AI email summaries: What 628 emails reveal about the layer between you and the journalist

Your pitch might never be read. Not because the journalist ignored it - which does happen, especially if your subject line doesn't show relevance to the topics they cover or areas of interest at that moment in time. But because an AI summarised it first, got it wrong, and the journalist moved on without ever seeing what you actually wrote.

AI is already embedded in how journalists on the receiving end of your pitches work. And in our analysis of 628 emails, up to a third of AI summaries misrepresented the original - the data, the information, or the expert commentary inside it.

We collaborated with digital PR tool BuzzStream to look at how AI inside the three largest email providers - Google, Apple and Microsoft - is summarising pitch emails: which elements of the original email carry the most weight, and what that means for how PRs should structure pitches from here.

Why AI email summaries matter now

According to Gallop's Workplace Panel in Q4 2025, 26% of US employees use AI a few times a week and 12% use it daily; in the UK, the 2026 GOV.UK Survey on AI Skills highlighted that 21% say AI has increased their productivity at work.

Muck Rack's 'The State of Journalism 2026' report uncovered that among journalists specifically, AI adoption has climbed to 82%, with ChatGPT and Gemini leading the way. So AI is already embedded in the daily workflow of the people you're pitching to.

The same report found that 29% of journalists now receive 6-10 PR pitches on a normal work day, up from 25% the year before - and 14% of journalists record getting more than 20 pitches a day! At the same time, understaffing and time constraints worry 20% of the profession, and disinformation and misinformation was cited as one of the biggest issues of concern to 32% of journalists.

So journalists are inbox-heavy, concerned about misinformation, and already finding ways to increase productivity and efficiency - meaning the question is not whether AI is summarising your pitches, but how, and if those summaries are accurate or being classed as misinformation.

What 628 emails (and their AI summaries) reveal

We analysed emails for 13 different campaigns. Each campaign had 4-11 email templates, with each template isolating individual variables like subject lines, formatting, data placement and links. Every email was sent to two addresses per email platform (Google, Apple and Microsoft), and summaries were recorded and analysed to determine which variables had influence over the summary output.

Three findings stood out: platforms produce wildly different summaries, roughly one in three summaries misrepresents the original, and bullet points were most likely to influence what AI surfaces. Each has practical consequences for how a pitch should be written moving forward.

AI email summaries are dramatically different across platforms

The platforms don't just summarise differently, they produce fundamentally different outputs. So which email provider the journalist uses could impact how they receive your email.

Google (Gmail's Gemini) produces the shortest AI email summaries at just 29 words on average - meaning your entire pitch is being compressed into roughly two sentences. There's almost no margin for error when the summary is that compressed.

Apple sits in the middle at around 50 words. Whereas Microsoft (Copilot) generates the longest AI email summaries at an average of 157 words, giving it room to capture more detail but also more room to deviate from the original.

Microsoft Copilot was the only tool we saw pulling in information that wasn't in the email itself. Its built-in web-search capability grounds responses in Bing results - which can include the campaign you're pitching - unless an admin switches it off at organisation or group level. In practice, that means Copilot will sometimes summarise your pitch using context you never sent.

PRs and marketers need to think about the possibility of having emails both being summarised in brief, or with a longer summary.

One in three emails is misrepresented

AI summaries misrepresented information in up to one in three emails.

Apple had the highest error rate at 33%. Microsoft followed at 30%. Google was the most accurate at 11% of emails being misrepresented by the summaries - but even that means one in nine summaries contains something that wasn't quite what the email said.

The types of errors matter too. We saw:

• Overgeneralisation - some summaries make the email sound like it's just focused on one area, when it covers a broader range of data and hooks.
• Data misinterpretation - one example being confusion between similar data sets, where the AI conflated two distinct statistics into one misleading statement.
• Misrepresentation of expert commentary - one example being that an expert recommended "keeping your vehicle as charged as possible - around 40-80% is ideal", but the email summary surfaced "charging to 40-80%", which reads as only charging the battery by that amount. Very different and not something we'd want journalists to think our 'expert' would recommend.
• Misinterpretation of context - for example, the AI read the heading row of a city-ranking table and reported the top-ranked city as 'best for everything', when the table actually ranked separate categories (cost, commute, schools) independently.

These aren't dramatic fabrications, but they're subtle distortions, the kind that could make a journalist question your expert or if you're giving misinformation - and you'd never know it happened.

Bullet points are doing more work than ever

Of every formatting trick we tested bullet points were the most likely to influence the AI email summary output, being included in at least 64% of summaries across every platform:

• Microsoft Copilot - the heaviest influence at 92% inclusion.
• Apple Intelligence - pulled bulleted content 64% of the time.
• Google Gemini - 65% inclusion rate. Notably, it's the only platform that appears to have a preference over bullet point position, most often pulling from the first bullet.

Bullet points have always been useful for pitch emails - they break up a wall of text and let a busy journalist scan the key facts of your story quickly. But if journalists are starting to lean on AI email summaries to speed up reviewing their inbox, those bullets are doing more than just helping a human skim, they're actively shaping what AI flags as the important points of your pitch.

If AI is doing the first read of your email for a journalist, your bullet points are doing the first selling for you.

Platform by platform: how Google, Apple and Microsoft handle your pitch

The table below shows the average summary length, and the percentage of AI summaries, per platform, that featured specific variables from the pitch emails.

Four secondary findings round out the picture:

• AI email summarisation pulls disproportionately from the first half of the message. This isn't surprising if you think about how these models are designed to work - but it has real implications for anyone crafting pitches.
• Subject lines play a role too. Keywords like "new," "data," and the expert's position are pulled into summaries more frequently when used in subject lines vs only in the body of the email, with Microsoft Copilot email summary generation showing the strongest subject-line influence.
• Bold text has a weaker effect than you might expect, influencing only 25-48% of summaries depending on the platform.
• Tables have minimal impact - perhaps because they're usually further down the email.
• Graphics or attachments are essentially invisible to all three systems.

What this means for how PRs should write pitches

This isn't about gaming an algorithm, there's no set formula or way around AI email summaries. It's about understanding a new layer between your email and the person you're trying to reach, and writing with that layer in mind to increase the chances of the right things being surfaced and emails not being misinterpreted.

1. Front-load your most important information

It's always been crucial for PRs to get to the point right at the top of an email, so the journalist can easily see who, what and why - including why it's relevant right now. It's now even more essential to have anything you want surfaced in summaries at the top.

If your strongest data point, your most relevant hook, or your expert credentials sit in the second half of the email, there's a chance the AI email summary will never surface them.

2. Use bullet points deliberately

Bullet points have always been useful for pitch emails - they let a busy journalist scan the key facts quickly - now, they're actively shaping what AI flags as the important points of your pitch.

The findings make two things crucial:

• Lead with your strongest data point or information in the first bullet point.
• Keep bullet points as simple as possible. Make sure each bullet is phrased really clearly with context, and can be understood on its own, so it can't be confused or misinterpreted by AI.

This might sound like obvious advice, but complex bullet points can lead to the misinterpretation we're trying to avoid.

3. Keep data clear and contextualised

With AI misrepresenting data in roughly one in three emails, the way you present statistics matters more than it used to. Some key things to keep in mind:

• Include the context immediately alongside the figure, so every data point can be understood on its own - like a standalone point.
• Avoid layering multiple comparisons, or if you need them ensure they have clarified in the most simple way possible.

4. Craft your subject lines carefully and include key words

Whilst subject lines influenced under half of AI email summaries, key words like "new" and "data" appeared significantly more often in email summaries when it was included in the subject line than when it was just in the body of the email.

Including key words like these, that simply and clearly demonstrate what you're offering can help ensure the journalist receives that in their summary. Plus, remember the subject line is what gets the email opened in the first place.

If you're pitching expert commentary, we'd recommend using their position in the subject line too, to help sell their expertise. This also appeared more often in email summaries when used in subject lines than when solely used in the body of the email.

5. Don't rely on tables, graphics, or links to carry crucial information

Information and data from tables is included in fewer than half of AI email summaries, which could be because tables often sit slightly lower down emails too. They're not completely redundant though, and it appears the top 1-3 entries in a table are most favoured. But don't hide crucial information you want to highlight to the journalist in the first instance in tables, or it might get left out.

In addition, information from graphics and attachments don't appear to be included in summaries at all, and only Microsoft Copilot mentions links - even then, full URLs are cited just 3% of the time. This further emphasises the importance of including key selling points at the top of the email, yet still including these elements for those journalists who want to explore the story further.

In summary, PRs are now writing for three readers, not one

PR professionals have always written for two audiences - the journalist and their readers. Now there's a third, the AI that sits between the send button and the journalist's attention. And unlike a human reader, it doesn't ask for clarification when something is ambiguous. It just summarises - sometimes accurately, sometimes not - and moves on.

The good news is that this is a solvable problem. Our research gives us clear signals about what these systems prioritise and where they fail.

The shift isn't dramatic, it's structural. Front-load your strongest material, format for clarity rather than visual appeal, and accept that roughly a third of the time an AI is going to get something wrong - so write in a way that minimises the damage when it does. 

Ultimately, these are the same techniques that make for good PR pitching anyway - being clear, being relevant, putting the strongest material first. The AI layer just raises the stakes on getting it right.

If you want to dig deeper into how Aira's digital PR team approaches outreach strategy, or if this research has raised questions about your own pitch structure, we're always happy to talk through it.

Methodology

We tested 13 digital PR campaigns, creating 4-11 outreach email templates per campaign. Each template varied one element at a time to isolate its effect on email summaries. Variables tested included subject lines, bold key points, bullet points, tables, tailored introductions, graphics, press releases, and links to full campaign content, expert LinkedIn profiles, or About Us pages.

Every template was sent to two email addresses across each of the three major providers (Google, Apple, and Microsoft), allowing us to also assess whether the same provider generated different summaries for identical emails. All summaries were recorded on receipt.

Summaries were analysed against a set of research questions to identify the key factors shaping them and to assess whether they accurately represented the original emails. Claude AI supported part of the analysis, but methodology and raw outputs were manually reviewed to validate conclusions. Manual review also covered most research questions directly - including the impact of bold points, bullets, and tables, and any cases of misrepresentation.

Data was collected and analysed across March and April 2026. The complete collaborative study, with all data, can be found at https://www.buzzstream.com/blog/ai-generated-email-summary/.