Email Marketing Deliverability
Email Spam Triggers: Words and Practices to Avoid
Why the famous word lists are mostly obsolete, and the behavioural signals that replaced them.
On this page
The short version
- The circulated word lists are largely obsolete. Filters moved to behavioural and reputational scoring years ago, and legitimate senders use those words constantly without consequence.
- What does still get scored is structural: link and domain reputation, image-to-text ratio, mismatched or obscured links, and sudden changes in sending pattern.
- The strongest single predictor of being filtered is your own recipients' behaviour, not any property of the message.
Search for spam trigger words and you will find lists of several hundred terms — free, guarantee, act now, click here — presented as things to avoid. Those lists date from an era of rule-based filtering and have been recirculated ever since.
They are not entirely wrong; they are wildly disproportionate. Every major retailer sends email containing the word free every week, and it arrives. The words are a small input into a model whose dominant inputs are elsewhere.
Believing the lists is costly in a specific way: it produces contorted copy — "f-r-e-e", "gu@rantee" — which is itself a pattern filters recognise, because it is what people trying to evade filters do.
What is actually scored
The inputs below are the ones senders can observe having an effect. They divide into things about your sending history, things about your list, and a smaller set of things about the message.
Real signals, ranked
| Signal | Group | Weight |
|---|---|---|
| Complaint rate | Recipient behaviour | Very high |
| Engagement ratios across the list | Recipient behaviour | Very high |
| Authentication passing and aligned | Infrastructure | High, and binary |
| Hard bounces and unknown users | List quality | High |
| Spam trap hits | List quality | High |
| Reputation of linked domains | Message | Moderate |
| Image-to-text ratio | Message | Low to moderate |
| Link text not matching link target | Message | Moderate |
| Specific words | Message | Low |
The bottom row is the one the folk model is entirely about. It is the least of nine, and it only tends to matter when several of the others are already poor.
The message-level things worth attention
Link reputation is the underrated one. If your email links to a shortened URL, a domain with a poor history, or a redirect chain that ends somewhere unexpected, that counts against the message regardless of who sent it. Link to your own domain wherever you can, and avoid public URL shorteners entirely — their reputation is the aggregate of everyone using them.
Mismatched links matter for the same reason obfuscation does: text that says one destination while the anchor points at another is a phishing pattern. It also happens accidentally, when a template's placeholder text survives an edit.
Image-heavy emails carry weak signal on their own and strong signal at the extreme. An email that is a single image with a link on it has no text to evaluate, cannot be read with images off, and matches a long-standing evasion technique.
Sudden changes are themselves a signal
Filters model your normal behaviour. A sender who transmits fifteen thousand messages every Tuesday and then sends four hundred thousand on a Thursday has done something anomalous, and anomalies get scrutiny even when everything else is healthy.
The practical version: ramp volume increases rather than stepping them, keep your sending schedule roughly consistent, and treat a large one-off campaign as something to build up to over a couple of weeks.
Two emails with identical copy
Same words, same offer, different outcomes — because the words were never the variable.
Filtered
- Sent to a list with a 22% never-opened share
- DKIM signing with the platform's domain
- Links via a public URL shortener
- One large image, minimal live text
- Volume four times the usual
Inboxed
- Sent to engaged segments only
- Domain-aligned authentication passing
- Links to the sender's own domain
- Live text with images as enhancement
- Volume consistent with recent weeks
What to check instead of a word list
- Complaint rate per segment, not just overall
- The share of the list that has never engaged
- Authentication passing and aligned to your domain
- No public URL shorteners anywhere in the message
- Link text matches link destination
- The email has real text, not just images
- This send's volume is not a large jump from recent ones
Frequently asked questions
Should I run my email through a spam-score checker?
They are useful for catching mechanical faults — a missing unsubscribe, a broken authentication setup, a bad link. Treat the content score itself with scepticism, because it cannot see the thing that matters most, which is how your recipients behave.
Does the word "free" in a subject line hurt?
Not meaningfully, on its own, from a sender in good standing. It contributes more when it appears alongside several other weak signals, which is the general pattern: no single content signal decides anything.
What about excessive punctuation and capitals?
A subject in full capitals with several exclamation marks does carry weight, and it also reads badly, so there is no tension between the deliverability answer and the editorial one.