Backlink Quality Scoring Beyond DR and UR

TL;DR: Domain Rating (Ahrefs), URL Rating, Domain Authority (Moz), and the other single-number aggregate metrics that dominate link analysis are useful for triage and harmful for decision-making. The Google-side ranking system has used multi-dimensional graph features since at least the Hilltop and TrustRank era; the right operating posture is to score links on six independent dimensions (topical relevance, page-level traffic value, editorial quality, anchor distribution, velocity, decay) rather than to compress them into a single number. The single-number metrics are vendor-graph proxies for a target Google actually computes very differently.

A note on tools and brands. Ahrefs, Moz, Semrush, Majestic, and the named operating-case sources appear in this essay as the available reference points for vendor-graph metrics and published research. Quantitative claims framed as advisory-engagement observation come from anonymized partner operators, not from the named companies. Public claims from Ahrefs, Google, Moz, and Majestic are attributed inline.

---

The Single-Number Problem

The dominant metrics in link analysis (Ahrefs Domain Rating, Ahrefs URL Rating, Moz Domain Authority, Majestic Trust Flow and Citation Flow) are useful for one thing and dangerous for everything else. They are useful as triage filters: at a glance, a DR 80 site is roughly more authoritative than a DR 20 site, and the ordering is correct on average. They are dangerous as decision metrics because they collapse a high-dimensional graph problem into a single scalar, and the dimensions they suppress are the ones that determine whether a link actually moves ranking.

Ahrefs's own glossary documents the limitation. Domain Rating is computed from the link graph in the Ahrefs index, not from Google's index, and Ahrefs explicitly cautions against using DR as a standalone metric of site quality. The DR-to-rank correlation in Ahrefs's published research has consistently come in around r = 0.09 for individual queries, which is to say, weak. Sites with high DR can lose to sites with much lower DR for specific queries; sites with manipulated DR can score above genuinely better sites because DR is computable from links alone and links can be bought.

The single-number problem is structural. Any scalar metric that aggregates a graph property over an entire domain must trade off precision in one dimension against precision in another. Topical relevance, anchor-text health, editorial quality, traffic value, link velocity, and decay all matter for ranking. Compressing them into a single number lets the metric track the average, but the average is exactly the wrong statistic when individual links carry highly heterogeneous value.

This essay proposes the alternative: a six-dimensional scoring framework that scores each link on its own merits along independent axes. The framework is a research direction, not a validated metric, and the weights are calibrated against advisory partner data rather than published ground truth. The point is not to replace DR with another single number. The point is to refuse the compression.

Where DR Comes From and What It Compresses

To use vendor-graph metrics well, it helps to know how they are computed. The published methodology for Ahrefs Domain Rating is approximately the following: take the link graph of the Ahrefs index, compute a PageRank-style iterative score per page, aggregate to the domain level, and normalise to a 0-to-100 logarithmic scale. URL Rating (UR) is the same calculation but reported at the page level rather than the domain level. Moz Domain Authority uses a different model (a machine-learned predictor of Google ranking trained on a sample of SERPs) but reports on the same scale.

The compression is in two places. First, the underlying graph is the vendor's crawl of the web, not Google's index. Vendor crawls miss large portions of the link graph (especially newer links and links from regions the vendor crawls less aggressively) and over-represent links that the vendor surfaces aggressively. Ahrefs's index is large and well-maintained; Moz's index is smaller; Majestic's index is differently structured. None of them is Google's index.

Second, the aggregation strips out the dimensions that matter. A domain with 10,000 links from one topical neighbourhood is different from a domain with 10,000 links spread across hundreds of unrelated neighbourhoods, but DR cannot distinguish them. A domain with 10,000 links acquired at a flat rate over five years is different from a domain with 10,000 links acquired in two months and then nothing, but DR cannot distinguish them either. A domain whose links are mostly in editorial content is different from a domain whose links are mostly in sponsored placements or comment sections, but DR cannot distinguish them.

This is the single-number problem in concrete form. The dimensions DR compresses are exactly the dimensions Google's published spam policy and the academic graph-theoretic literature say matter most.

What Google Says About Link Quality

Google's published guidance on link quality is unusually direct compared to most ranking-signal documentation. The Search Quality Rater Guidelines, the Spam Policies for Google Web Search, the Webmaster Guidelines, and the public statements from Mueller and Illyes converge on a small number of operating principles.

The first principle is topical relevance. A link from a topically related page is treated differently from a link from a topically unrelated page. The mechanism, in graph terms, is the reasonable-surfer model and its successors: Google's link-weighting algorithms have, since at least the original PageRank patent updates, given more weight to links that a hypothetical user is likely to actually click in context. Topical relevance is the strongest predictor of click likelihood, and click likelihood is the proxy for editorial endorsement.

The second principle is editorial intent. Google's spam policy explicitly prohibits link schemes including paid links without rel="sponsored" or rel="nofollow", excessive link exchanges, large-scale guest-post networks, and links acquired through manipulation of the editorial process. The 2022 Link Spam Update and the 2024 follow-ups have meaningfully tightened detection of these patterns. SpamBrain, Google's spam-detection system, analyses relational patterns across the linking domain, the linked domain, the topic cluster, the anchor distribution, and the historical behaviour of the network. The unit of analysis is not the link; it is the link in its graph context.

The third principle is link decay. A link present at the time the algorithm assessed the page contributes value; a link removed contributes nothing. Google's caching infrastructure and rolling reassessment mean that links can lose value over time even without being removed, if the linking page itself loses authority or relevance.

The fourth principle, increasingly emphasised in 2024 and 2025 statements, is the source-page's own ranking signals. A link from a page that itself ranks well, attracts qualified traffic, and demonstrates editorial standards is worth materially more than a link from a page that ranks for nothing and has no measurable traffic. This is the part that traffic-aware metrics (Ahrefs Page Traffic, Semrush Estimated Traffic) get closer to, but still imperfectly.

The six dimensions in the framework below derive from these published principles plus the broader academic graph-theoretic literature on link quality.

The Six-Dimensional Framework

Each link should be scored independently on six axes. The axes are chosen to be approximately orthogonal: scoring high on one does not mechanically imply scoring high on another. The framework is operational: it produces a per-link vector rather than a per-link scalar, and decisions about which links to pursue, which to disavow, and how to interpret a competitor's profile become more legible.

A link's overall value is best treated as a vector in this six-space, not a scalar. For some operating decisions, a weighted sum is unavoidable (which of 200 prospects should we email first?), but the weights should be situation-specific rather than fixed. A site building topical authority in a narrow niche should weight relevance heavily; a site building brand visibility should weight traffic value and editorial quality; a site under manual-action risk should weight anchor distribution and velocity.

The weights above are not universal. They are operational starting points that have produced reasonable triage in advisory engagements, and they should be revised against the operator's specific risk posture, competitive context, and current ranking.

Topical Relevance: The Underweighted Dimension

Topical relevance is the dimension that vendor-graph metrics most systematically underweight. The reason is structural: DR, UR, and DA aggregate across all referring pages without conditioning on topic. A link from CNN to a SaaS landing page is weighted the same in DR terms as a link from a SaaS publication to the same landing page, even though Google's link-quality system treats them very differently.

The academic foundation for topical link quality goes back to the Hilltop algorithm (Bharat and Mihaila, 2001), which proposed that link weight should be conditioned on whether the linking page is a recognised "expert" page on the topic of the linked page. Topic-sensitive PageRank (Haveliwala, 2002) generalised this by computing PageRank separately for each topic vector and combining them according to query context. Both papers anticipated the operating shift that happened inside Google over the following decade: link weight became conditional on topical context.

The operating signal is consistent. In partner data, links from publications within the linked site's vertical (defined by ODP-style topic taxonomy or by topic-vector cosine similarity) produced two to four times the ranking effect of links from publications outside the vertical, conditional on equal DR. The variance is large, and we have no published ground truth to validate this rigorously, but the directional pattern has held across multiple verticals.

The implication is operational. When evaluating a link prospect, the first question is not "what is the DR?" but "is the linking publication topically relevant to the linked page?" A DR 40 publication squarely within the niche is typically more valuable than a DR 80 publication outside it. Most link-building tools, including the major SaaS platforms, do not surface this question well by default.

Page-Level Traffic Value: Going Beyond Domain Aggregates

Page-level traffic value is the dimension that has improved most in vendor tooling over the past five years. Ahrefs's Page Traffic and Semrush's Estimated Traffic both attempt to estimate the organic traffic of the specific linking page, not just the domain. The estimates are imperfect (vendor traffic estimates are typically within a factor of two of ground truth, with systematic bias by vertical), but they are far more useful than domain-level aggregates.

The argument for page-level traffic is straightforward. A link sits on a specific page. That page either attracts qualified users or does not. A page that ranks for high-value queries and attracts 10,000 monthly organic visitors is a different kind of link source than a page that ranks for nothing and attracts no traffic, even if both pages live on the same domain.

Google's link-quality system, by Mueller's repeated descriptions, weights links by the source page's own ranking signals. The exact form of the weighting is not published, but the operating implication is that page-traffic-weighted link metrics are closer to what Google computes than domain-aggregated metrics. Ahrefs's "Page Traffic" plus a domain-trust signal is meaningfully closer to the right model than Domain Rating alone.

A useful operating metric is the link's "estimated organic value" (EOV): the source page's estimated traffic, weighted by the topical relevance of the page to the link target, weighted by the link's position in the page (above-the-fold editorial body is materially more valuable than footer or sidebar), weighted by an estimate of click-through to the linked URL. EOV is a vendor-noisy estimate, but it is closer to the truth than DR.

Editorial Quality and the Placement Question

Editorial quality is the dimension most resistant to automation. A link can be inside an editorial body paragraph, where a working journalist or author intentionally included it as a reference, or it can be in a sidebar widget, footer block, sponsored placement, comment section, or programmatically generated list. Google's link-quality system, by published policy, treats these as fundamentally different.

The published spam policies are explicit. Paid links without rel="sponsored" or rel="nofollow" are link-scheme violations. Links in widely-syndicated infographics, where the embedded link is part of the artefact, are downweighted. Links in comment sections are typically nofollowed by the host platform and contribute nothing in the dofollow link graph. Links inside footer-block link exchanges are flagged. Links in dropdown navigation across thousands of pages on a partner site are treated as a single linking entity, not as thousands of independent endorsements.

In partner data, the share of a profile that is genuinely editorial (placed by an author in the body of a piece written about a related topic) correlates more strongly with ranking improvement than total link count. The teams we have advised that focused exclusively on editorial placements typically built smaller link profiles but earned more ranking value per link.

The operational difficulty is that classification at scale is hard. Vendor tools provide some signals (anchor placement, surrounding text length, page type) but the final call typically requires manual inspection of a representative sample. A scalable middle-ground is to classify a sample of the profile manually, train a lightweight classifier on the sample, and run the classifier across the full profile, accepting that classification noise will be material at the page level but tolerable at the aggregate.

Anchor Distribution: The Manipulation Signal

Anchor text distribution is the single dimension where over-optimisation is most directly detectable and most likely to trigger algorithmic or manual action. The published Google guidance has been consistent for years: anchor text that is excessively exact-match for commercial keywords is the manipulated-link signature, and it is the signature SpamBrain is best at recognising.

The natural anchor-text distribution for an organic site has a characteristic shape. Branded anchors (the site name or its variants) dominate, typically 40 to 70 percent of the profile. Naked URL anchors (https://example.com or example.com) make up 10 to 25 percent. Generic anchors (click here, read more, source) make up 5 to 15 percent. Partial-match anchors (keyword variations embedded in longer sentences) make up 10 to 20 percent. Exact-match anchors (the precise commercial keyword the page targets) are typically 1 to 5 percent.

The manipulated profile inverts this. Exact-match anchors are over-represented (often 20 to 40 percent), branded anchors are under-represented, and the distribution looks like the page was deliberately built to rank for a specific keyword rather than to attract organic editorial coverage. SpamBrain's anchor-distribution detector is by all accounts strong at recognising this signature.

The chart represents typical shapes, not universal ones. Some genuinely successful editorial sites have higher exact-match shares because the editorial language naturally uses the keyword (review sites linking to products often use the product name as the anchor). The interpretation requires context. A 35-percent exact-match share is suspicious in a B2B SaaS profile; it is normal in a movie-review aggregator.

The operating implication is to audit the anchor distribution at the linked-URL level, not just the domain level. A domain can have a healthy aggregate distribution while a single high-priority page has a dangerously over-optimised distribution targeting one commercial keyword. The aggregate hides the page-level problem.

Link Velocity: Why Acquisition Pace Is a Signal

Link velocity, the pace at which a domain acquires new referring domains over time, is a signal both Google and the major spam-detection systems use. The reasoning is that natural link acquisition follows a recognisable temporal pattern: bursts of links around newsworthy events (product launches, major announcements, media coverage), followed by gradual baseline accumulation, with no sudden flat-rate plateaus or zero-to-spike anomalies.

Managed or paid campaigns produce different patterns. A guest-post campaign that fires 200 placements in a single quarter and then nothing looks unnatural in the velocity series. A private blog network that drip-feeds 50 links per month, every month, for two years, looks unnatural because real coverage does not occur at a constant rate. Spamming campaigns that produce hundreds of low-quality links in a week and then disappear are the most obvious case.

In partner data, velocity-related issues have been a recurrent finding. Sites that ranked well for years and then suddenly lost ranking have, on inspection, often shown a velocity spike in the prior three to six months, even when the spike was the product of an inadequately-disclosed paid campaign rather than outright manipulation. The detection on Google's side, by all available evidence, has tightened materially through the 2022 to 2024 spam-update cycle.

The operating implication is that link campaigns should be paced to look organic. This is not a prescription for slowness; large genuine PR events do produce spikes. It is a prescription for distinguishing genuine spikes (tied to a verifiable external event) from constructed spikes (no external event, just campaign machinery).

Decay: The Time Dimension

The last dimension is decay. A link's value is not constant over time. Three decay mechanisms operate.

First, link removal. The linking publication may take down the article, restructure the site, change ownership, or have an editor decide to retire the link. In a 2017 longitudinal study, Hindrik Kuzee and colleagues at Distilled (now Brainlabs) reported that approximately 8 to 12 percent of editorial links acquired through digital PR were no longer present 24 months later. The decay rate varies by vertical: technology and finance publications churn more than legacy print-derived publishers.

Second, source-page decay. The linking page itself may lose ranking, lose traffic, or fall out of Google's index. A link from a page that no longer attracts qualified users contributes less than a link from a page that still ranks well. The page-traffic-weighted link metric introduced earlier captures this dynamically.

Third, source-domain decay. Whole domains can lose authority. A publication that loses its editorial team, becomes a thin-content site, or gets penalised will see its outbound links lose value across the board. Tracking the trajectory of source-domain organic traffic is a leading indicator of this kind of decay.

The bottom row is the most penalising one. The link-building tactic with the worst long-run economics is large-scale guest posting on networked content sites: the placements are easy to acquire and they decay rapidly because the host sites are themselves transient. The value-per-link, net of decay, is typically a small fraction of the apparent value at acquisition.

The operating implication is that link campaigns should be evaluated on a multi-year horizon, not on a 12-month horizon. A campaign that delivers 50 links with 90 percent two-year value retention is materially better than a campaign that delivers 200 links with 30 percent retention, even if the first costs more per link.

Putting It Together: A Composite Score Without Compression

The framework so far has refused to compress the six dimensions into a single number. That refusal is fundamental, not stylistic. The dimensions are independent enough that compression destroys information. But for triage, a composite is sometimes unavoidable: you cannot manually inspect every link in a 10,000-link profile.

The pragmatic compromise is a weighted-vector display rather than a scalar. For each link, produce the six-dimensional vector. For prioritisation, compute a context-specific weighted sum, but always retain access to the underlying vector. The decision tree below captures the logic for the most common operating decisions.

Decision path: What is the right action on this backlink?

Is the anchor text within the natural distribution for the linked URL?

• If yes: Is the source page topically relevant and does it attract real traffic?
  • If yes: Is the link in editorial content rather than footer, sidebar, or sponsored block?
    • If yes: Outcome: Retain and value highly. This is a healthy, high-value link. Track decay over time.
    • If no: Outcome: Retain but discount in value estimation. Non-editorial placement reduces weight.
  • If no: Outcome: Retain but discount. The link is not actively harmful but contributes little. Do not invest more in similar prospects.
• If no: Is the page targeted by the over-optimised anchor a high-priority ranking page?
  • If yes: Outcome: Investigate. If the anchor pattern is from a managed campaign, consider disavow and pace the future campaign more carefully.
  • If no: Outcome: Monitor but do not act yet. Single over-optimised anchors on low-priority URLs are usually noise.

The tree captures the operating principle: not every link needs a binary keep-or-disavow decision, but every link should have a vector and a documented rationale for how it fits the broader profile.

What Competitor Analysis Looks Like Without Single-Number Compression

The most common use of vendor-graph metrics is competitor analysis: a team looking at competitors' link profiles to understand the gap. Done with DR alone, this analysis produces misleading conclusions. Done with the six-dimensional framework, it produces a different and more actionable picture.

The right competitor analysis starts by mapping each competitor's profile across the six dimensions. The output is not "we are at DR 45 and they are at DR 62." The output is "their topical relevance is similar to ours, their editorial-quality share is materially higher, their anchor distribution is healthier, and their velocity is half ours." This produces a target list of specific dimensions to close, rather than a generic "build more links."

In partner data, the most common competitive gap turns out not to be link count but topical concentration. The leading competitor is typically over-represented in a small set of high-relevance publications; the trailing operator is under-represented in those publications and over-represented in broader, less-relevant ones. Closing the gap is not about acquiring more links; it is about acquiring the specific in-vertical links that the leader concentrates in.

When the Single-Number Metrics Are Still Useful

The single-number metrics are not useless. They are useful for two specific operations.

The first is triage at very large scale. When evaluating a list of 50,000 prospects from a scraping pipeline, sorting by DR is a faster initial filter than sorting by a six-dimensional vector. The decision is not "is this link worth pursuing?" but "is this link plausibly worth a closer look?" DR answers that triage question well enough for the top of the funnel.

The second is communication with stakeholders who are unfamiliar with the framework. A CFO who wants a quarterly KPI for the link-building program is not going to engage with a six-dimensional vector. A DR average or a DR-weighted referring-domain count is a defensible summary. The risk is that the summary becomes the metric the team optimises, which is the failure mode the rest of this essay was about. The right operating posture is to use DR for communication and the six-vector for decision-making, while making sure they do not diverge so far that the team's actions stop being justifiable to stakeholders.

The Academic Literature on Link Quality as a Graph Problem

The single-number aggregate metrics inherit their structure from the original PageRank algorithm (Page, Brin, Motwani, and Winograd, 1998, Stanford technical report). The original paper modelled the web as a directed graph, computed the stationary distribution of a random surfer's location, and used the resulting per-node score as a quality signal for ranking. PageRank in its original form treated all links as equivalent in weight, with the surfer choosing the next link uniformly at random from the current page's outbound links.

The limitations of uniform-weight PageRank were apparent almost immediately, and the academic literature has spent the subsequent quarter-century proposing successor models. The major lines of work are worth knowing because they map directly onto the dimensions in the framework above.

Hilltop (Bharat and Mihaila, 2001) proposed that link weight should be conditioned on whether the linking page is an "expert" page on the topic of the linked page. Expert pages were defined as pages with substantial outbound links to authoritative pages on the topic. The contribution of a link from an expert page on the topic was weighted more heavily than the contribution of a link from a generic page. Hilltop is the conceptual ancestor of the topical-relevance dimension in modern link-quality work, and Google has been reported to have integrated Hilltop-style logic into the broader ranking system since the 2003 to 2004 era.

Topic-Sensitive PageRank (Haveliwala, 2002, WWW) generalised the random-surfer model to compute multiple PageRank scores, one per topic vector. The query's topic determined which PageRank score was used at ranking time. The model anticipated the topical-relevance dimension at a more mathematically principled level, and although Google never confirmed exactly which form of topic conditioning it uses, the principle of topic-conditioned authority has been visible in observed ranking behaviour for years.

TrustRank (Gyongyi, Garcia-Molina, and Pedersen, 2004, VLDB) proposed using a seed set of trusted pages and propagating trust along the link graph to identify the trust score of all other pages. The idea was a counterweight to spam: trust flows away from authoritative starting points and decays through long chains, which means manipulated link networks downstream of seed pages accumulate less trust. TrustRank-style logic is the ancestor of the editorial-quality and source-page traffic dimensions in modern frameworks, and the SpamBrain detector almost certainly incorporates trust-propagation models.

Anti-TrustRank and BadRank (Wu and Davison, 2006) approached the problem from the other side: identify a seed set of known-bad pages and propagate distrust along outbound links. The combination of TrustRank and Anti-TrustRank produces a two-sided authority signal that is more robust to manipulation than either alone.

Reasonable Surfer (US Patent 7,716,225, granted 2010) is the Google patent that most directly informs the modern link-weighting architecture. The patent describes a model where the random surfer is not uniform across outbound links; instead, link weight is conditioned on the link's position on the page, the surrounding anchor text, the user attention model for the link's visual prominence, and the surfer's predicted click probability. The patent essentially formalises the editorial-quality and topical-relevance dimensions in a probabilistic framework.

The Link Spam Update arc (2012 Penguin, 2022 Link Spam Update, 2024 follow-ups) is the operational expression of the academic work. Google's spam-detection system has incorporated graph-relational features (anchor distribution across the network, velocity patterns, topical coherence of linking neighbourhoods, source-page quality signals) at increasing sophistication. The 2022 Link Spam Update marked the public emergence of SpamBrain's graph-relational detection at scale, and the subsequent updates have continued to tighten detection.

The pattern across the lineage is clear. The academic and patent literature has been pointing at multi-dimensional link quality for nearly the entire history of web search. The vendor-graph single-number metrics, which arrived in their current form between 2008 and 2014, are a market response to operator demand for simple numbers, not a description of how Google's link-quality system actually works. The framework in this essay attempts to align operator practice with the long-running academic and Google-internal understanding rather than with the simplified vendor abstractions.

Operationalising the Framework

The framework is implementable with current vendor tools plus modest internal engineering. The minimum viable pipeline is:

The first step is a backlink dump from Ahrefs, Semrush, or Majestic, ideally combined to cover crawl gaps. The data per link should include source URL, target URL, anchor text, first-seen date, last-seen date, source-page metrics (estimated traffic where available), and source-domain metrics.

The second step is enrichment. For each link, compute a topical-relevance score (cosine similarity of the source page text to the target page text, or a coarser topic-taxonomy match if compute is constrained). Classify the placement (editorial, sidebar, footer, sponsored, comment) by URL-pattern heuristics plus a manual sample for ground-truthing.

The third step is the per-target-URL anchor histogram. For each target URL with more than 20 inbound links, compute the anchor distribution and compare to the natural-distribution reference. Flag pages where the exact-match share exceeds the natural-distribution upper bound (typically 5 to 10 percent depending on vertical).

The fourth step is the velocity series per domain. Plot new referring domains per week or month over the trailing 24 months. Look for anomalous spikes, flat-rate plateaus, and recent step-changes.

The fifth step is the decay tracker. Re-check the link sample monthly (or quarterly for larger profiles) and flag removed links, links on pages that have lost source-page traffic, and links on domains that have lost overall traffic.

The whole pipeline is doable in a few weeks of engineering for a mid-sized link profile. The output is dashboards that surface the six dimensions independently and a per-link vector that supports both triage and detailed inspection. The single-number metrics remain available for communication, but they are no longer the decision input.

The teams that score links on six dimensions buy themselves the ability to disagree with Domain Rating in writing. The teams that only score on DR have no language for the disagreement, so they cannot make the case for the high-relevance, lower-DR link they should pursue or against the high-DR, low-relevance link they should not. The framework's contribution is grammar, not metric.

Key Takeaways

1. Single-number aggregates (DR, UR, DA) compress out the dimensions that determine link value. Ahrefs's own published correlation between DR and ranking position is approximately r = 0.09, characterised by Ahrefs as weak. The compression is structural, not a calibration problem to be solved.
2. Six dimensions are approximately orthogonal. Topical relevance, source-page traffic value, editorial quality of placement, anchor-text distribution, link velocity, and decay risk each capture independent information. Scoring on the vector preserves decision-relevant detail that the scalar destroys.
3. Topical relevance is the most underweighted dimension. In partner data, in-vertical links from moderate-DR publications consistently outperform unrelated-vertical links from high-DR publications, often by 2 to 4 times the ranking effect. The DR metric does not see this; the vector does.
4. Anchor distribution is the manipulation signal Google detects best. Natural profiles dominate with branded anchors (40-70 percent) and use exact-match anchors sparingly (1-5 percent). Manipulated profiles invert this. SpamBrain's anchor-distribution detector targets the inversion, and over-optimised anchors trigger algorithmic and manual penalties more reliably than any other dimension.
5. Decay should be priced into campaign economics. Net value retention over a 24-month horizon varies from approximately 85-92 percent for legacy publishers to 15-35 percent for guest-post and syndicated-content placements. The campaign with the best apparent unit economics often has the worst net-of-decay economics.
6. Use single-number metrics for triage and communication; use the vector for decisions. DR is acceptable for filtering 50,000 prospects to 5,000, and for explaining the program to a CFO. It is not acceptable as the primary decision metric. The vector is the decision input; the scalar is the summary.