How Modern ISP Content Scoring Systems Work

• July 2022
• Engineering Memo · External Release

The mental model of email spam filtering as a content scoring system — where specific words, punctuation patterns, and image ratios generate scores that determine inbox vs spam placement — is outdated. Modern ISP filtering at Gmail, Yahoo, and Microsoft evaluates messages through multi-dimensional models that weight sender reputation, authentication quality, and recipient engagement history far more heavily than message content. Content analysis still occurs, but it is interpreted in the context of these other signals rather than scoring independently.

This note explains how modern ISP content scoring actually works, which content signals still matter, which have been superseded by reputation signals, and what the correct approach to "content optimisation for deliverability" looks like given the actual filtering architecture.

The Multi-Signal Filtering Model

Modern ISP filtering systems — particularly Gmail, which has invested most visibly in ML-based classification — evaluate messages across five signal categories simultaneously, with different weights applied to each category based on the sender's characteristics and history:

1. Sender identity signals (authentication): Does the message pass SPF, DKIM, and DMARC? Is the sending IP listed in the domain's SPF record? Is the DKIM signature aligned with the From: domain? These are binary-ish checks — authentication either passes or fails. Failed authentication produces significant negative weight; passed authentication is essentially neutral (it enables, rather than improves, reputation assessment).

2. Sender reputation signals: What is the domain's historical complaint rate? What is the IP pool's reputation tier? What is the domain's historical engagement rate? These signals carry the highest weight in the classification decision for senders with established history. A sender with High domain reputation at Gmail receives substantial inbox-routing preference regardless of message content. A sender with Low domain reputation routes to spam regardless of content quality.

3. Recipient engagement signals: Has this specific recipient previously engaged with email from this sender? Have they marked previous messages as spam? Have they unsubscribed or moved messages to spam? These individual-level signals personalise the classification outcome — the same message from the same sender may route to inbox for recipients who have historically engaged with it and to spam for recipients who have historically ignored it.

4. Infrastructure signals: Does the sending IP have a PTR record? Is FCrDNS aligned? Is the sending infrastructure consistent with the sender's established pattern (same domain, similar IP range, consistent EHLO)? Infrastructure inconsistencies are mildly negative signals that contribute to classification alongside the higher-weight signals above.

5. Content signals: Does the message contain URLs associated with spam or phishing? Does the content structure match patterns associated with legitimate senders in this category? Is the image-to-text ratio within the range expected for this type of message? Content signals are the fifth category — they contribute to classification but are outweighed by signals 1–4 for established senders.

Content Signals That Still Matter

While content signals are secondary to reputation in modern filtering, they are not irrelevant. The content signals that carry genuine filtering weight in 2022:

URL reputation. Links to domains associated with phishing, malware, or spam are strong negative signals that can override positive reputation signals. Gmail's Safe Browsing database, Spamhaus DBL, and SURBL are queried for URLs in the message body. A legitimately branded email from a high-reputation sender that happens to link to a known malicious domain will be blocked regardless of the sender's reputation. This is the most impactful content signal in modern filtering, and the one most worth monitoring — ensuring that all URLs in campaign emails are to legitimate, reputation-clean domains.

HTML structure consistency. Messages with HTML that matches the structural patterns established by the sender's previous messages are treated more consistently than messages with dramatically different structures. A sender who consistently sends multi-part HTML with a single header image and several text sections, and then sends a message that is entirely image-based with no HTML text, has changed their message structure in a way that some filters treat as anomalous. The consistency of HTML structure across a programme's campaigns is a minor positive signal.

Image-to-text ratio in edge cases. Entirely image-based messages with no HTML text content provide no text for content classification. This is less a spam signal per se and more an absence of a positive signal — the filter has no content to evaluate and relies more heavily on reputation signals for classification. For High-reputation senders, entirely image-based messages typically still route to inbox because reputation signals dominate. For Medium or lower-reputation senders, the absence of text content removes a potential positive signal that might have contributed to inbox routing.

Phishing-pattern text in subject or body. Specific language patterns associated with phishing campaigns — urgency language combined with requests for credentials, impersonation of financial institutions, certain combinations of trigger words that consistently appear in phishing messages — remain effective content signals because they are predictive of malicious intent regardless of sender reputation. Legitimate senders do not use these patterns, so they function as reliable signals that the ML model has identified as predictive of spam/phishing regardless of sender reputation.

Content Signals That No Longer Matter (As Much)

The content signals that dominated early-era spam filters and continue to be cited in outdated deliverability guidance:

Specific words and phrases ("FREE", "CLICK HERE", exclamation marks). These rule-based triggers were the foundation of first-generation spam filters (SpamAssassin-era) and are still referenced in outdated deliverability guides. Modern ML-based filters have largely superseded word-list scoring. A legitimate sender with High domain reputation can use "FREE SHIPPING THIS WEEKEND!!!" in a subject line without material inbox placement impact. An illegitimate sender with Low domain reputation who carefully avoids all trigger words still routes to spam. The word-content signal is small relative to the reputation signal and is outweighed by it in virtually all cases for established senders.

Specific image dimensions or image-to-text ratios (within normal ranges). While entirely image-based messages lack text content (as noted above), normal ranges of image-to-text ratios within multi-part HTML messages do not produce meaningful filtering signals. A message that is 70% image content and 30% text routes the same as one that is 30% image and 70% text for a High-reputation sender. The image ratio advice from the early filtering era — "keep images under 40% of total content" — has not been relevant to modern filtering for several years.

ALL CAPS in subject lines. All-caps in the subject line generates no meaningful filtering signal at modern ISPs. It may reduce open rates because recipients perceive it as aggressive or unprofessional, but it does not route messages to spam.

What "Content Optimisation for Deliverability" Actually Means Today

Given the actual signal weighting in modern filtering, "content optimisation for deliverability" in 2022 means something different from what it meant in 2010. Today, it means:

Optimise content for engagement, not for filter avoidance. Content that recipients genuinely want to open and click generates positive engagement signals that are the most impactful positive reputation inputs. A message that avoids every theoretical filter trigger but recipients don't open will produce weak engagement signals that contribute to reputation erosion. A message that some outmoded guides would flag for content issues but recipients actively engage with produces strong positive signals. Optimise for the human, not the filter — the filter evaluates what the human does with the message, so optimising for human engagement automatically optimises for the filter signal.

Maintain URL reputation hygiene. This is the content dimension with the highest genuine filtering weight. Audit every URL in campaign emails before sending — confirm destination domains are clean, that redirect chains don't pass through reputation-impaired domains, and that tracking link domains are not shared with spam-associated senders. URL reputation issues are the most likely content-level cause of deliverability problems for senders with otherwise good reputation.

Maintain structural consistency. Don't dramatically change email structure between campaigns without understanding the filtering implications. Major structural changes — moving from text-heavy to image-heavy, changing the header template, adding new content sections — should be tested in a small segment before full deployment to confirm they don't affect delivery rates or complaint rates.

The conclusion from modern filtering architecture is consistent with the broader deliverability message: content quality matters, but through the engagement signals it generates, not through the filter-avoidance rules it satisfies. The most effective "content optimisation for deliverability" is producing content that recipients genuinely value — which produces the engagement signals that modern reputation-based classification systems reward above all other signals, including content signals specifically.

The Plain Text Alternative Part

Multi-part MIME messages that include both an HTML part and a plain text alternative part have been recommended deliverability practice since the early days of HTML email. The plain text alternative serves several purposes: it provides readable content for email clients that don't render HTML (rare but still present in some corporate environments), it provides text content for classifier systems to evaluate alongside image content, and it is included in DKIM signing for the complete message.

For modern filtering, the plain text alternative's deliverability benefit is modest — the primary classification signal is reputation, not content, and the plain text part provides content for classifiers that already have the HTML content to evaluate. Where the plain text alternative matters most is in ensuring that the DKIM signature covers the complete multipart message — if the sending system generates HTML-only messages without the plain text alternative, the message structure is simpler and the DKIM signature is over a simpler message, which is not inherently problematic but is a deviation from the multi-part standard that most ISPs associate with well-configured legitimate senders.

MailWizz generates multi-part messages with HTML and plain text alternatives when both are configured in the campaign. Generating the plain text alternative manually for each campaign is time-consuming; many programmes auto-generate the plain text from the HTML by stripping HTML tags. The auto-generated plain text is often not well-formatted (long run-on lines from block HTML tags, broken link text), but this is sufficient for the technical purpose — providing a text alternative that is readable and structurally complete.

Subject Line Length and Preview Text

Subject line length recommendations have shifted over time in deliverability guidance, and it is worth clarifying what is actually known vs what is speculation. For deliverability (inbox vs spam routing): subject line length does not materially affect spam classification in modern filtering. Longer subject lines are not more likely to route to spam than shorter ones, and vice versa. The spam classification happens based on reputation signals, not character count.

For engagement (open rates): subject line length does affect open rates in client display — many email clients truncate subject lines at approximately 60 characters in the preview pane. Content that is cut off in the preview pane may communicate less effectively than content that fits. The optimal subject line length from an engagement perspective is determined by where the most important information appears and whether it is visible before truncation — a per-programme decision based on the typical subject line content, not a universal rule.

Preview text (the line of text visible in the inbox view beneath the subject line on many email clients) is not a standard email header — it is generated from the beginning of the message body. Programmes that place hidden preview text at the top of the HTML template (using a small, hidden text element with CSS visibility:hidden or similar) can control what appears as preview text. This is a standard practice that ISP filters do not penalise — hidden preview text using CSS is not equivalent to hidden spam content, which typically uses techniques like white text on white background that are clearly deceptive.

Testing Content for Deliverability: What Tools Can and Cannot Tell You

Content testing tools — services that score a message against spam filter rules and report which rules triggered — have significant limitations that operators should understand before interpreting their output. Tools based on SpamAssassin rules (which many commercial content testing tools use under the hood) are testing against the rule-based scoring of a filtering system that most major ISPs no longer use as their primary classifier. A SpamAssassin score of 3 does not predict Gmail inbox routing; it predicts how the message would score against rules that Gmail has largely superseded with its ML-based classification.

Seed testing services — which send the actual message to seed addresses at Gmail, Yahoo, and Microsoft and report where it lands — are more predictive of actual inbox placement but have their own limitations: seed addresses may have different engagement histories from real recipients, seed networks can become known to ISPs and may be treated differently, and seed results represent a point-in-time snapshot that may not reflect the sender's typical recipients' inbox placement experience.

The most reliable content-deliverability testing approach: send the campaign to the real list, monitor real engagement and complaint rates in the first hours, and track Postmaster Tools spam rate in the 48 hours after send. The real recipient population's behaviour is the most accurate signal of the content's actual performance. Pre-campaign seed testing provides directional guidance, but the real population's response is the definitive measurement.

Content Personalisation and Filtering

Personalised content — messages where the body includes recipient-specific data such as the recipient's name, recent purchase history, or location-based content — has a relationship with filtering that is worth understanding. Personalised content is not inherently more likely to reach the inbox than non-personalised content; filtering is based on sender reputation and recipient engagement signals, not on content personalisation.

Where personalisation matters for deliverability is through its effect on engagement. Well-executed personalisation typically produces higher open rates and higher click rates because the content is more relevant to the specific recipient. These higher engagement rates contribute positively to the reputation signals that drive inbox placement. The deliverability benefit of personalisation is mediated through engagement, not direct — personalisation → higher engagement → better reputation signals → better inbox placement. The mechanism is indirect but real over time.

Poorly executed personalisation — for example, "Dear [FIRST_NAME]" failures where the variable substitution doesn't work correctly, or personalisation based on stale or incorrect data — can generate complaint rates from recipients who find the error off-putting. This complaint-generation effect damages reputation through the same engagement signal channel, in the negative direction. Quality control of personalisation data and template rendering is therefore both a user experience concern and a deliverability concern.

The full picture of content and deliverability in modern email filtering is this: content affects deliverability almost entirely through its effect on recipient engagement signals (opens, clicks, complaint rates), rather than through direct content scoring by ISP filters. The exception is URL reputation, which is a genuine direct content signal. Everything else in the "content deliverability" conversation is actually about engagement quality — which is a much more tractable optimisation objective than trying to reverse-engineer ISP content scoring rules that change constantly and are outweighed by reputation signals for established senders. Optimise content to generate recipient value and positive engagement; the filtering outcomes will follow from the engagement signals that high-quality, valued content produces.

The Promotions Tab: Classification vs Filtering

Gmail's Promotions tab is a frequent source of confusion in discussions about content and filtering. Messages routed to the Promotions tab are not being spam-filtered — they are being categorised. The classification into Primary, Social, Promotions, or Updates tabs is a separate system from the spam vs inbox classification, and it is driven primarily by message content signals (commercial language, promotional offers, list-unsubscribe headers, bulk sending infrastructure) rather than sender reputation signals.

A message routed to Promotions has passed all spam filtering — it is in the inbox (Promotions is part of the inbox category in Gmail's architecture, not a spam folder). The deliverability concern for Promotions is not the same as the deliverability concern for spam — a message in Promotions is accessible to the recipient and generates engagement signals when opened or clicked; a message in spam may never be seen at all.

Content signals that influence tab categorisation include: the presence of List-Unsubscribe headers (associated with bulk senders → Promotions), commercial language in the subject or body (sale, discount, offer → Promotions), bulk sending infrastructure signals (sending from ESPs, high volume sending), and the sender's established tab for a specific recipient (if the recipient has previously moved messages from this sender to Primary, Gmail remembers this). Reducing Promotions tab routing is possible through content changes that make the message appear more personal and less commercial, but this approach should be evaluated against whether the content changes affect engagement rates — some messages naturally belong in Promotions because they are commercial promotions, and recipients check Promotions regularly for this type of content.

The distinction between Promotions routing and spam routing matters for how operators prioritise deliverability work. If the programme's problem is Promotions tab routing — visible because open rates are lower than expected but Postmaster Tools shows High domain reputation — the intervention is content-based: more personal writing style, reduced commercial language density, List-Unsubscribe header consideration. If the programme's problem is spam folder routing — visible because open rates are very low and Postmaster Tools shows Medium or Low domain reputation — the intervention is reputation-based: list hygiene, complaint rate reduction, engagement-based segmentation. Conflating the two leads to applying content interventions to a reputation problem, which produces no improvement.

Understanding the modern content scoring architecture — the five-signal model, the relative weights, the distinction between content that directly affects filtering (URL reputation) and content that indirectly affects filtering through engagement (everything else), and the separate classification question of Promotions vs Primary tab routing — gives email operators the mental model needed to diagnose content-related deliverability issues correctly and apply interventions at the correct layer. The widespread outdated guidance on content-based deliverability optimisation continues to focus attention on signals that modern filtering has made marginal, while the actual levers — reputation management, engagement optimisation, URL hygiene — receive insufficient attention. This note is an attempt to rebalance that attention toward the signals that actually determine filtering outcomes in the current filtering environment.

The programmes that consistently achieve excellent inbox placement are not those that spend the most time on content optimisation — they are those that invest in the reputation signals that modern filtering systems actually weight: engagement quality, complaint rate management, list hygiene, and authentication correctness. Content is a vehicle for generating these signals, not a direct deliverability lever in itself. Once this distinction is clear, the entire approach to "deliverability-friendly content" shifts from avoiding filter triggers to creating genuinely valuable content that recipients engage with — which is both better for deliverability and better for the email programme's commercial outcomes.