How to Optimise a Press Release for AI Search Visibility
Last Updated on 9 September 2026 by Francois Marsteau
AI assistants now answer questions that used to start with a search. When someone asks ChatGPT or Google’s AI Mode about a company, a product or a market, the answer is assembled from sources the model can parse, date and attribute. A press release can be one of those sources — or it can be invisible.
The short version: AI engines cite releases that state verifiable facts, carry a clear date and a named author, and are structured so a machine can extract a single claim without reading the whole page. Keyword density has nothing to do with it.
What we measured on our own site
We publish press releases and editorial guides, and we track what AI features do with them. Over three months, Google’s AI features reported 157,963 impressions across our pages. One editorial page accounted for 114,532 of them.
Three findings are worth knowing before you invest in this.
AI citations do not behave like search results. On our most-cited query, we recorded 88,219 impressions at an average position of 1.5 — and zero clicks. Not a low click-through rate. Zero.
Being cited is not the same as being visited. Any provider promising traffic from AI visibility is describing something they have not measured.
Citations land in unexpected markets. Our AI impressions came mainly from Brazil, Argentina, France and Malaysia. English-speaking countries accounted for 8.8 % of them. AI answers are generated wherever the question is asked, and the language of your release does not decide where it is read.
Editorial content is cited far more than releases. Across the same period, 167 of our English press releases appeared in AI answers, averaging 28 impressions each over three months. A single editorial guide drew 114,532. If your goal is to be cited, a release announces; a guide explains — and models quote explanations.
What makes a release citable
Six things, in the order they matter.
- A single, checkable claim in the first two sentences. Models extract short spans. If your news is “we opened a factory in Lyon employing 200 people from September 2026”, write that, in those words, at the top. If the reader has to reach paragraph four, so does the machine.
- A visible date, and one date only. Publication date, event date, effective date — label each. An undated document is treated as an undated document, and undated documents lose to dated ones.
- A named author or spokesperson, with a role. “A company spokesperson” is not attribution. “Marie Dupont, Chief Operating Officer” is. Attribution is what lets a model present your statement as someone’s statement rather than an anonymous claim.
- Numbers with units and sources. 20,000 m². 200 employees. €14 million. Percentages with their base. A model that can quote a figure will quote it; one that finds only adjectives has nothing to carry.
- Structure a machine can walk. Headings that describe their section, short paragraphs, one idea each. Tables for anything comparative. Bullet lists only where the content is genuinely a list.
- A boilerplate that identifies you unambiguously. Legal name, sector, location, website. This is how a model links your release to your company rather than to a similarly named one.
What does not work
Keyword stuffing — and it now carries a risk
Google’s spam policies explicitly name links with optimised anchor text in widely distributed press releases. Writing for keyword density does not improve AI visibility and can damage your search visibility.
Publishing the same release in twenty places
Duplication does not multiply citations. Models converge on one source, usually the one with the clearest attribution.
Adjectives
“Innovative”, “leading”, “revolutionary” are unquotable. They carry no information a model can present as fact, and they are the first thing a journalist deletes.
Locking the facts inside a PDF or an image
If your key figures are in a graphic, they do not exist for a language model, and often not for a search engine either.
Where technical work actually helps
Most of what makes a page citable is editorial. A few technical elements do matter, and they are the publisher’s job rather than yours.
| Element | What it does |
|---|---|
| Structured data | NewsArticle with datePublished, author and publisher lets a machine read the metadata without inferring it. Every release published through 24Presse carries it. |
llms.txt | Tells AI crawlers what a site contains and where the canonical version of each document lives. |
| Crawler access | GPTBot, ClaudeBot, PerplexityBot and Google-Extended must be allowed to fetch the page. Many sites block them without realising. |
| Speed | A page that takes several seconds to render is fetched less often. |
If you distribute through a provider, these are worth asking about. If you publish on your own newsroom, they are worth checking.
How to know whether it works
Two measurements, both free.
Google Search Console → Performance → Search Appearance → AI features. This reports impressions from AI Overviews and AI Mode separately. Watch the impressions, and watch the clicks separately — as our own data shows, the two move independently.
Your analytics referral report. Traffic arriving from chatgpt.com, perplexity.ai or an “AI Assistant” channel is a direct signal. Over thirty days, that channel brought us 115 sessions — modest in volume, ahead of DuckDuckGo and Facebook combined.
Neither number will be large at first. Both are real, and both are ahead of where most of this industry is looking.
Frequently asked questions
Not directly, in our measurements. We recorded 88,219 AI impressions at position 1.5 with zero clicks. AI visibility builds presence in answers; it does not behave like a search ranking.
No. What a model needs — a clear claim, a date, a named source, concrete figures — is what a journalist needs. Writing for AI visibility mostly means writing well.
Not reliably. Duplication tends to consolidate on a single source rather than multiply citations.
It helps, and you usually do not control it — your distribution provider does. Ask whether releases are published with NewsArticle markup including a publication date and a named author.
The ones your buyers use. In our own referral data, ChatGPT dominates. Google’s AI Mode generates the largest impression volume but the fewest visits.
There is no fixed delay. Citation depends on the model’s index and on the question being asked, not on a publication schedule.
Publish something worth quoting.
If your announcement contains a fact worth repeating, a date, and someone willing to be named, it is already most of the way there.
Send your press release — published with NewsArticle structured data, open to AI crawlers, and distributed to journalists who cover your subject. Or see how long distribution takes before you plan your date.