AI Summarises Your Newsletter. The Ad Still Gets Billed.

When an AI assistant reads a newsletter and hands the subscriber a three-sentence summary, the sponsor's placement is almost never in those three sentences. The tracking pixel still fired, because the assistant fetched the content. The impression was still recorded. The advertiser was still billed.

Every system in the transaction reports success. The ad reached nobody.

This is no longer a forward-looking problem. Google began rolling AI-generated summaries across Gmail at scale in January 2026, and reported click-through rates across email marketing have moved in a direction consistent with what you would expect: opens holding or rising while clicks soften. If you sell or buy newsletter advertising, the first reader of your email is now frequently not a person.

This guide covers what actually breaks, which metrics survive, why native text formats gain a structural advantage they did not previously have, and what publishers and advertisers should change now.

What the assistant layer actually does to an ad

An email advertisement is content positioned inside a message, on the assumption that a human moves through that message in sequence and encounters the placement along the way.

Every part of that depends on the human reading the message rather than a compressed version of it.

When an assistant ingests the newsletter and produces a summary, three things happen simultaneously and they are not visible to each other.

The fetch registers as an open. The assistant loads remote content to read the message, which fires the tracking pixel. Depending on the platform, that may be logged identically to a human open. This is the same mechanical problem that Apple's Mail Privacy Protection introduced, arriving from a second direction — and it compounds rather than replaces it, as covered in our guide to Gmail and Apple MPP tracking.

The advertisement is compressed out. A summariser optimising for what the user asked for treats sponsored content as noise. That is correct behaviour from the reader's perspective and fatal from the advertiser's. Nobody chose to remove the ad. It disappeared as a property of compression.

The subscriber reads something. They received the information they wanted. They simply received a version of the newsletter with the advertising removed by a process no party in the commercial transaction agreed to.

What the data shows so far

Three things are worth establishing before the argument, because the scale changed faster than most publishers noticed.

The rollout is already broad. Google began pushing AI-generated summaries across Gmail during January 2026, reaching a user base measured in billions of inboxes rather than a limited test cohort. Gmail is the dominant client for most consumer newsletter lists, which means this is not a segment of your audience — for many publishers it is the majority.

Click-through rates have moved. Reporting on email marketing performance following the Gemini rollout has shown average click-through declining, with figures around 4.35% falling to roughly 3.93% cited across industry coverage. That is a relative decline near 10% on a metric that had been broadly stable, over a period with no other obvious cause.

Auto-opens are being recorded as opens. Generating a summary requires fetching message content, which loads the tracking pixel. Some platforms log that identically to a human open. So the same event that suppresses a click inflates an open.

I would treat the specific percentages as directional rather than precise, because attribution across an industry-wide shift is genuinely hard and the figures come from aggregated reporting rather than a controlled study. The direction is not in dispute and the mechanism is straightforward.

What matters for anyone selling or buying newsletter inventory is that this is a present-tense change with observable effects, not a forecast.

Why this is harder to handle than ad blocking

The instinctive comparison is browser ad blockers, and the comparison misleads in a specific way.

Ad blocking is adversarial and detectable. A publisher can measure it, respond to it, and in some cases negotiate with the reader about it. There are two recognised parties with opposed interests.

Summarisation is a helpful feature performing exactly the function the user enabled. Nobody in the chain intends to remove the advertising. There is no adversary to negotiate with because there is no adversarial intent, and the reader would be genuinely worse off if the feature stopped working.

It is also invisible in a way blocking never was. A blocked ad produces a measurable absence. A summarised-away ad produces a fired pixel, a clean delivery report, and an invoice that reconciles perfectly.

The metric scissor, and how to spot it in your own data

Here is the pattern to look for, because it is diagnosable before it becomes a commercial problem.

Machine fetches inflate opens. Summarisation suppresses clicks. Those two effects push your headline metrics in opposite directions at the same time, which produces a distinctive signature: a widening gap between open rate and click rate with no corresponding change in content, list composition, or send cadence.

Work a simple illustration. A newsletter sends 40,000 emails and historically records a 42% open rate and a 2.6% click-to-delivered rate. Assistant fetching adds machine opens; assistant summarisation removes some share of the human clicks.

A year later the same newsletter reports a 47% open rate and a 2.1% click-to-delivered rate. Opens up five points. Clicks down half a point, which is a 19% relative decline.

Read either number alone and you reach the wrong conclusion. The open rate says engagement improved. The click rate says something is wrong with the content. Read together, the divergence is the signal, and the cause is upstream of anything the publisher did.

The diagnostic to run: plot open rate and click-to-delivered on the same chart, monthly, for the last twenty-four months. Healthy newsletters show these moving broadly together. A scissor pattern — one rising while the other falls, sustained over quarters — is the fingerprint of machine reading, and it is worth separating from a genuine content decline before you act on it.

What a click decline actually does to your sponsorship revenue

The scissor is a measurement problem until it becomes a renewal problem, and the arithmetic on that transition is worth running.

Take a placement sold at a flat $2,000 against a list delivering 20,000 emails per send.

At a historical click-to-delivered rate of 2.6%, the sponsor receives 520 clicks. Their cost per click is $3.85.

After the decline described above, the same placement at 2.1% delivers 420 clicks. Cost per click: $4.76.

That is a 24% increase in what the sponsor pays per click, on an unchanged invoice, from a publisher who did nothing wrong.

If the sponsor's downstream conversion rate holds — and there is no reason it should not, since the people still clicking are still people — their cost per acquisition rises by the same 24%. For most advertisers that is enough to move a placement from comfortably inside their acceptable range to marginally outside it.

Here is the part that makes this expensive rather than merely unfortunate. The sponsor will not diagnose this. They will see a campaign that underperformed against a prior benchmark, conclude the newsletter's audience declined in quality, and not renew. The publisher will see a non-renewal with no stated reason.

Neither party has the information to identify the actual cause, because the mechanism sits between the send and the reader and is invisible to both. This is the same structural failure as the seasonal delivery gap covered in our piece on Q4 flat-rate pricing: a silent change in delivered value, absorbed by the advertiser, surfacing months later as a lost account.

The defence is disclosure. A publisher who proactively shows a sponsor the click-rate trend, explains the cause, and adjusts pricing or offers additional delivery keeps the relationship. A publisher who invoices normally and waits does not.

Which metrics survive

The answer is the same one that survived MPP, which is not a coincidence. Each successive layer of machine mediation degrades the same metric and leaves the same one standing.

Click-to-delivered survives. An assistant summarising a newsletter does not click the sponsor link. A click still requires a decision by a person. Both numerator and denominator remain clean.

Note the denominator specifically. Click-to-open rate is now corrupted from two directions — MPP inflated the denominator, and assistant fetching inflates it further. Click-to-delivered is the version that holds. The reasoning is covered in more depth in advertiser reporting for newsletter ads.

Downstream conversion survives, and gains importance. If an advertiser can match a click through to a signup or a purchase, that chain is entirely human. Attribution setup without cookies is covered in attribution 101.

Reply rate survives and is under-used. Replies require a person, and they carry a second benefit: they are among the strongest positive signals a mailbox provider weighs when deciding inbox placement, which matters for reasons covered in why deliverability is an ad revenue problem.

Open rate does not survive as a comparative metric. It remains marginally useful for tracking a stable list against itself. It is unreliable for comparing publishers and unusable for pricing, which is a problem given how many rate cards still rest on it.

Native text formats gain an advantage they did not have before

This is the most practically useful consequence and it points the same direction as several existing arguments, which is usually a sign it is real.

An image banner is, to a summariser, an object with no semantic content. There is nothing to read, nothing to weigh for relevance, and nothing to carry into a summary. Alt text is the only textual signal, and alt text is typically a brand name.

A native text placement, written in the publication's voice and integrated into the editorial flow, is at least legible as content. Whether the summariser retains it depends on whether it reads as relevant to what the user wanted.

Which introduces a genuinely new criterion for ad copy: write something a summariser would consider worth mentioning.

That sounds like a technical trick and it resolves into ordinary editorial quality. Copy that a human reader would find genuinely useful — a specific claim, a concrete number, a distinct piece of information — is copy a summariser is more likely to retain, because both are selecting for informational value.

Copy that is pure brand assertion carries no information, which means there is nothing for a summariser to preserve even if it wanted to.

This stacks on top of the reasons native already outperformed banner: it renders when images are disabled, it survives dark mode inversion, and it does not shift the image-to-text ratio that spam filters weigh. Those are covered in ad creative formats that convert and newsletter creative specs. MailAdx native ad units are built as text-first placements for this stack of reasons rather than for any one of them.

What publishers should do now

Five things, roughly in order of return.

Instrument the scissor. Track open rate and click-to-delivered together, monthly. You cannot manage a divergence you are not plotting, and this is the earliest available signal.

Move your rate card off open rate. If your pricing rests on a number that is inflated by two independent machine-reading effects, you are exposed the moment a sophisticated buyer runs their own reconciliation. Leading with click-to-delivered is defensive and it is also a differentiator, because most publishers still quote the inflated figure. Our guide to building a media kit covers what to present alongside it.

Shift ad load toward native text. For the rendering reasons, the deliverability reasons, and now the summarisation reason. Placement guidance is in publisher ad unit best practices.

Put the sponsored section where a summariser is likeliest to see it as content. Early placement, inside the editorial flow, with a clear label. A sponsored block appended after a sign-off reads as an attachment to both humans and machines.

Rebuild engagement-based automation on clicks. If your re-engagement or sunset workflows trigger on opens, machine fetching means those conditions will increasingly never become true. You will retain dormant subscribers indefinitely while correctly sunsetting engaged ones. This is an afternoon of work and it prevents years of silent list decay.

What advertisers should do now

Stop buying on open-based CPMs where you can. Impression pricing anchored to a metric inflated by machine reads transfers the entire risk to you. Pricing models and their trade-offs are covered in newsletter ad pricing models.

Brief for information, not for positioning. If the copy contains a specific claim, a number, or a concrete offer, it has a chance of surviving compression. If it contains a value proposition, it does not.

Reconcile against your own analytics. Compare the publisher's reported clicks to sessions arriving on your landing page. A gap that widens over successive flights with the same publisher is worth a conversation, and most advertisers never run this check.

Weight campaign evaluation toward outcomes. Conversions require a human. Impressions increasingly do not. Setting up conversion tracking is covered in our guide to the newsletter advertising ROI calculation.

Ask the publisher what their scissor looks like. A publisher who can produce a twenty-four-month chart of open rate against click-to-delivered is running a tighter operation than one who cannot, and the answer tells you how to weight their reported numbers.

The move I would avoid

There will be pressure to build countermeasures. Detecting assistant fetches and serving them differently. Structuring content to resist compression. Obfuscating sponsored sections so a summariser cannot cleanly identify and drop them.

I think that is a mistake, for a reason that has nothing to do with technical feasibility.

It puts publishers in an adversarial position against a feature their subscribers actively chose. That is the ad-blocking war restarted, against a counterparty with better tooling, more legitimacy, and the reader on their side. Channels that have fought their audience's tooling have generally lost, and lost slowly enough to waste a great deal of effort on the way down.

The durable position is the opposite: make the advertising something a summariser has a reason to retain, because it contains information a reader wants. That is the same standard a good editor would apply, which is a reassuring sign that it is the right one rather than a workaround that expires with the next model release.

What this does not change

Worth stating plainly, because the framing above could read as more alarming than the situation warrants.

Newsletter advertising's core advantage is unaffected. The audience opted in, the publisher has a first-party relationship, and no third-party cookie is involved anywhere in the chain. Those properties are why email advertising is structurally durable as privacy restrictions tighten, and an assistant layer does not touch any of them. That argument is set out in cookieless newsletter advertising.

Summarisation also concentrates value rather than destroying it. A reader who relies on a summary for routine mail and reads specific publications properly is telling you something useful about which publications those are. Newsletters people actually want to read are the ones least likely to be summarised away, which makes editorial quality a measurement advantage rather than only an audience one.

And the effect is uneven. B2B and professional newsletters — where the reader is subscribed for a specific purpose and reads deliberately — are considerably less exposed than general-interest consumer mail. The B2B newsletter advertising guide covers why that audience behaves differently.

Frequently asked questions

Do AI email summaries count as opens?

In many cases yes, because generating a summary requires fetching the message content, which loads the tracking pixel. Whether your platform distinguishes these from human opens varies by provider. The practical consequence is that open rate is now inflated by two independent machine-reading effects — Apple's Mail Privacy Protection and assistant fetching — and should not be used for comparison or pricing.

Will my sponsor's ad appear in an AI summary?

Usually not, and there is no mechanism to require it. A summariser optimising for the reader's intent treats sponsored content as lower priority than editorial content. Native text placements carrying specific, useful information have a better chance of retention than image banners, which contain no text for a summariser to evaluate at all.

Should I stop selling on CPM?

Not necessarily, but you should be clear about the denominator. CPM priced against open-based impressions is exposed to machine-read inflation. CPM priced against delivered sends is not, because deliveries are a server-side fact. Moving your basis from opens to sends removes the exposure without abandoning impression pricing entirely.

Is this worse for image ads or text ads?

Considerably worse for image ads. An image banner is semantically empty to a summariser — there is nothing to read and nothing to carry forward. A native text placement is at least legible as content and can be retained if it carries information the reader wants. This adds to the existing rendering, dark mode, and spam-filter arguments for text-first formats.

How do I tell machine reading from a genuine engagement decline?

Plot open rate and click-to-delivered together over twenty-four months. A genuine content problem moves both down together. Machine reading produces a scissor — opens flat or rising while clicks fall. If you see the scissor with no change to your content, cadence, or acquisition mix, the cause is upstream of anything you did.

Is this the same problem as Apple Mail Privacy Protection?

Related but distinct, and they compound. MPP pre-fetches remote content on delivery, which inflates opens and tells you nothing about whether a human read the message. Assistant summarisation fetches content to read it, which also inflates opens, and additionally suppresses the click because the reader receives a compressed version. MPP broke the open metric. Summarisation breaks the open metric again and reaches the click metric as well.

Should I move sponsored placements higher in the newsletter?

It helps on the margin and it is not a solution. A summariser processes the whole message rather than the first portion, so position matters less than it does for a scrolling human reader. What position does affect is whether the placement reads as part of the editorial flow or as an appended block, and content integrated into the body has a better chance of being treated as content. A sponsored section after your sign-off reads as an attachment to both humans and machines.

Does this affect deliverability?

Not directly. What it affects is your ability to detect a deliverability problem, because the open rate that would normally signal a placement issue is being propped up by machine fetches. A newsletter sliding into the promotions tab now shows a smaller open-rate decline than it should, which delays diagnosis. Watch click-to-delivered for the real signal.