Most people picture inbox filtering as a pretty simple process: an email arrives, something checks it against a list of known bad things, and it either passes or gets sent to spam. That mental model made sense maybe fifteen years ago. It does not really describe what is happening today.
What inbox providers are actually doing is making probabilistic judgements about every piece of mail that hits their systems. They are not just asking whether an email looks like spam. They are trying to answer a much harder question, which is whether this particular sender has a track record of sending things this particular recipient actually wants. And the signals they are using to answer that question have become significantly more sophisticated than most senders appreciate.
Here is what we know, what we suspect, and where the whole thing is heading (our opinion).
Authentication is the entry ticket, not the finish line
Before getting into the more nuanced stuff, it is worth being clear about where authentication fits in. SPF, DKIM and DMARC are not a competitive advantage. They are the minimum requirement to be taken seriously. If your authentication is misconfigured or missing entirely, inbox providers will treat your mail with deep suspicion regardless of how good everything else looks, and there is very little you can do about placement until the foundations are right.
We covered authentication properly in Part 2 of this series, so if that is an area you are not confident about, that is the place to start. For the purposes of this post, we are assuming authentication is in place and correctly aligned. The question then becomes what happens after that.
How reputation works in practice
Once an inbox provider has confirmed that a message is genuinely from the sender it claims to be from, it looks at what it knows about that sender's history. This is what people mean when they talk about sender reputation, and it operates at two levels.
Your domain builds a reputation over time based on how recipients have responded to your mail. Gmail, Outlook, Yahoo and the others track this across every mailbox they manage, which in Gmail's case means billions of individual data points. A domain with a long history of sending mail that people engage with, open, click on and keep, is treated quite differently from a domain that generates complaints, high deletion rates, or low engagement. That history is not forgotten quickly, and it takes real effort to rebuild if it has been damaged.
Your sending IP also carries a reputation, and this creates a complication we will come back to in a moment when we talk about shared infrastructure.
The engagement signals we know they look at
Inbox providers pay attention to how recipients engage with your mail once it has been delivered. The signals we know matter are open rate, click rate, and whether recipients reply to messages.
Open rate carries weight as a signal, though it has become a less reliable metric for senders specifically since Apple introduced Mail Privacy Protection in 2021. The way MPP prefetches images inflates open rates in Apple Mail, making it harder for senders to know whether a real person actually opened a message. Inbox providers track opens at the mailbox level rather than through tracking pixels, so they have a cleaner picture than senders do.
Click rate is a stronger signal of genuine intent. Someone who clicks a link in an email is telling the inbox provider, in the clearest way available, that the message was relevant enough to act on. That carries real weight.
Replies tend to be underestimated by most senders. When a recipient replies to a marketing or lifecycle email, they are effectively vouching for the sender to their own inbox provider. It is one of the clearest positive signals that exists, and if your emails are structured in a way that never creates an opportunity for a reply, you are leaving that signal on the table.
The signals we suspect but cannot confirm
This is where it gets more speculative, though not by much. The signals below are not things inbox providers have published detailed documentation about. They are logical inferences from how sophisticated filtering systems work, and they are consistent with patterns that deliverability specialists observe in practice.
Think about how often a recipient goes back and re-reads an email. A message that gets revisited suggests it contained something worth returning to. One that was opened and dismissed in two seconds tells a completely different story. Whether inbox providers measure this formally or simply infer it from other patterns, the underlying logic holds.
Dwell time works along similar lines. An email that someone spends real time reading is probably more valuable to that recipient than one that gets closed immediately. There are only so many ways to infer content quality from the outside, and time spent is one of the more reliable proxies.
Folder behaviour is worth thinking about too. Moving an email to a specific folder suggests it was wanted. Archiving it rather than deleting suggests the same. These are positive signals about intent even when there is no explicit click involved.
Forwards are probably one of the strongest signals of all. When someone forwards an email to another person, they are actively vouching for it to a third party. That level of secondary engagement is rare, which is likely why it carries disproportionate weight when it happens.
On the negative side, consistently deleting emails without opening them tells a story about how a recipient feels about a sender's subject lines or timing. At the individual level it means very little, but at scale, across thousands of recipients, it becomes a meaningful pattern.
Deleting without reading, which means opening briefly and then deleting immediately, is a worse signal. The recipient made an effort and still decided the message was not worth their time.
Bulk deletion is perhaps the most damaging behaviour of all. When someone selects multiple emails from the same sender and deletes them together in a single action, they are making a pretty unambiguous statement. Inbox providers can observe this behaviour directly and it would be naive to assume they do not act on it.
None of this is confirmed. Inbox providers do not publish their filtering algorithms in detail, and they have very good reasons not to. But if you were building a filtering system with access to all of this behavioural data, you would use it. It is far more predictive of whether a recipient wants a message than any technical check could ever be.
The shared IP problem
If you are sending on a shared IP address, your reputation is partly determined by what everyone else on that IP is doing, and you have no control over any of it.
Inbox providers see the full traffic pattern coming from an IP. They see complaint rates, bounce rates, and engagement signals across all the mail originating from that address. If another sender in the same IP pool is generating spam complaints or behaving badly, your mail is affected. You have no visibility into what they are sending, no ability to influence their practices, and no way to isolate yourself from the consequences.
That said, if you are a smaller sender on shared IPs, this is not a reason to panic. Shared infrastructure is perfectly sensible for businesses that are not yet sending at the kind of volume that warrants a dedicated IP, and the risk is manageable as long as you are using an ESP with strong terms and conditions that does not onboard poor-quality senders. A reputable ESP is selective about who it puts on its shared infrastructure precisely because it has a commercial interest in keeping those IPs clean. If you are with a provider that takes this seriously, you can have reasonable confidence that the other senders you are sharing with are operating to a similar standard. The problem tends to arise with cheaper or less selective providers where the onboarding bar is low and the shared pool reflects that.
For anyone sending at meaningful commercial volume where deliverability has real financial consequences, a dedicated IP with a proper warm-up programme is worth serious consideration. But for smaller senders on reputable shared infrastructure, the focus should be on your own sending practices rather than worrying about the neighbours.
What Google is doing with AI, and why it matters
The filtering picture described above has been evolving gradually for years. What is happening at Google right now is a bigger shift, and it is worth understanding clearly rather than treating it as background noise.
Google is applying Gemini to Gmail in ways that go well beyond filtering. The inbox itself is becoming an intelligent, curated experience rather than a simple list of messages ordered by time of arrival.
The Purchases folder is one visible example of this. Gmail now surfaces transactional emails in a dedicated area and presents them as visual cards showing shipping status, delivery dates and order details. This is clearly useful for recipients, and it also represents Google making curatorial decisions about how transactional mail gets presented, rather than simply passing it through.
The Promotions tab is undergoing a more consequential change. Promotional emails have always appeared in chronological order, most recent at the top. Google is rolling out relevance-based ordering in the Promotions tab, which means emails will be sorted based on what Gemini predicts the recipient is most likely to want to see. Recency, which senders have always been able to control by choosing when to send, matters less. Relevance, which depends on engagement history and content quality, matters more.
The practical implication is significant. A well-timed promotional email that would previously have sat at the top of the tab simply because it just arrived may now appear lower down if Gemini predicts low engagement likelihood. A less timely email from a sender with a strong engagement history may appear higher. This is Google making a deliberate bet that a better consumer experience is worth the disruption to senders who have relied on timing rather than relevance.
The direction of travel here is clear. Google is building an inbox that actively prioritises what individual recipients actually want. Senders who have invested in genuine relevance, healthy list management, and strong engagement signals are better positioned in that environment than those who have relied on volume and send-time optimisation.
Not sure how your sending programme looks from the inbox provider's perspective? A free 15-minute check-up takes a look at your domain reputation, authentication setup, and the signals you are generating before they become a problem. Book a free check-up.
Microsoft and Yahoo are watching the same kinds of signals
Microsoft and Yahoo have not made announcements equivalent to Google's Gemini integration, but they are not standing still. Microsoft's filtering systems, which cover Outlook.com and the large share of business email that runs through Microsoft 365, use machine learning to classify mail and assess sender reputation. The signals they use are broadly consistent with Google's approach, with engagement behaviour, complaint rates, and sending patterns all feeding into placement decisions.
Yahoo and its associated properties process a significant share of consumer email, particularly in the US. Their filtering has historically been more sensitive to complaint rates than some other providers, which makes list quality particularly important for senders whose lists include a meaningful proportion of Yahoo addresses. If you are generating spam complaints at a rate that would be borderline acceptable elsewhere, you are likely to feel the effects more acutely with Yahoo.
The underlying principle across all of these providers is the same. They are all trying to answer the same question on behalf of their users: is this email something this person actually wants? Every piece of data available to help answer that question more accurately is data they will use.
What this means for how you approach sending
Continuing to send to people who have not engaged with your mail for a long time is not neutral. Those recipients are generating negative signals each time they delete without reading, and those signals accumulate. Segmenting out unengaged contacts and either running a deliberate re-engagement programme or suppressing them is not just good list hygiene, it is reputation management.
The bigger shift in thinking, though, is about relevance. We are in a world now where relevance is everything, and that means sending email that is genuinely valuable to the person receiving it rather than convenient for the business sending it. Knowing your target audience and the different avatars within that audience, and being able to use whatever data points you have about them, allows you to understand what specific content is likely to resonate in relation to your brand or offering. That understanding is what makes dynamic content worth investing in, because when you can serve content that speaks to a specific type of customer rather than everyone at once, engagement follows naturally.
The flip side of relevance is giving recipients some control. Preference centres, where people can choose how often they hear from you and on what topics, and strict sunset policies that remove people who have stopped engaging, both work in your favour. Counterintuitively, sending less and to a smaller engaged audience almost always outperforms sending more and to everyone. One of the first pieces of advice we give clients is to stop sending so much, stop sending the same thing to the whole list, and start thinking seriously about the different types of customers they have and what each of them actually wants to receive. The question to ask is not what do we want to send, but what does this person want to get from us.
Transactional mail reliability matters more than most teams appreciate. Google's AI features are built on the assumption that transactional mail is something recipients genuinely want. If your transactional mail generates poor engagement or complaints because of inconsistent quality or poor targeting, those signals go into the same reputation pool as your marketing mail. This is also a good illustration of why sending different types of email from separate subdomains is worth doing properly. Having transactional mail on its own subdomain, with its own IP or set of shared IPs, is responsible architecture. It means the reputation of your transactional stream is not affected by the performance of your marketing campaigns, and vice versa. It also helps inbox providers like Google route transactional messages correctly, making it less likely that an order confirmation or password reset ends up sitting in the Promotions tab when it should be landing in the primary inbox.
The tools that give you some visibility
You cannot see everything inbox providers can see about your sending. But there are tools that give you a partial view, and monitoring them regularly is far better than having no data at all.
Google Postmaster Tools, available at postmaster.google.com, shows domain reputation, IP reputation, spam rate, and authentication pass rates for mail delivered to Gmail. It requires domain verification and is free to use. It is one of the most useful data sources available to any sender with meaningful Gmail volume.
Microsoft SNDS, the Smart Network Data Services tool at sendersupport.olc.protection.outlook.com, provides IP reputation data and complaint rate information for mail delivered to Microsoft properties. Also free, and worth checking on a regular basis.
Yahoo Sender Hub at senders.yahooinc.com provides reputation information and sender guidance for Yahoo Mail and associated domains. If Yahoo addresses make up a meaningful portion of your list, the data here is worth paying attention to.
None of these tools shows the full picture of how filtering decisions are made, and none of them surfaces the engagement signals we discussed earlier. But they are the closest thing to a window into how major inbox providers see your programme, and the information they surface is genuinely actionable when something is starting to go wrong.
In this series
Part 1: What sender reputation actually is. What Is Sender Reputation and How Do You Protect It?
Part 2: SPF, DKIM and DMARC in plain English. SPF, DKIM and DMARC Explained (And Why Getting Them Wrong Is Costly)
Part 3: How inbox providers actually filter mail (this post)
Part 4: Engagement signals: what counts and what doesn't (coming next)
Part 5: Sending architecture for growing programmes
Part 6: Monitoring that catches problems early
