A tracking pixel can't tell you who loaded it. A security scanner and your dream prospect trigger the exact same request. So how does any tool honestly claim to know a human opened your email?
By not looking at the pixel. By looking at everything around it.
Why is this hard in the first place?
Because the raw event carries almost no meaning on its own. One pixel load says "something fetched this image." That's it.
And a lot of somethings fetch images. Apple Mail Privacy Protection pre-fetches tracking pixels on Apple's servers, so every Apple Mail recipient looks like an opener even when they never read the email. Corporate security scanners like Mimecast, Proofpoint, and Microsoft Defender open every link and image in incoming mail hunting for malware, and 15 to 25 percent of B2B recipients sit behind one. Gmail serves images through its own proxy, which strips the direct device signal.
Add it up and raw open rates on B2B lists typically run 2 to 3 times inflated. A reported 70 percent open rate is often 25 to 35 percent real reads. Counting pixel loads stopped being tracking a while ago. Now it's just counting.
What signals give humans away?
Behavior. People and machines both load pixels, but they don't load them the same way.
Real people open in patterns. They open at normal hours, their hours, in their timezone. They open on real devices with real client signatures. Sometimes they come back to the same email a day later and re-read it, which is one of the most human moves there is. No scanner re-reads your proposal because it's mulling the price.
Machines have their own tells. They fire instantly after send, often within seconds. They frequently hit many links at once, because a scanner checking for malware doesn't pick its favorite link, it follows all of them. They come from network origins that don't look like a person's device.
So the useful evidence lives in three places: request timing, network origin, and client signature. Any single signal can mislead. Together, they usually don't. The glossary entry on [email open rate](/glossary/email-open-rate) covers why the headline number needs this context.
How does the model use those signals?
It grades instead of guessing. Outsolvi scores every open on a ladder from Tier 1, high-confidence human, down to Tier 5, bot or scanner.
An open at 2pm on a Tuesday from a real device, on a thread the same contact re-opened yesterday? That grades high. A burst of fetches two seconds after send, touching every link in the email, from infrastructure that isn't a person's network? That grades as a machine.
The easy cases get handled with certainty. Apple MPP pre-fetches are the clearest machine-open class there is, and they stay filtered, full stop. The genuinely ambiguous cases, like Gmail proxy opens, get judged on their full signal picture rather than blindly counted or blindly discarded.
Then comes the floor: opens below 25 percent confidence are excluded from your counts entirely. Your dashboard shows opens the model would stand behind, not everything that twitched. The complete grading approach is written up on the [methodology page](/methodology), with the tier system detailed on the [confidence scoring feature page](/features/confidence-scoring).
How accurate is it?
Accurate enough to act on, and honest about its limits. The model agrees with human-rated classifications 95 to 98 percent of the time. False positives at Tier 1, the tier your follow-up decisions should key off, run under 2 percent.
Just as important: it keeps learning. Email providers change behavior. Proxy patterns shift, scanners get stealthier, new clients appear. A filter written as a fixed rule two years ago is quietly wrong today. A model that retrains as providers change stays right.
What if you disagree with a call?
Look at the evidence yourself. The [Diagnose view](/features/diagnose-open-evidence) lets you click any open and see why it counted or was filtered, in plain language.
This matters more than the accuracy stats, honestly. Plenty of tools show you a number. Very few will show their work on a single open when you challenge it. When you [compare trackers](/compare), that's the test worth running: pick one open on the dashboard and ask why it's there.
What does this mean for your open rate?
It'll drop. And that's the point.
A graded open rate is lower than a raw one because the machines are gone. What's left is the number you can actually use: real people, really reading, worth a follow-up. No score at all would be better than a wrong score, but you don't have to settle for either.
Outsolvi runs $7 a month billed yearly, or $12 monthly, with a 14-day free trial, no credit card, and a 7-day money-back guarantee. Try it against your current tracker for two weeks and see how many of your "opens" were ever people at all.