Five months of AI Overviews data shows opens up and clicks down: recipients read summaries. How to structure email so the model's version still does your job.
In January we said we would wait for data before writing the practical guide to Gmail's AI Overviews. Five months in, the data has a shape. Industry benchmarks show open-adjacent metrics climbing toward the mid-40s in percentage terms while click-through rates slipped, from roughly 4.4% to under 4% in aggregate measurements since the rollout, the signature of recipients extracting what they need from summaries without visiting the message. The summary is now a primary reading surface. This article is about writing email that still works when a model reads it first.
What summarizers do to typical marketing email
A summarizer extracts what it judges central and discards the rest, and typical marketing email is built almost entirely of the rest: atmospheric openers, brand storytelling, three CTAs competing for primacy, and the actual offer buried in a hero image. Run that through a model and the recipient gets something like "a promotional message from Brand X about their products," which is technically accurate and commercially worthless. The failure is not the model's. A message with no extractable point summarizes as pointless, and the summarizer is just the first reader honest enough to say so.
Structure that survives compression
The summarization-proof message
- 1
One message, one point
Decide the single sentence you need the summary to contain, and build the email around it. Multi-topic digests summarize as topic lists; single-purpose messages summarize as themselves.
- 2
State the point in the first hundred words, in text
Subject, preheader, and opening paragraph should each carry the core: what, for whom, until when. Models weight early, explicit content; so do skimming humans, which is why this was always the advice.
- 3
Put the specifics in plain text
Dates, prices, discount amounts, deadlines, and names survive extraction verbatim when they exist as text. "30% off until Friday, June 26" travels through a summary intact; the same offer living only in a banner image does not exist to the model.
- 4
Make the CTA singular and self-describing
One primary action, labelled with what it does ("Book your renewal call", not "Learn more"), stated in text near the point it serves. Summaries frequently surface the action; give them one worth surfacing.
- 5
Keep HTML semantic and images decorative
Real headings, real paragraphs, real lists parse cleanly. Anything essential that lives only in images, or in layout tricks that scramble reading order, is invisible or garbled at the summary layer, exactly as it always was for screen readers.
What not to do
The temptation industry forums discovered within weeks: prompt-like instructions embedded in messages, hidden text addressed to the model, keyword stuffing aimed at summary inclusion. Treat all of it as radioactive. Hidden-content tricks are classic spam-filter signals with decades of detection behind them, providers are explicitly hardening summarizers against manipulation, and a summary that misrepresents your mail to a recipient is a complaint generator even when the trick works. The arms race framing from our AI content piece last year applies unchanged: you are not going to out-prompt the company that builds the model, on their infrastructure, against their incentives.
Measuring in the summary era
The metrics shift requires reading discipline. A falling click rate with stable placement and stable conversions means summaries are absorbing intermediary clicks, not that the program is sick: some recipients now act on the offer (visit the site directly, use the code) without the tracked click. That makes conversion and revenue per send the honest program metrics, with clicks demoted to diagnostic status alongside opens. Watch the Gmail segment separately, where the effect concentrates, and A/B test against downstream outcomes rather than click-through, since the click is now a sometimes-skipped step rather than the funnel's front door.
The strategic reading of five months of data: AI mediation punishes padding and rewards clarity, which is a redistribution most senders should welcome. The messages losing engagement are the ones that needed a scroll to find their point; the ones holding steady already led with it. Write the email whose three-line summary is the email, and the model becomes a distribution channel instead of a filter.
Frequently Asked Questions
Should we write shorter emails now?
Do plain-text emails summarize better than HTML?
Is there any legitimate way to influence what the summary says?
Does this apply outside Gmail?
Key Takeaways
- Five months of data shows summaries absorbing readership: opens up, click-through structurally down
- Messages survive summarization when one point, with concrete specifics, lives early and in plain text
- Single self-describing CTAs and semantic HTML give the model something worth surfacing
- Prompt injection and hidden-text tricks are spam signals in a losing arms race; structure is the legitimate lever
- Judge programs on conversions and revenue per send; the click is now an optional step, not the funnel door
Related articles
Half-Year Review: Email in the AI Inbox Era
Six months that rearranged the reading layer: Gemini in Gmail, Microsoft rejecting outright, DMARC finally a standard. What the first half of 2026 means for senders.
Deliverability Postmortems: Learning From Incidents
The incident is over, placement recovered, and the pressure to move on is enormous. The blameless postmortem practice that converts each incident into prevention.
Per-Provider Playbooks: Operating Gmail, Microsoft, and Yahoo Differently
The big three score senders differently, expose different data, and fail in different ways. One operating playbook per provider beats one averaged strategy.