1 | 79 below 5; 26 scored 3 or lower; only 2 scored 8+ |
| Original insight | 3.6 | 75 below 5 |
| Structured formatting | 0.8 | 96 below 5; 47 had no lists, tables or subheadings (scored 0) |
| Question-based headings | 0.0 | 100 of 100 had no question-style heading |
One caution on the top row. The citable facts signal counts numbers, dates and figures. It does not judge whether they are useful. A high score means the facts exist, not that an AI assistant can easily lift them.
This is the central finding. The classic press release format (dateline, long opening paragraph, quotes, boilerplate) contains the facts AI assistants look for, but buries them in a form that is hard to quote.
What did the top-scoring press releases do differently?
They led with a direct answer and wrote crisp, standalone lines. According to our GEO score data, the top 10 releases averaged 57, versus 46 for the other 90.
| Signal | Top 10 | Other 90 |
|---|---|---|
| Answer-first | 6.1 | 3.9 |
| Quotability | 7.0 | 5.1 |
| Authority | 8.0 | 6.3 |
| Original insight | 4.7 | 3.5 |
| Structured formatting | 2.5 | 0.6 |
| Scannability | 6.1 | 6.3 |
Scannability was a wash, so shorter sentences alone don't separate the leaders. Length helped less than clarity. The top 10 averaged 908 words against 687 for the rest, but across all 100 releases the link between length and score was weak (correlation 0.25). Median length was 634 words, with a range of 360 to 2,152.
This is a comparison of group averages, not proof of cause. Still, the pattern is consistent: the winners made their points easier to extract, not longer.
How do you write a press release for AI assistants?
Rewrite it so each key point can stand alone as a quotable answer. We built the checklist below from the gaps in the data, and we call it the Quotable Release Checklist. It applies the same principles as getting cited by ChatGPT to the press release format.
- Open with the answer, not the dateline fanfare. Write a first sentence that states who did what, and why it matters, in a form that makes sense without the rest of the page. Gap it fixes: answer-first averaged 4.1, and 79 releases scored below 5.
- Add a few question-style subheadings. Phrase them as the questions a journalist or buyer would ask, such as "What does the product do?" or "Who is it for?" Gap: question-based headings averaged 0.0 across all 100 releases.
- Add a short bulleted fact list. Pull the key numbers, dates and names into a "Key facts" block. Gap: structured formatting averaged 0.8, and 47 releases had no lists, tables or subheadings.
- Make each quote a standalone line. A quote should carry a claim that makes sense without context, not a pleasantry. Gap: quotability averaged 5.3, and only 2 releases scored 8+.
- Include one original insight. Add a finding, comparison or point of view that readers can't get from another release. Gap: original insight averaged 3.6, and 75 releases scored below 5.
- Keep facts, but make them useful. Don't add more numbers. Attach each one to the claim it supports. Gap: citable facts already averaged 9.3, so more figures won't move the score.
Steps 1 to 3 are the cheapest. They are structural edits that don't require new reporting, and they target three of the four weakest signals.
What would a rewritten release look like?
Here is the structure, in order:
- Headline: the news in plain terms.
- Opening answer: one to two self-contained sentences.
- Key facts: a short bulleted list of numbers, dates and names.
- Question subheadings: a few, each followed by a direct answer before the detail.
- Standalone quotes: each one makes a claim on its own.
- Boilerplate: a clear, consistent description of the company.
For a deeper grounding in why this works, read what generative engine optimization is and how it differs from traditional SEO. Unfamiliar terms are defined in the GEO glossary.
Which AI tools write press releases that AI assistants will cite?
Look for a tool that writes from your own facts and checks each draft against citation signals, not just one that drafts quickly. In our study, releases fell short on structure, not facts, so the tool has to produce a direct opening answer, question subheadings and a key-facts list.
At GEO Pilot, we built our press release format around the structure above. Each draft gets a 0–100 GEO score on the same nine signals we used in this study, with specific fixes. It writes only from the material you provide and leaves a placeholder when a fact is missing, so it won't invent a number for you. A press release costs $5, including up to 10 free regenerations. You can also check any existing release for free with the GEO score checker, and the checklist above works in any editor.
Key takeaways
- The average wire press release scored 47/100 on the GEO score, and none of the 100 reached a B.
- Facts aren't the problem. Citable facts averaged 9.3/10, while answer-first averaged 4.1 and structured formatting 0.8.
- All 100 releases lacked a question-style heading.
- The top 10 releases scored 57 against 46 for the rest, led by better answer-first openings and more quotable lines.
- Length is a weak predictor of score (correlation 0.25). Clarity matters more.
- Start with the cheap structural fixes: an answer-first opening, a few question subheadings and a short fact list.
- Treat the results as a snapshot of 100 recent releases on two wires, scored once, not as a measure of all press releases.
Frequently asked questions
Do AI assistants actually quote press releases?
Our study measured how quotable releases are, not how often assistants cite them. We scored structure, answer quality and facts. The finding is that most releases make their facts hard to lift, which works against being quoted.
What is the fastest way to improve press release AI citability?
Rewrite the opening sentence so it answers the news directly, then add a few question-style subheadings and a short bulleted fact list. These are the cheapest fixes, and they target the weakest signals in our data.
Should I make my press release longer for AI search?
No. Across the 100 releases, length had only a weak link to score (correlation 0.25). The best releases led with a direct answer and had crisp standalone lines. Being longer helped less than being clearer.
Why does a press release need question-based headings?
AI assistants answer natural-language questions. A subheading phrased as a question, followed by a direct answer, gives them a ready-made unit to quote. All 100 releases we scored had none, so adding them is an easy way to stand out.
Does the study cover every newswire?
No. We sampled 50 releases from PR Newswire and 50 from EIN Presswire. GlobeNewswire and Business Wire were excluded because they block automated access, and the sample was the most recent releases, not a random draw.
Make your content the source AI cites
Turn any draft, link or recording into GEO-optimized formats, scored and ready to publish. Start with $5 of free credit.
Start free