Agency methodology for small-business website projects

Overview

Agency methodology for small-business website projects is the practitioner-level workflow by which a research-first agency translates foundation research into a built, measured, maintained website. It is a companion to Knowledge-base-backed website methodology, which establishes the principles — atomic notes, sourced claims, confidence labels, and the five-stage pipeline from capture to maintenance. This page documents how those principles operate inside an engagement: how discovery is run, how foundation research is bounded, how structured content gets designed, how the build sequence resists scope creep, and how post-launch measurement is decoupled from the legacy single-metric model. The decision rules and patterns described are documented practitioner standards for SMB (small and medium-business) work, not universal claims about every web project.

The methodology assumes three things about the operating environment in 2026. First, the cost floor for "real" web functionality has dropped far enough that database-backed tools, customer portals, and live-data features are within SMB reach. Second, Google is no longer the only success surface — AI citation, branded search, and direct traffic now matter independently. Third, the relationship structure between agency and client determines whether the compounding upside of a research-first build ever materializes. Each of those assumptions has consequences for how the workflow is sequenced.

Stage A — Discovery

Discovery is the first paid stage of an engagement. Its purpose is not to write a creative brief but to establish whether the project is buildable, who will use it, and what the realistic payback envelope looks like. A documented practitioner standard is that scoping begins with the question how does this business acquire customers? rather than what does this business want the site to do? — because the acquisition mode determines which of the four working-surface capabilities (structured data, interactive functionality, live data, account state) actually serves the business, and most businesses need at least one but not all four.

Source: Atomic rule "Most businesses benefit from AT LEAST ONE of the four working-surface capabilities, NOT all four; match the capability to acquisition / service mode," citing Zippia 2026 (27% of US SMBs run with no website at all), B2B Lead Finder 2026 (trades 45–56% no website), and the four-capability taxonomy. Confidence: Industry-consensus. Caveat: A sole operator booked out on word-of-mouth may legitimately need none of the four capabilities; the working surface should not be pushed when it does not serve the acquisition mode.

Scoping the four capabilities

The four-capability taxonomy is the discovery-stage filter. Each capability is independent; bundling them as a checklist is the most common scoping failure. A documented practitioner standard is to name, in writing, the capability the build serves and to state explicitly that the other three are not in scope.

• Structured queryable data — searchable, filterable inventory or records (parts catalogue, project gallery, document library).
• Interactive functionality — calculators, configurators, request flows that perform a useful operation.
• Live, frequently-updated data — pricing, availability, dashboards driven by current information.
• Account state — customer logins, saved jobs, invoices, project status.

The decision rule for whether to build a searchable catalogue is documented as a three-part check: catalogue size (would the visitor otherwise scan past a few hundred items?), change frequency (does the data update often enough that hand-edited documents become error-prone?), and attribute richness (do items differ along multiple independent dimensions people genuinely filter on?). If two or three of those are true, a structured catalogue is justified.

Source: Atomic rule "Build a searchable, structured catalogue when records are numerous, change often, or carry several independent queryable attributes," citing Baymard list-usability data (67–90% vs 17–33% abandonment) and the NN/g faceted-search refinement definition. Confidence: Industry-consensus on the magnitude evidence. Caveat: No single source quantifies the three-leg interaction jointly; the legs operate as a checklist, not a validated cutoff.

The corresponding don't-build rule is equally documented. A static list is sufficient when inventory is small (commonly cited as under ~200 items), stable, and shallow. Three independent search-software vendors — Hawksearch, Prefixbox, and Luigi's Box — converge on the ~200-item threshold against their own commercial interest, which strengthens the threshold as a planning heuristic.

Source: Atomic rule "Skip faceted search when the inventory is small (~under 200 items), stable, and shallow; a static list is fine," citing Hawksearch, Prefixbox, Luigi's Box, and NN/g on faceted UI interaction cost. Confidence: Industry-consensus. Caveat: The 200-item number is a rule of thumb, not a validated cutoff; the skip recommendation is most defensible when all three conditions (small, stable, shallow) are true.

Payback framing

Discovery is also where realistic financial expectations are set. The documented practitioner standard for SMB website payback is to quote a range — 6 to 24 months — with non-trivial probability of never paying back. Single-number payback estimates do not survive contact with the compounding unknowns of vertical, competition, content, authority, customer lifetime value, churn, attribution decay, and AI disruption.

Source: Atomic rule "Quote payback as a 6-24-month range with non-trivial failure probability — never as a point estimate." Confidence: Verified as a documented practitioner standard. Caveat: The range is a framing discipline, not a forecast; precise payback for a specific business is not forecastable.

Capability versus outcome

A discovery-stage honesty discipline distinguishes capability from outcome. The agency can build the capability — a flag, a dashboard, a structured catalogue. Whether the client captures the resulting edge depends on whether the client changes behaviour and has authority to act on the information. The BDC / MIT digital-transformation model is explicit on this: technology investment drives revenue, but transformation-management capabilities — clear strategy, training, continuous-improvement culture — drive profit.

Source: BDC, citing the MIT (Westerman / Bonnet / McAfee) digital-transformation model, accessed 2026-06-21. Confidence: Verified; consensus across the digital-transformation literature. Caveat: "Transformation management capabilities" is a fuzzy construct that resists clean measurement; BDC's framing names it without quantifying it.

Anchoring the discovery conversation in the Canadian SME baseline is also documented. BDC research finds that only about one in five Canadian businesses has reached a high level of digital maturity, while more than half remain low — so most peers of any given prospect are not capturing the information-asymmetry edge either.

Source: BDC digital-maturity research (bdc.ca), accessed 2026-06-21. Confidence: Verified; primary research from a federal Crown corporation. Caveat: Cross-sectional, not causal; mild mandate-aligned incentive flag because BDC has a built-in interest in encouraging SME digital adoption.

Stage B — Foundation research

Foundation research is the second stage and the one most often skipped. It is the atomic, sourced, dated, confidence-labeled note-taking that powers everything downstream — see Knowledge-base-backed website methodology for the principles and Editorial discipline and sourcing for the citation discipline that governs how claims are made.

The foundation-research stage produces internal notes, not draft marketing copy. Each note carries a Source, a Confidence label drawn from a documented taxonomy (Verified / Industry-consensus / Single-source / Estimated / Author's view / Contested / Stale), and a Caveat where the claim's primary source has limits. The practitioner pattern is that every objective claim that later appears on a public-facing surface should be one click away from a primary source — and the only way that becomes operationally affordable is if the sourcing is done at the research stage, not retrofitted after marketing copy is written.

Source: Research brief "Confidence Levels, Sources, and Dated Claims," citing Edelman 2025 (7 in 10 globally believe leaders deliberately mislead), Profound on ChatGPT pulling 7.8% of citations from Wikipedia, Perplexity API design (title + URL + date + snippet for every source), and the FTC's reasonable-basis substantiation doctrine. Confidence: Industry-consensus. Caveat: The AI-citation premium for sourced content is plausible but not yet rigorously proven at RCT level; the supporting studies are largely vendor-produced.

Bounding the research stage

The risk at this stage is unbounded scope. A documented practitioner discipline is to bound foundation research by the actual decisions it informs. The atomic-note pattern (one idea per note) is the operational unit; clustering into briefs happens at synthesis, not during note-taking. Foundation research is done when the synthesis outline can be assembled without further research, not when the research feels complete.

Why this stage justifies its cost

Foundation research is also the stage that protects against two documented failure modes. The first is content that does not get cited by AI systems because it is generic and templated. The peer-reviewed lift in generative-engine answers comes from edits to visible page text — body-text citations, quotations, and statistics — not from schema markup. The second is content that does not earn links because it is not worth linking to: Moz's Beginner's Guide to Link Building is direct that "all link building campaigns must start with something worth linking to. It's very difficult to build links to low-value webpages."

Source: Aggarwal et al. (KDD 2024), "GEO: Generative Engine Optimization" — over 40% visibility lift in generative-engine answers from edits to visible page text, not schema. Confidence: Verified; peer-reviewed. Caveat: Methodology is explicit that schema is not the lever; there is no independent or primary evidence for the corresponding schema claim.

Source: Moz, Beginner's Guide to Link Building (moz.com). Confidence: Industry-consensus. Caveat: Moz sells SEO software (Link Explorer); the guide is itself a linkable asset and lead magnet.

Stage C — Structured-content design

Once foundation research is in hand, content design begins. Two documented patterns govern this stage: information architecture for multi-vertical service businesses, and editorial structure for AI-citation and E-E-A-T (Experience, Expertise, Authoritativeness, Trust) signals.

Information architecture

The dominant URL pattern for multi-vertical B2B service firms is /industries/<slug>/ — observed at Crowe, BDO USA, HUB International, Dentons (as /industry-sectors/), and Grainger. The /for-farmers/ audience-based pattern was observed on zero of eight sampled multi-vertical service sites; that pattern belongs to SaaS marketing, not service-business information architecture.

NN/g's primary finding on audience-based navigation is a warning, not an endorsement — it "will often degrade usability" unless categories are mutually exclusive, jargon-free, with substantially unique content per section. Small service businesses rarely meet that bar. The dominant working model for multi-vertical firms is an Industries × Services matrix with both axes in the top navigation. For small service firms (four to ten people) the scalable pattern is hub-and-spoke — one hub page per vertical, shared services pulled in via reusable blocks — rather than a full matrix.

Source: Research brief "Information architecture for service businesses with multiple verticals," citing observed URL conventions at Crowe, BDO USA, HUB International, Dentons, Grainger, RelaDyne, E.H. Wolf; NN/g audience-based-nav guidance; Google's John Mueller on near-duplicate consolidation. Confidence: Industry-consensus on the URL convention; the hub-and-spoke vs matrix distinction is a documented practitioner pattern, not a controlled study. Caveat: NN/g and Baymard have no dedicated article on the "industries we serve" pattern specifically; the closest primary research is the audience-based-nav warning. Mobile navigation patterns for multi-vertical sites are underexplored in current research.

The same brief documents that multi-step forms outperform single-step only above approximately seven fields; below that the lift evaporates (Zuko). The widely circulated "300% lift" multi-step claim is misattributed to ConversionXL — the correct primary attribution is Venture Harbour (Marcus Taylor) portfolio testing.

Editorial structure for AI citation

A documented practitioner standard for top service and FAQ pages is to answer the customer's actual question directly in the first 100–150 words, with specific local detail — pricing ranges for the area, named staff with credentials, real project specifics — and without marketing preamble. This serves two regimes at once: Helpful Content / E-E-A-T rewards demonstrable first-hand experience, and AI Overviews preferentially surface concrete, extractable answers.

Source: Atomic rule "Restructure top service and FAQ pages to answer the customer's question directly in the first 100–150 words, with specific local detail," citing the August 2022 Helpful Content Update (which added the extra "E" for Experience in December 2022) and Pew / Ahrefs data on AI Overview click loss. Confidence: Industry-consensus. Caveat: AI Overview click-through rates remain disputed between vendor studies; the lead-with-the-answer pattern is robust regardless of which click-rate figure is closest to correct.

Service pages must also show evidence of first-hand experience — author bylines, real credentials, real project detail — because generic templated AI-spun content has become a ranking and citation liability rather than neutral.

Source: Atomic rule "Demonstrate first-hand experience explicitly on service pages — named staff, real credentials, real projects — not slogans," citing the Helpful Content Update and Google's stated March 2024 goal of cutting low-quality unoriginal content by 45%. Confidence: Industry-consensus.

A further documented editorial discipline is that body-text citations, quotations, and statistics — not schema markup — are the AI-visibility lever. The peer-reviewed evidence is on the body-text side; investing additional editorial effort in schema beyond baseline JSON-LD for entity disambiguation is not supported by the primary literature.

Source: Aggarwal et al. (KDD 2024). Confidence: Verified. Caveat: Schema remains valuable for entity disambiguation and rich-result eligibility; the rule is that it is not the AI-citation lever, not that it should be skipped.

Stage D — Build

The build stage executes against the structured content design. Documented practitioner patterns at this stage govern stack selection, build-versus-rent decisions, scope discipline, and the longevity properties that determine whether the resulting site lasts six years or two.

Stack selection by use case

A documented six-tier framework matches stack to use case and budget. The framework operates as a default; deviations are noted in writing. See Default web stack (Astro on Cloudflare) for the canonical Astro-on-Cloudflare reference build.

• Brochure / service / contractor site, $3–8k, single editor — WordPress + Gutenberg + Kadence theme. Native, free, fast, accessible, no recurring license risk.
• Marketing site for established SMB, $5–15k, frequent edits — WordPress + Gutenberg + custom block theme + ACF Blocks for bespoke modules.
• Content-heavy, SEO-critical, AI-citation-sensitive, $8–20k — Headless WordPress (WPGraphQL) + Astro on Cloudflare Pages. Server-rendered HTML, zero JavaScript by default.
• Web app, authenticated flows, dashboard, $15k+ — Next.js + headless CMS (WordPress, Sanity, or Payload) + Vercel or Cloudflare.
• Tiny brochure, $0–1.5k, no developer access — Squarespace, or WordPress + Gutenberg + the default Twenty Twenty-Five theme.
• Existing legacy Elementor / Divi the agency inherits — do not rebuild reflexively. Audit each; optimize in place if traffic is significant and the site performs.

Source: Reference entry "Alternative-stack recommendations by use case and budget." Confidence: Documented practitioner framework. Caveat: Page builders are not uniformly bad in 2026 — Divi 5's block format and Elementor's Gutenberg interop suggest the category is fixing some of its lock-in problems. The strongest "shortcode lock-in" argument applies to legacy installs.

The case against page-builder defaults

Page builders dominate WordPress and degrade it. Elementor alone runs on approximately 18.6% of all WordPress sites (W3Techs April 2026); combined with Divi, WPBakery, Beaver Builder, and Bricks, page builders touch a clear majority of WP installs. The architectural costs — DOM bloat, asset weight, lock-in, accessibility gaps, AI-extraction friction — compound across every site they build. Elementor's own engineering blog notes that Flexbox Containers ship "40% less HTML"; Bricks Builder produces "40–60% less code output than Elementor for equivalent layouts."

The "non-developer can edit it" promise is largely fictional in practice. Agencies use page builders because they speed up the agency's build, justify monthly retainers, and let junior staff ship — not because clients actually edit. Most maintenance hours go to plugin conflicts, breaking updates (Elementor 3.24, 3.26), and security incidents (Bricks CVE-2024-25600, exploited approximately 24 hours after the patch shipped).

Source: Research brief "The Case Against Page Builders," citing W3Techs April 2026 (Elementor 18.6%), Elementor engineering blog (Flexbox Containers ship 40% less HTML), Bricks Builder documentation, and the Bricks CVE-2024-25600 timeline. Confidence: Industry-consensus on architectural costs. Caveat: Page builders were the right answer for 2018–2022 and genuinely democratized the web. The case applies to 2026 defaults for new builds, not retroactive criticism of every existing site. Per-builder Core Web Vitals breakdowns are mostly agency-published; Web Almanac 2024 did not republish per-builder splits.

Owning the stack

A documented practitioner standard is that "ownership" in 2026 is a layered claim — domain, DNS, code, content, data, infrastructure, analytics, and customer relationships each live in different places, and most small businesses unknowingly cede control of three or four of those layers to their agency or platform. The defensible position is not "build everything custom" but a small, portable stack: registrar account held in the business's own name as the Registrant, Git repository under the client's organization, open content formats, exportable database, and a posture that treats any platform incapable of producing a clean export as a hostage situation in slow motion.

The lock-in costs are documented. Approximately nine of ten web migrations fail to improve SEO (single-source — flagged), 50% organic traffic loss is common, and the average post-migration recovery is 523 days across a study of 892 domain migrations. The WP Engine / Automattic dispute (September 2024 → ongoing), Shopify's forced sunset of checkout.liquid (August 2024 / August 2025 / June 2026), and Squarespace's $180M absorption of Google Domains (September 2023) are three case studies that prove platform "ownership" is conditional.

Source: Research brief "Owning your stack," citing Search Engine Journal's 892-migration study (Jan 2025, 523-day recovery), the WP Engine / Automattic case, Shopify's checkout.liquid sunset notices, and the Squarespace–Google Domains acquisition. Confidence: Verified on the case studies; Industry-consensus on the 523-day figure. Caveat: The "9 of 10 migrations fail" framing is single-sourced (Numen Technology blog citing unnamed analysis). Public writing should lead with the 523-day stat instead. Domain-hostage cases are pattern-documented but rarely litigated.

The platform-decay framing — Cory Doctorow's "enshittification" — is useful as a diagnostic but not as a law. Some platforms (Stripe, Cloudflare, Postgres, Linux) have not enshittified despite being enormous; the pattern applies most cleanly to vertically-integrated, two-sided, advertising-funded platforms with switching costs.

Source: Doctorow, Pluralistic (January 21, 2023); American Dialect Society 2023 Word of the Year; Macquarie Dictionary 2024 Word of the Year. Confidence: Verified. Caveat: The pattern is rhetorical, not empirical.

Rent the commodity, build only what is differentiated

The cost floor for "real" web functionality has dropped substantially. Storage moved from $0.15/GB (S3 launch, 2006) to approximately $0.02/GB (2025); entry compute from $0.10/instance-hour (EC2 launch, 2006) to approximately $0.005/hour; internet transit from $1,200/Mbps (1998) to approximately $5/Mbps (2010). The decade 2004–2014 is when the spine got cheap, on dated pre-AI timelines. AI-assisted coding (2022–present) is a recent accelerant at the margin and should not anchor the falling-cost argument.

The ceiling, however, did not fall. Custom client portals run $20,000–$50,000 to build with $10,000–$25,000/year maintenance, and approximately 80% of software features are rarely or never used. The commodity floor is cheap; the bespoke labor to assemble, customize, integrate, and maintain a genuinely custom tool remains the dominant and still-substantial cost.

Source: Research brief "The falling cost floor of 'real' web functionality for SMBs" (June 2026), citing S3 launch pricing, EC2 launch pricing, the bandwidth-transit drop documented at 1,200 → 5 $/Mbps (1998–2010), and SPP 2025 on custom-portal cost ranges. Confidence: Verified on the infrastructure cost trajectory. Caveat: Build-vs-buy SaaS vendor cost pages (SPP, Agency Handy, Space-O, Kavara) are used for ceiling figures but flagged for vendor incentive.

The documented practitioner standard that follows from the falling-floor / stable-ceiling pattern is: rent the commodity parts (Stripe, Auth0, Algolia, RDS, Lambda) and build only what is genuinely differentiated — the client's actual business logic. Each commodity service replaces a person-month-or-more bespoke build, and building any of them in-house pays twice (once at build, again at maintenance) for a worse outcome.

Source: Atomic rule "Rent the commodity parts; build only what is genuinely differentiated logic," citing Stripe (2011), Auth0 (2013), Algolia (2012), Elasticsearch (2010), RDS (2009), and Lambda (2014). Confidence: Industry-consensus. Caveat: "We want our own auth" is almost always a red flag and should surface the pre-Auth0 password-hashing entry and the security implications.

The corresponding scope-discipline rule is that approximately 80% of software-product features are rarely or never used (Pendo 2019; corroborated at 64% by Standish Group). Every unused feature is build cost plus maintenance cost plus bug surface area plus cognitive load. The default answer to "could we also add…" is "not in v1."

Source: Atomic rule "Build features customers will actually use; ~80% of software features go unused," citing Pendo 2019 and Standish Group. Confidence: Industry-consensus.

Page speed as a moat

Core Web Vitals (CWV) are necessary but not sufficient in 2026. Google's official "Good" thresholds remain LCP ≤ 2.5s, INP ≤ 200ms, CLS ≤ 0.1. The widely-repeated SEO-blog claim that March 2026 tightened LCP to 2.0s and "elevated" INP is unconfirmed by Google and contradicted by current documentation. More than half of the mobile web still fails CWV: Web Almanac 2025 reports 48% mobile / 56% desktop origins pass all three on July 2025 CrUX data.

The platform spread is the moat. Duda 83.63%, Shopify 75.22%, Squarespace 67.66%, WordPress 43.44%, with Elementor sub-segments performing worse (25–35%). Agencies who deliver durable CWV on WordPress — or skip it for Astro / Next.js / Hugo — are doing structural work that freelancers using stock themes cannot replicate.

Source: Research brief "Page Speed as a Moat" (piece 9 of 15), citing developers.google.com (Dec 10, 2025) on Good thresholds, Web Almanac 2025 on July 2025 CrUX pass rates, and CWV Technology Report platform breakdowns. Confidence: Verified on thresholds and pass rates. Caveat: CWV is Chrome-only field data; Safari only added LCP + INP in Safari 26.2 (December 2025). Low-traffic SMB sites often show no URL-level CrUX data — the moat may be invisible at very small scale. Per Dan Taylor (Search Engine Land, January 13, 2026, n=107,352), CWV is a gate, not a growth lever — severe failure suppresses AI Overview citation; Good → great does not lift it.

Built to last

Most SMB websites get rebuilt every two to four years (Orbit Media: average across the Inc 5000 is two years four months). Orbit's own client base — sites under continuous care — averages six years four months between rebuilds. Lifespan is a function of architecture and maintenance, not a fixed property of websites.

What kills sites in three to four years is documented: plugin abandonment and vulnerabilities (96% of WordPress vulns in 2024 were in plugins; 1,614 plugins removed from the .org repo for unpatched issues); performance decay (median desktop page now 2.56 MB and rising); accessibility lawsuits (5,000+ ADA filings in 2025; 95% of sites fail basic WCAG); page-builder lock-in (Divi 4 → 5 is a one-way migration); platform decay (Squarespace acquires Google Domains; Shopify sunsets checkout.liquid); and content decay plus link rot (Ahrefs' 14-billion-page study: 96.55% of pages get zero Google traffic).

Survivors — Daring Fireball (since 2002), Berkshire Hathaway (since 1997), Craigslist, Pinboard, GOV.UK Design System, Stack Overflow — share a common pattern: content separated from presentation, minimal dependencies, URL stability as a design principle, boring architecture as a feature.

Source: Research brief "Built to Last" (piece 5 of 15), citing Orbit Media on rebuild cadence, Patchstack on WordPress plugin vulnerabilities, HTTP Archive Web Almanac on median page weight, ADA filing data, and Ahrefs' 14B-page traffic study. Confidence: Verified on the named figures. Caveat: The recycled "average website lifespan = 2y7mo" stat is unverified at primary source (the HubSpot citation could not be located). No primary surveyed dataset of Canadian SMB rebuild cycles exists. Some rebuilds are legitimately driven by business pivots, capability growth, or compliance forcing functions — not all rebuilding is decay-driven.

Stage E — Post-launch measurement and maintenance

Post-launch is where the relationship structure of the engagement determines whether the compounding upside of a research-first build ever materializes. See SEO J-curve and new-site ramp for the canonical J-curve framing of new-site ramp.

Decouple from rank as the sole metric

A documented practitioner standard is that Google organic rank can no longer serve as the sole success metric for client SEO or content programs. The decoupling is multi-source and large: BrightEdge reports approximately 17% AI Overview citations come from organic top 10; Ahrefs documents a page-one overlap collapse from 76% to 38% in six to seven months; approximately 80% of AI-cited URLs are not in Google's top 100 across non-Google surfaces; Moz reports 88% of AI Mode citations are outside the top 10. Pew behavioural data confirms that the search session increasingly resolves on the SERP itself — citation matters more, raw click volume less.

The documented client-reporting structure leads with rank for high-intent commercial queries, then AI-citation rate on tracked queries, then branded and direct traffic trend.

Source: Atomic rule "Decouple from Google rank as sole success metric; AI citation, AI-Mode mentions, and direct-traffic / branded search now matter independently," citing BrightEdge, Ahrefs, Moz, Pew Research, and Seer. Confidence: Industry-consensus on the decoupling; per-vendor figures vary. Caveat: The Seer correlation between citation and CTR is real but not causal — report both citation and CTR separately.

Monitor AI Overview and pack presence

For every active client, the documented practice is to track the priority keyword set on three dimensions: whether an AI Overview appears, whether the client is cited inside it, and the local-pack position. Snapshots happen regularly, not just at engagement start. AI Overview prevalence is moving fast and uneven across query types — Whitespark Q2 2025 reports approximately 15% on local-intent queries and 92% on informational; Local Falcon reports 17.2% on commercial and 58.3% on informational. Recommendations grounded in last-quarter's data go stale quickly.

Source: Atomic rule "Monitor priority keywords for AI Overview presence + local-pack position," citing Whitespark Q2 2025, Local Falcon 2025, Semrush 2025 swing data, and the AI Overview ads rollout to Canada (Dec 19, 2025). Confidence: Industry-consensus.

Google Business Profile first

For Ontario service-business clients in 2026, the documented practitioner standard is that Google Business Profile (GBP) optimisation precedes content, paid, and AI-surface work. AI Overviews appear on only approximately 15% of simple local-intent queries; the local pack appears on approximately 93% of local-intent queries. With only three slots in the pack since the July 2014 Pigeon update, GBP is the single highest-leverage local-SMB asset and feeds AI local answers too.

Source: Atomic rule "Google Business Profile is still the primary local-lead surface for 'near me' queries (Ontario, 2026)," citing Whitespark Q2 2025 (15% local-intent AI Overview prevalence), Local Falcon 2025 (17.2% commercial), and the Pigeon update (July 24, 2014) three-pack shrink. Confidence: Industry-consensus.

Substantive freshness, not cosmetic

For pages where the working surface depends on currency, or where AI-citation matters, the documented refresh cadence is at least every 60 to 90 days, and the refresh must be substantive. Cosmetic date bumps do not count. Ahrefs reports AI-cited URLs are on average 1,064 days old versus 1,432 days for non-cited; BrightEdge reports approximately 1.9× appearance lift for pages updated within 60 days; AirOps reports 90+-day stale pages are approximately 3× more likely to lose citations; Amsive reports approximately half of AI citations are under 13 weeks old.

Source: Atomic rule "Keep frequently-updated content fresh SUBSTANTIVELY, not cosmetically; cadence at least every 60–90 days for the pages that matter," citing Ahrefs, BrightEdge, AirOps, Amsive, and Google's QDF (Query Deserves Freshness) conditional ranking factor. Confidence: Industry-consensus.

The relationship structure determines compounding capture

Whether the engagement captures the long-tail upside of a research-first build is a function of the relationship structure. Industry data documents that retainer-based agencies average 56-month client lifespans versus 24 months for project-based engagements. A 24-month project-based engagement ends roughly at the point a healthy site would be entering compounding — the asset gets handed over (or abandoned) just as it becomes valuable.

The same industry data documents annual client churn rates by service line: SEO 38%, PPC 49% (highest), full-service 25% (lowest). The SEO figure is consistent with the J-curve's invisible-window dynamic — clients abandon during the trough, converting a paper loss into a realised loss before compounding can occur.

Source: Focus Digital, 2026 Marketing Agency Churn Report. Confidence: Single-source. Caveat: Focus Digital is itself a marketing agency that sells retainer engagements, so the report has vendor incentive. The 56-vs-24 contrast is directionally consistent with the broader J-curve case but should not be quoted as an industry-wide measured figure. The principle — relationship structure matters for capturing compounding — is independently defensible.

Pricing transparency tradeoffs

A documented tension at the post-launch stage is whether to publish pricing. Publishing prices exposes cost structure to competitors and may force back-tracking on tailored services when scope exceeds what a public number implied. The strategic-exposure argument applies most strongly to differentiated-service businesses; for commodity or near-commodity services, the disqualification-via-public-estimator benefit may dominate.

Source: BlackCurve, "Should You Publish Your Prices Online?"; Legal Futures, "The pros and cons of publishing prices." Confidence: Industry-consensus.

The historical pattern that frames the methodology

The methodology is the latest application of a recurring pattern across every major Google shift since 1998. Each shift changed what got surfaced: from links (1998), to paid slots (2000 / Feb 2016), to local listings (2004–2014), to quality content (Panda / Penguin), to mobile pages (2015), to natural-language intent (Hummingbird / BERT), to experience-backed content (Helpful Content), to AI summaries (2024+). The three-step dynamic repeats: Google changes the surface, early adapters capture the new surface, laggards lose the old one. The defensible lesson for a research-first agency is the historical one — businesses that adapt early to being clear, specific, credible, and citable are the ones AI systems will surface, just as they were the ones that won every prior shift.

Source: Synthesis across atomic entries documenting each Google shift (1998 PageRank through 2024 AI Overviews). Confidence: Verified on each constituent shift; the pattern is interpretive but well-corroborated. Caveat: The pattern does not support the claim that every change was Google forcing businesses to spend more on ads — the dates and products are real; the motive of coerced spend is mostly unproven interpretation.

Cross-references

The methodology described here operationalizes the principles in Knowledge-base-backed website methodology. Its citation discipline is the editorial pattern documented at Editorial discipline and sourcing. Stack selection in Stage D defers to the canonical reference build at Default web stack (Astro on Cloudflare). The post-launch measurement framing in Stage E assumes the J-curve dynamics documented at SEO J-curve and new-site ramp — particularly the invisible-window stage during which most project-based engagements end before compounding begins.