Sparkitecture: give a roofing website a useful job

Research and positioning brief · source checks October 3–4, 2026 (America/Vancouver).

The argument we can defend

Information discovery can finish before a website visit. Independent research finds lower click rates in specific AI-summary contexts and declines for some informational publishers. That does not establish that every website loses traffic, that SEO causes the decline, or that roofing businesses have the same results.

Our design proposition: give the visit a job. Help a homeowner investigate their situation, understand a tradeoff, make a choice, and leave prepared for the next conversation. A useful experience is a testable response to changing discovery habits; it is not a guaranteed traffic or conversion remedy.

Sparkitecture combines four pillars: Spark (mystery, surprise and delight), Value (a useful outcome), Education (learning by investigating) and Realm (a coherent world). Attention → Interest → Decision → Action describes the authored visitor journey through those pillars. Attention opens an answerable question and plants a clue. Interest lets the visitor investigate. Decision reveals a constraint or tradeoff that changes a choice. Action produces a useful, portable result.

What Zombie UI means here

Katie Dill's linked Lenny's Podcast talk, published September 25, 2026, criticizes generic zombie UI and advocates clear point of view, user understanding and care in details. Its relevance is design judgment: fast generation still needs intent, editing and standards. The talk is a viewpoint, not a trial of Sparkitecture.

The original page metadata and description were verified. Native transcript loading returned no segments. A secondary transcript informed the detailed research notes; exact words or timestamps should not be attributed as independently verified original transcript. Neither Dill nor Lenny's Podcast is represented as endorsing RoofClaw.

Our operating definition of Zombie UI is an interface that looks complete but gives the visitor no meaningful way to understand, decide or act. A lifeless interface can be animated. A useful experience can be quiet and simple. The test is what changes for the visitor.

Ten sources, their methods and their limits

Pew Research Center: 8% vs. 15%

Google users are less likely to click on links when an AI summary appears in the results · 2025-07-22.

In Pew's March 2025 U.S. browsing panel, people clicked traditional results on 8% of Google visits with an AI summary, versus 15% without. Summary-source links were clicked on 1% of summary visits.

Method: 900 U.S. adults; 68,879 unique Google searches during March 2025. Search results were recreated April 7–17 to classify AI summaries and links.

Limits: Observational association, one month, Google only. Different queries and recreated results can affect the comparison. This is not a measured traffic loss for roofing websites.

Ahrefs: 58% lower CTR

Update: AI Overviews Reduce Clicks by 58% · 2026-02-04.

Ahrefs found AI Overview presence correlated with a 58% lower position-one desktop clickthrough rate relative to its modeled counterfactual.

Method: 300,000 keywords, split between 150,000 with AI Overviews and 150,000 informational queries without. Aggregated Search Console desktop CTR compared December 2023 with December 2025.

Limits: Selected keyword sample, desktop only, observational comparison. The 58% is relative to a modeled counterfactual, not 58 percentage points or all website traffic. No roofing conversion outcome was studied.

SparkToro / Similarweb: About 68% zero-click

In 2026, Less than One Third of Google Searches Still Send a Click · 2026-06-08.

SparkToro's Similarweb-based estimate found 68.01% of U.S. Google browser searches ended without a click in January–April 2026.

Method: Desktop/mobile web clickstream panel; combined estimates assume two-thirds mobile and one-third desktop. Mobile sessions end after ten seconds of inactivity; mobile paid-click share is modeled.

Limits: Excludes Google's mobile search app. Historical comparisons use different data providers and panels. Zero-click can mean a useful answer, reformulation or abandonment; it does not mean a lost customer.

Cloudflare: 9% fewer March referrals

The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals · 2025-08-29.

Google referrals to Cloudflare's news-customer cohort fell approximately 9% in March 2025 compared with January; April was approximately 15% lower.

Method: Cloudflare network observations of news-related customers across the Americas, Europe and Asia. Fixed customer cohorts were used for crawler-growth comparisons.

Limits: News publishers are not all websites or roofing businesses. AI launches coincided with the changes but this does not isolate causality. Crawl requests and human referred visits measure different things.

Wikimedia Foundation: Roughly 8% fewer human pageviews

New user trends on Wikipedia · 2025-10-17.

Wikipedia reported human pageviews down roughly 8% year over year in recent months of 2025, after updating its bot detection. It linked the trend to changes involving AI answers and social media.

Method: First-party pageview measurements across Wikipedia language versions; revised bot classification for March–August 2025.

Limits: Classification rules changed, so Wikimedia advises care in interpretation. The explanation is an attribution, not a controlled causal estimate. This is a major informational publisher, not a roofing-company traffic study.

Google: Aggregate clicks reported stable

AI in Search is driving more queries and higher quality clicks · 2025-08-06.

Google said total organic clicks were relatively stable year over year and click quality had increased, while traffic shifted between sites. It defined quality as users not quickly clicking back.

Method: Google statement based on internal platform data; no underlying reproducible dataset or detailed method supplied in the post.

Limits: An interested platform's aggregate statement. Per-query CTR can fall while total clicks stay stable as query volume grows. A quality-click proxy does not establish leads, sales or learning.

Google Search Central: No special AI markup required

AI features and your website · Live documentation; checked 2026-10-04 UTC.

Google recommends existing SEO fundamentals for AI features: crawl access, helpful native text, good page experience and structured data matching visible content. Supporting links need indexing and snippet eligibility.

Method: Official implementation guidance for Google AI Overviews and AI Mode; not an experiment measuring design performance.

Limits: Applies to Google's features. Eligibility does not guarantee crawling, indexing, ranking, AI citation or traffic. Google requires no special AI file or schema.

Nielsen Norman Group: Explore, then decide

Calculators and Quizzes: User Expectations · 2024-03-22.

Usability research describes calculators and quizzes as exploratory decision aids. Inputs should meaningfully change outputs, and forcing contact details before revealing results can frustrate users.

Method: Qualitative user-research observations summarized with examples. The public article does not provide a full sample frame or a quantitative conversion estimate.

Limits: Supports useful interaction design, not a promise that every quiz increases retention or sales. Experienced users can reject generic outputs; decision support is one input to judgment.

Nielsen Norman Group: Make motion explain a change

The Role of Animation and Motion in UX · 2020-01-12.

Page Laubheimer describes restrained animation as useful feedback for recognized actions, state transitions, relationships and available interactions. Motion unrelated to the visitor's task can distract or overwhelm.

Method: Expert usability guidance with interface examples and references to attention research; not a controlled landing-page conversion experiment.

Limits: Supports purposeful motion, not a claim that animation improves engagement or revenue. Animation supplements clear navigation and should be brief and restrained. A useful reduced-motion route remains necessary.

Shiri Melumad and Jin Ho Yun / PNAS Nexus: Seven experiments; active synthesis mattered

Experimental evidence of the effects of large language models versus web search on depth of learning · 2025-10-28.

In learning-and-advice tasks, participants using LLM summaries reported shallower learning and wrote advice that was shorter, less original and received lower recipient willingness-to-adopt ratings than advice from participants exploring web links. Some conditions held the information constant or compared tools on the same platform.

Method: Seven randomized online/laboratory experiments, including four in the main text and three conceptual replications. Topics included gardening, healthier living and financial scams; outcomes included self-report, advice-text analysis and recipient ratings.

Limits: These tasks do not measure commercial conversion or Sparkitecture. Several learning outcomes were subjective or inferred from advice. The abstract reports 10,462 participants while the introduction gives 10,426; avoid repeating the unresolved total. Results may differ for factual lookup, advanced domains or other learning measures.

How apparently conflicting findings fit together

A click rate is a proportion of searches; aggregate clicks are a total. If search volume grows, lower per-query click rates can coexist with stable total clicks. Traffic can also move between sites. Pew, Ahrefs, Cloudflare, Wikipedia and Google measure different populations and outcomes. They should not be added into one percentage or presented as roofing-specific forecasts.

Zero-click is broader than AI: direct answers, reformulation, abandonment and other behaviors can produce it. A publisher losing pageviews does not prove lost buyers. A local roofing query with high purchase intent is not identical to a general informational query. The practical next step is to inspect the roofing company’s own search impressions, clicks, referrals and qualified conversations.

CCdocs: a memorable story earned editorial attention

The DeepSeek-versus-Opus page makes a technical comparison legible through eight boxing rounds, illustrated characters, native explanations and math, optional AI clips, and shareable material. Its cartoon hype character is explicitly fictional. The point to borrow is the structure: tension, evidence, an economic constraint and a concrete decision.

Tom's Hardware covered the Call Center Doctors comparison on October 1, 2026 and linked directly to the supplied page. The creator reports that the coverage arrived within two days. A public page-launch timestamp was not found, so the interval remains creator-reported.

This is a specific example of earned editorial attention and a page designed for participation. It does not establish that design caused the coverage, how many readers arrived, how long they stayed, which controls they used, or whether they converted. Those are separate analytics questions. The model comparison is one organization’s workload and configuration, not a general ranking. Optional video playback was not verified during source research.

A roofing page that demonstrates its own mechanism

Start with “What should I prepare before a roofing inspection?” A static answer lists general advice. The experience asks whether the visitor has a leak, a storm concern or a planned replacement. It changes the observations and questions they should gather and gives them a preparation brief. The same mechanism can help an owner compare page scopes or explain a complex service.

The visitor’s choice must change the output. Ask only for inputs that affect the result. Keep preparation separate from diagnosis, price estimates, warranty promises and insurance decisions. Provide the result before asking for contact details. Use a clear existing booking destination when the visitor is ready.

Vito’s purpose is to explain what changed and why it matters. On this demonstration page, his guidance is authored and scripted, not a live model response. His character, optional movement and visuals never carry the only copy of the meaning.

SEO and AEO implementation

Use a specific title, description, one H1, a plain-language definition near the top, answer-led headings, descriptive internal links, and visible FAQ answers. Key evidence and meaning remain native text. Structured data describes the visible page; it does not claim ratings, customer outcomes or a live autonomous agent. The public source report is readable without JavaScript.

Google’s AI-features guidance requires no special AI schema or file. Normal search fundamentals remain relevant: crawl access, indexed and snippet-eligible content, useful text, a good page experience and accurate structured data. Eligibility does not guarantee indexing, ranking, answer-engine citation or traffic. The optional llms.txt is a convenience document, not a ranking mechanism.

The canonical RoofClaw route is /2027-website-design, with /sparkitecture as an alias. The source report, definitions and FAQ answers remain readable native text.

What to measure after a real launch

Compare an equivalent static route with the useful interactive route. Define the visitor task and success criterion before collecting outcomes. Measure actual interaction exposure, completed preparation tasks, blueprint exports, verified completed bookings and qualified attended conversations. Keep source/search impressions, AI citations, referred visits and business outcomes separate.

Longer time on a page can mean confusion. A CTA click is not a booking. A download is not later use. A build, a source link or a press mention does not prove learning, engagement, sales or revenue. No audience outcome has been measured for this demonstration.

Next move

Book the existing RoofClaw demo and bring one homeowner question, the result you want visitors to leave with, and your saved blueprint. Explore a RoofClaw configured with the Sparkitecture harness and agree the delivery scope. No price or delivery-time promise is made here.