Writing for the agent: when your customer’s AI reads your email first

In this article we look at what is actually changing in the inbox, what the parallel to search means in practice for marketing teams in the Netherlands, and how to write an email that holds up under both human and machine reading. We close with five principles and a note on where rich, themed campaigns still belong.

The agent layer is here

The change has happened quickly. Apple Intelligence email summaries have been available since iOS 18.1 and now run automatically on supported devices. Gmail’s Gemini-powered summary cards appear above messages the algorithm decides are long or complex enough to compress, and Google is adding AI to more parts of Gmail, from prioritization to reply drafting. Outlook Copilot is mainstream in enterprise. Third-party clients like Superhuman AI, Shortwave, Notion Mail and Arc Mail go further: the agent ranks, archives, replies and escalates without the user reading the original at all.

The dentsu 2026 CMO Navigator captures the broader picture. Nine in ten CMOs report that emerging AI capabilities are already reshaping their strategy, and the report calls this the “Algorithmic Era” (source). Email is one of the next places where AI will become visible to consumers, and where the impact for marketers is most immediate.

What changes when the first reader is a model

When a model reads first, three things change. Each has direct implications for how marketing email is briefed and written.

Subject line and preheader become decisive (again). In the early days of email marketing, subject line and preheader were treated as the most important 100 characters in a campaign. Dynamic preview rendering and image-led emails reduced their relative weight. Today, AI summarizers use the subject and preheader as primary signals for what to highlight, and how to summarize. Witty subject lines can still drive opens with human readers, but if they obscure the key message they will hurt summarization accuracy.

Structure beats flowing text. Hierarchical, scannable content gives AI summarizers more reliable signals. According to Attentive’s analysis of Gmail’s Gemini behavior, the first 100 to 200 characters of body copy drive the summary (source). The implication: place the value at the top and structure the rest. Long-form storytelling will be compressed to its core claim. If the core claim is buried, the model identifies a different one.

The proposition has to be extractable. If your offer cannot be stated in one sentence, the model will produce its own version. That version may not match your intended takeaway. Preheaders that contradict subject lines, headlines that compete with each other, and relevant info sitting in the bottom half all fail this test. Write the one-line summary you want first, then check that the email actually leads with it.

Lessons from search, lessons from commerce

Search has already been through this transition. SEO became GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) once LLMs entered search. dentsu has commercialized GEO as a service and has published on the Algorithmic Era of search, with new KPIs like True AI Share of Voice and a GenAI Visibility Index (source).

Email is going through the same transition. AI summarizes rather than ranks, but the marketer’s job is the same as in search: write content that is readable for the human recipient and for the AI that prepares what the recipient sees. Some practitioners call this Email Engine Optimization or agent-readable email. The label matters less than the underlying change.

Commerce shows what this looks like when it goes further. When Merkle tested AI shopping agents, the agents recommended products without visiting any brand websites. The agent chose where to source from, and brand sites were only one option; the recommendation came from what the LLM already knew (source). The brand had little role in the decision. Email works differently: the AI is given one specific email and asked to summarize it, so the brand’s words are the only input. But the principle is the same. Content not written for both readers will be summarized in ways the brand cannot predict.

What good looks like: five principles

These five principles apply to almost any campaign and are part of our standard processes.

1. Lead with the substance, not the warm-up. The first 100 to 200 characters of body copy now decide how the model summarizes the email. “Hi, we hope you are well. We have something exciting to share…” trains the AI to produce a summary about pleasantries. Open with the value, the offer, the deadline, or the news. Approach the opening line as the input to a summary, not as the introduction to a letter.
2. Structure first, flowing stories second. Use clear sections, descriptive headings, short paragraphs and explicit markers (“Offer ends Friday at 23:59”, “What’s included:” followed by a bulleted list). AI systems parse structured content well and extract meaning from long paragraphs less reliably. The closer your email looks to a well-marked-up document, the more reliably the model surfaces what you intended.
3. Make the proposition extractable. If a model could only return one sentence about your campaign, what would you want it to say? If you cannot answer that question quickly, neither your readers nor their AI assistants will. Test extractability before sending: write the desired one-line summary first, then check that the email actually leads with it. Or use your LLM to do so.
4. Subject and preheader are content, not packaging. Brief them as carefully as the body. Avoid treating the subject as a teaser and the preheader as the punchline. AI summaries are more accurate when subject, preheader and opening line align. What feels repetitive to a human reader can help the model identify the correct takeaway.
5. Test the summary before you send. Before approving any campaign, run it through Apple Intelligence, Gmail’s summary card and Outlook Copilot. If the summary the model generates is not the takeaway you intended, the email is not ready to send, regardless of the visuals. It is a quick step that few email marketeers have built into their workflow. We have built it into ours: agent-readability checks are part of the standard QA we now apply across client campaigns.

We apply this thinking to our own work for clients, where AI-readability is no longer a side experiment but a core part of how we build campaigns. These principles describe what Merkle calls content optimized for human engagement and machine comprehension. The wording comes from a different channel (commerce), but the underlying logic applies to email as well.

Where rich, themed campaigns still belong

These principles do not call for every email to become a plain-text functional message. Brand storytelling, re-engagement, visual product discovery and loyalty narratives still benefit from rich design, dynamic content and creativity. These formats still work, provided the principles above are applied to them.

The constraint is that visual richness needs to be balanced with a textual and structural layer the AI can read. A visually heavy email that lacks a clear textual lead-in, descriptive alt text and a single extractable message will summarize badly. A heavily visual email with those elements in place can still summarize well, and still serve the human reader who scrolls.

Some email types are more exposed to AI summarization than others. Long, narrative-led brand emails are most affected, because the elements that AI removes are often the ones that carry the message. Transactional confirmations are largely unaffected. Promotional emails sit between the two: they perform worse when the offer is buried under creative, and better when the offer is clearly visible alongside the visuals. For each campaign type, decide whether the AI summary is decisive or supporting, and brief accordingly.

Two short cases

1: What the public market shows

Two patterns are visible in the public newsletter market.

The first is the daily-brief format used by senders like Morning Brew: a tight summary at the top, deep links below. This format works well for AI summarization, since the summary card and the email’s own intro tend to say the same thing.

The second is visible in the major Dutch news brands. NRC, de Volkskrant and FD have all moved towards plain, factual subject lines and preheaders that summarize the story directly rather than tease it. Compared to two years ago, the visual density of these emails is lower, and the textual hierarchy is sharper. These programs are adapting in advance of broader AI summarization in mainstream consumer mail clients.

2: A small experiment any team can run this week

Take ten of your most recent campaigns and do three things with each:

1. Forward each to a Gmail address running Gemini summaries and screenshot the summary card.
2. Open each on an iOS device with Apple Intelligence enabled and screenshot the summary.
3. Paste the HTML into Outlook Copilot and ask for a one-sentence summary.

Then compare the summaries to your campaign brief. Count how many of the ten match the intended takeaway. In a first test, most teams find that fewer than half of their emails produce the summary they intended. Seeing your own emails summarized by three different models makes the principles above concrete. The gap between intended and actual summary tells you whether a campaign is optimized for the agent layer in email.

What this means for our processes

For marketing teams, the implications are structural rather than only stylistic. Email has joined search and product content as a channel that must serve two readers at once: the human recipient and the agent that determines what the recipient sees. Treating this purely as a copywriting question would underestimate the change. This way of writing needs to be built into briefing templates, QA checklists and approval steps, not left to individual copywriters.

Key takeaways

• The first reader of a marketing email is increasingly a model, not a human.
• Subject line, preheader and the first 100 to 200 characters of body copy drive the summary.
• Structure, an extractable core message, and alignment between subject, preheader and opening line support both human and machine reading.
• Rich, themed campaigns still work, provided they include a clear textual layer the AI can parse.
• Add a step to test the AI summary before sending.

When the first reader is a model, write for both.

This is an article by EMAS 2026 Gold Sponsor Merkle.

Curious how to put this into practice? Visit EMAS 2026 on June 11 at Circa Amsterdam. Be inspired by international speakers and discover how to take your email marketing to the next level. Tickets and programme: shop.emas.nu

Joost van Dam

Senior Strategist | Merkle Nederland.

• LinkedIn