← All posts Outbound

How to write cold email that survives an AI filter

Inbox providers now score intent, not just spam words. The copy that gets through in 2026 looks different from the copy that worked three years ago.

The old advice was about keywords: avoid "free", avoid excessive punctuation, do not use the word "guarantee". That advice describes a filtering generation that has largely been replaced.

Modern inbox filtering evaluates sender reputation, engagement history, structural patterns across many recipients and, increasingly, whether the message reads as bulk regardless of its wording. You cannot word your way past a system that is looking at whether four hundred near-identical messages left one domain this morning.

What the filters are actually reading

Roughly in order of weight: how recipients have historically treated mail from this domain, whether this specific recipient has ever engaged, structural similarity across the batch, link and tracking-pixel patterns, and only then content.

Which means most "copywriting for deliverability" advice is optimising the smallest term in the equation.

Write for the reply, not the filter

The useful reframing: the strongest deliverability signal you can generate is a reply. Replies tell the provider this is wanted mail, and they compound — a domain with a healthy reply rate gets more latitude on everything else.

So the copy goal is not "avoid triggering the filter". It is "get a response from a smaller number of better-chosen people", which happens to be the same thing.

The strongest deliverability signal you can generate is a reply.

Six things that hold up

Vary the structure, not just the fields

Mail-merge variation fools nobody. Four hundred messages with the same three sentences and a different company name are structurally identical, and structural similarity is precisely what batch detection looks for.

Real variation means different sentence counts, different orderings, different openings — genuinely different messages that happen to make the same argument. This is where language models earn their place in the stack, and it is a much better use of them than writing the pitch.

The test that matters

Seed inboxes across the major providers, send the actual campaign to them alongside the real list, and check placement — inbox, promotions or spam — per provider. Do it every campaign, not once at setup. Placement drifts, and reply rate will tell you eventually but placement tells you today.

Want this running on your pipeline?

We design and operate the system — outbound, paid and the data layer underneath. Book a call and we'll map it against your funnel.

Book a Strategy Call