What are the core principles of growth-driven design, and what evidence supports them?

"Every plan starts with a focus metric that you want to improve," and the highest-impact ideas for that metric go into a time-boxed sprint [4]. That sentence plus "Optimizations are not blind guesses" is the entire published decision rule [4]. What counts as proof, how long measurement runs, and what to do with an inconclusive result are all left to whoever is running the sprint, so the strength of the principle depends on evidence from outside the method.

Jakob Nielsen measured four iterative redesigns and reported a median improvement in overall usability of 165% from the first version to the last, about 38% from one version to the next [6]. Two parts of that finding usually get dropped when it is quoted. Several projects had at least one usability attribute score lower after a revision, and in the cash register case the third version tested worse than the second, which is why Nielsen sets three versions as a minimum, not a target [6]. The systems were software interfaces measured on fixed task sets in the early 1990s, and his 2006 note on a later set of web redesigns puts average measured improvement at 135% [6]. The result supports iterating on a design; it does not forecast a gain for any page of yours.

Kohavi, Deng, Longbotham and Xu, writing from thousands of controlled experiments at Bing, LinkedIn, Amazon and Microsoft properties, put the success rate of ideas at Bing at roughly 10 to 20%, with successful changes moving key metrics 0.1% to 1.0% once diluted to overall impact [7]. Their third rule of thumb aims directly at the case studies agencies publish about this method: quality varies, many reported tests were never peer reviewed or properly run, and a result that held on one site often fails on another [7]. Their sixth rule is that simple experiment designs beat complex ones, because a complex design hides bugs and makes attribution harder, which is the same reason the method caps a sprint at one focus metric [7].

A test that clears significance still does not tell you why the number moved. NN/g's position is that A/B testing gives data only on the element being tested and nothing about the reasoning behind the result, so it should not be the first method used on a design problem [10]. That gap costs the most on problems the focus-metric loop never surfaces: qualitative sessions regularly expose visitors who do not believe a claim on the page, and NN/g puts effect sizes for trust and information problems at 100% or more against the 1 to 2% a button test moves [10].

The fourth step of the sprint cycle sends what the sprint learned to marketing, sales and service, which the method describes as cross-department collaboration that helps the growth team adjust for performance [4]. Luke Summerfield, who created growth-driven design, states the same point on that page: a team should not get stuck optimizing existing content, and should expand into helping other departments hit their goals with the website as a tool [4]. No measurement of what transfer produces appears anywhere on the site, and it is the only step in the cycle that ships no artifact, so it is the easiest one to quietly drop.

The case for protecting it rests on the type of finding it moves. NN/g's qualitative work identifies visitors who will not do business with a company because the site undermines its credibility as a recurring and large effect [10]. A finding like that is not a page fix. It changes which proof a sales team leads with on a call and which objection marketing answers in an email, which is value the sprint's own focus metric will never record.

The 80/20 rule is one of them, because the section heading "Run an 80/20 Wishlist Analysis" and the instruction to distill the launch pad down to the 20 percent of the wish list that will have the most impact come from IMPACT's guide, written by an agency that sells website builds [9]. growthdrivendesign.com states the underlying idea without a ratio: sprints focus on "the 3-5 pages that can drive the greatest impact," and the remaining pages get "a 'fresh coat of paint'" during migration [3].

Two different claims travel under the 80/20 label, and merging them produces the invalid reasoning that shows up in most summaries of this method. One is a filter applied to a wish list of page and feature ideas, a prioritization convention with no measurement behind it [9]. The other is an attribution claim about where a website's results come from, messaging and buyer research against custom visual design, which is Lean Labs' estimate from its own projects and belongs in the contributor section below [11]. A filter tells you what to build first. An attribution claim would need evidence that the split holds, and none is published by anyone.

"It should go up quickly, and it will definitely not be perfect" is also IMPACT's description of the launch pad, aimed at avoiding analysis paralysis [9]. The method's own launch pad page carries a counterweight that the popular version leaves out, under the heading Quality Over Speed: "We lean towards launching quickly, however, speed should not compromise your ability to provide user value and the happiness of the team/stakeholders" [3]. The wish list brainstorm of "50 to 150 (or more) new ideas" and the position that buyer personas are "a non-negotiable step" are IMPACT's as well [9]. Persona and customer research are part of the method's strategy stage [1][2], but the stop-reading-now framing belongs to the agency.

The outcome numbers attached to the method are survey answers, not results. The methodology page labels its own figures "Based on the '2017 state of GDD' survey responses" and describes them as what "Agencies that used GDD on HubSpot Websites (vs. Wordpress) reported": 14.34% more leads after six months and 12.56% higher revenue, alongside 60 days to launch against 108 and client happiness scored 7.7 against 6.3 [1]. Agencies were reporting on their own projects, on one CMS against another, in a survey that is now nine years old, with no control over which clients chose which model. Kohavi's third rule covers the general case, that published test results vary in quality and often fail to transfer [7]. Those figures record what practitioners said about their own work.