Research brief: SMB widget capture layer — what owners can vs cannot self-report (June 2026)

Status: Synthesised June 2026. One of six briefs supporting the Candid SMB digital-difficulty self-assessment widget. Sister briefs: Research brief: SMB widget spend benchmarks — feasibility of a "digital-minus-ads" % of revenue (June 2026), Research brief: SMB widget difficulty-to-work mapping — three tiers of work for three sizes of gap (June 2026), Research brief: SMB widget presentation layer — tiered results without overclaiming (June 2026), Research brief: SMB widget market difficulty — six ranked factors (June 2026), Research brief: SMB widget vertical difficulty — two-axis tiering by industry (June 2026). Cluster entry point: Research cluster: SMB digital-difficulty self-assessment widget (six briefs, June 2026).

TL;DR

• An SMB owner can reliably self-report concrete, observable, countable, and recent facts about their own behavior (whether they run ads, website age, update frequency, review count) — but cannot reliably self-rate quality, skill, "optimization," or relative competitive standing, because those judgments are systematically inflated by well-documented biases. The widget should be built almost entirely from observation/counting tasks, not self-ratings.
• The single most dangerous question type for this widget is the self-rating of relative competitive position, because (a) the better-than-average effect makes most people place themselves above the median, (b) entrepreneurial overconfidence runs specifically through overplacement (belief in superiority relative to others), and (c) this bias is strongest precisely in markets perceived as easy — exactly the markets where the widget most needs accurate difficulty signals.
• The fix is well-supported by survey methodology: convert every judgment into an observation or count. Item-specific / behavioral questions measurably outperform abstract evaluative ("rate your SEO," "agree/disagree") items, with the largest documented quality gap coming from reduced systematic (method) bias rather than random noise. Tiered outputs (the widget's design) are also more defensible than precise scores given irreducible self-report error.

---

PART 1 — THE LEAD SPLIT: CAN vs CANNOT

What an SMB owner CAN answer reliably

These share four properties: observable, countable, recent, and behavioral (about their own actions, not evaluations).

Question type	Why it works
Do you currently run paid ads? (yes/no, and on which platforms) — CAN — "Do you currently run paid ads?" (behavioral, binary, current)	Binary, behavioral, current state — high recall accuracy
How many Google/Yelp reviews do you have? (count, in bands) — CAN — "How many Google/Yelp reviews do you have?" (counting from own profile)	Directly observable on the owner's own profile right now
What is your average star rating?	Observable number, not a judgment
When was your most recent review? (this week / month / 3+ months) — CAN — "When was your most recent review?" (recent, distinctive, checkable)	Recent, distinctive, checkable
How old is your website? (this year / 1–3 yrs / 3+ yrs / none) — CAN — "How old is your website?" (concrete date-based fact)	Concrete date-based fact
When did you last update/add content to your site? — CAN — "When was the site last updated?" (recent behavioral event)	Recent behavioral event
Do you have a Google Business Profile? Is it claimed? — CAN — "Do you have a Google Business Profile? Is it claimed?" (binary state)	Observable binary state
How many employees do you have?	Countable (with caveats)
Search your main service + your city. How many businesses appear above you? — CAN — "Search your main service + city. How many businesses appear above you?" (live counting task)	Converts a judgment into a counting task performed live
How many of the top results are paid ads vs. organic?	Observable counting task

What an SMB owner CANNOT answer reliably

These share one fatal property: they are subjective evaluations of quality, skill, or relative standing that the owner has no instrument to measure and strong motivation to inflate.

Question type	Why it fails
"How strong is your SEO?" — CANNOT — "How strong is your SEO?" (no instrument; owner has no access to the metric)	Owner has no access to the underlying metric; only 26% of business owners claim even a "good" or "expert" understanding of how Google ranks results (Fractl SEO knowledge-gap survey — only 26% of business owners claim "good" or "expert" SEO understanding)
"How well-optimized is your website?"	Requires expert audit; pure self-rating
"Rate your domain authority / keyword difficulty"	Owner does not know these and cannot estimate them
"How do you compare to your competitors?" — CANNOT — "How do you compare to your competitors?" (abstract → max overplacement)	Triggers overplacement + competition neglect — biased upward
"How difficult is your market?"	Underestimated, especially when the market is in fact easy to enter
"How good is your online reputation?" (as a rating)	Better-than-average effect; convert to review count/rating instead
"Rate your digital marketing maturity 1–10"	Abstract evaluative scale; low reliability, high method bias
Agree/disagree: "My business has a strong online presence"	Acquiescence bias (Acquiescence bias — over 100 studies; affects ~10-20% of responses) + self-enhancement

The design rule: Anything the owner must evaluate should be re-expressed as something they can observe or count. The widget's job is to supply the evaluation (the tier) from objective inputs — never to ask the owner for it. Codified as R1 — Convert every judgment into an observation or counting task.

---

PART 2 — DETAILED FINDINGS BY RESEARCH QUESTION

RQ1 — Self-report validity in general

The authoritative synthesis is Tourangeau, Rips & Rasinski, The Psychology of Survey Response (Cambridge University Press, 2000), which models the response process as four stages: comprehension → retrieval → judgment → response. Each stage introduces error. The framework is corroborated across the broader CASM (Cognitive Aspects of Survey Methodology) literature and Schwarz (1999, American Psychologist, "Self-reports: how the questions shape the answers"). See Tourangeau, Rips & Rasinski (2000) — the four-stage response model.

• Recall/memory error: Accuracy of autobiographical recall depends on event properties — recent, significant, and distinctive events are recalled far more accurately than distant, routine, or vague ones (Tourangeau/Rips/Rasinski 2000, chapters on "Factors Affecting Recall of Autobiographical Events" and frequency reporting). Design implication: ask about recent, distinctive events ("last review this month?") not lifetime or vague aggregates. — see R4 — Ask about recent, distinctive events — not vague aggregates.
• Self-reported behavior diverges from logged behavior: Parry et al. (2021, Nature Human Behaviour), a pre-registered meta-analysis of 106 effect sizes, found self-reported digital media use "correlates only moderately with logged measurements" and "self-reports were rarely an accurate reflection of logged media use." In a representative primary study, Junco (2013, Computers in Human Behavior) found students "spent an average of 26 min (SD = 30) per day on Facebook, significantly lower than the average of 145 (SD = 111) minutes per day obtained through self-report" — a ~5.6× overestimate, with self-report correlating with logged use at only r = 0.59. See Parry et al. 2021 — pre-registered meta-analysis: self-report ≠ logged media use and Junco 2013 — Facebook self-report overstates logged use by ~5.6x. Even sincere self-reports of one's own behavior are heavily inflated when drawn from memory; counts the owner reads off a screen are far safer.
• Self-reported objective business facts also diverge from administrative records. A matched register-and-survey study (Cabral & Gemmell, 2021, International Tax and Public Finance) found self-employment income underreporting of roughly 20% against tax-return data but only about half that (~10%) when based on survey data, the gap reflecting survey measurement error. A US Census/IRS-linked study (Imboden, 2023, CES Working Paper) found median self-employment income of $37,500 in survey data versus $18,500 in administrative records — survey reports roughly double the administrative figure. See Imboden 2023 (US Census/IRS) — survey self-employment income roughly DOUBLE the admin figure. Implication: even "objective" figures like revenue carry large error; favor binaries and bands over precise self-reported numbers, and read-off-screen counts over recalled ones.
• Social-desirability bias: Tourangeau & Yan (2007, Psychological Bulletin, "Sensitive Questions in Surveys") review shows misreporting on sensitive topics "is quite common and largely situational" — respondents over-report desirable behaviors (voting) and under-report undesirable ones. Self-administration (as in a web widget) reduces but does not eliminate this. See Tourangeau & Yan 2007 — social-desirability misreporting is common and situational.
• Acquiescence bias: Over 100 studies show respondents tend to agree with statements regardless of content; estimates of the affected proportion run ~10–20%. See Acquiescence bias — over 100 studies; affects ~10-20% of responses. Design implication: avoid agree/disagree framing entirely — codified as R2 — Never use agree/disagree statements in the widget.
• Question wording/scale effects: Schwarz, Hippler, Deutsch & Strack (1985, Public Opinion Quarterly) showed respondents reading a TV-use scale centered on low values reported less usage (and rated TV more important) than those given a high-value scale — respondents infer the "normal" value from the scale and anchor to it. See Schwarz et al. 1985 — response-option ranges leak information and anchor the answer. Response-option ranges leak information; use neutral, behaviorally realistic bands or open counts.

RQ2 — Confidence vs. accuracy (calibration)

• The hard-easy effect: Lichtenstein, Fischhoff & Phillips (1977; reviewed 1982) established that overconfidence INCREASES as task difficulty increases — people are overconfident on hard items and can be UNDERconfident on easy ones. See Lichtenstein, Fischhoff & Phillips 1977 — the hard-easy effect. This dovetails with Moore & Healy 2008 (Moore & Healy 2008 — the three forms of overconfidence (estimate / place / precision)): on EASY markets the bias runs through overplacement (vs others); on HARD markets it runs through absolute overestimation.
• Calibration can improve with framing: Gigerenzer, Hoffrage & Kleinbölting (1991) and Juslin (1994) showed that representative / ecological item selection and frequency framing can sharply reduce or eliminate measured overconfidence — though critics (Griffin & Tversky 1992) note this partly confounds item selection with difficulty. See Gigerenzer/Juslin 1991-1994 — frequency framing reduces overconfidence (with Griffin-Tversky 1992 critique). Design implication: ask for frequencies / counts ("how many…") rather than subjective probabilities ("how likely…").
• Confidence does not reliably track accuracy for subjective judgments; the relationship is task-dependent and often weak. See Confidence does not reliably track accuracy for subjective judgments. The widget must therefore not use confidence as a proxy for knowledge or skill.

RQ3 — Entrepreneur/business-owner overconfidence (the central risk)

This is the most decision-relevant block for the widget.

• The taxonomy: Moore & Healy (2008, Psychological Review, "The Trouble with Overconfidence") distinguish three forms: overestimation (of one's absolute performance), overplacement (belief one is better than others), and overprecision (excessive certainty). Critically: on easy tasks people underestimate their absolute performance but overplace — believe they are better than others; on hard tasks the reverse. See Moore & Healy 2008 — the three forms of overconfidence (estimate / place / precision).
• Entry/competition overconfidence: Camerer & Lovallo (1999, American Economic Review, "Overconfidence and Excess Entry") demonstrated experimentally that excess market entry is driven by overconfidence about relative skill, and that entry was highest when participants knew success depended on skill and self-selected in — "reference-group neglect": entrants ignore that competitors also self-selected. See Camerer & Lovallo 1999 — excess entry driven by overconfidence about RELATIVE skill.
• What drives entry is overplacement, not absolute confidence: Cain, Moore & Haran (2015, Strategic Management Journal, "Making sense of overconfidence in market entry") reconciled the puzzle — "entry into different markets is not driven by confidence in one's own absolute skill, but by confidence in one's skill relative to that of others." People choose markets they perceive as easy, and easy markets maximize overplacement. See Cain, Moore & Haran 2015 — OVERPLACEMENT (not absolute confidence) drives entry; easy markets pull entrants in.
• The easy-market asymmetry (the key finding for this widget): Moore & Cain (2007, Organizational Behavior and Human Decision Processes) — on easy tasks/markets, overplacement is greatest, so entrants flood in believing they will beat competition they have not actually assessed. See Moore & Cain 2007 — overplacement is GREATEST on easy tasks/markets. Implication: when an owner is in an "easy" vertical, their self-assessment of competitive position is most inflated — precisely where the widget must override self-rating with observation. Codified as R6 — In easy-looking markets, override owner optimism with observed counts.
• Entrepreneurs specifically: Cooper, Woo & Dunkelberg (1988, Journal of Business Venturing 3:97–108) surveyed 2,994 entrepreneurs and found "they perceived their prospects as very favorable, with 81% seeing odds of 7 out of 10 or better and a remarkable 33% seeing odds of success of 10 out of 10," and they rated their own odds higher "than other new business owners with similar ideas." Notably, "those who were poorly prepared seemed just as optimistic as those who were well prepared" — even though roughly half or more of new firms do not survive their first several years. See Cooper, Woo & Dunkelberg 1988 — 81% of entrepreneurs see odds 7/10+; 33% see 10/10. Busenitz & Barney (1997, Journal of Business Venturing) found entrepreneurs show significantly more overconfidence and representativeness bias than managers in large organizations — see Busenitz & Barney 1997 — entrepreneurs more overconfident than corporate managers.
• The better-than-average effect: Svenson (1981) found ~93% of US drivers rated themselves above the median for skill; Alicke et al. (1995) and many replications confirm robustness. Crucially, the effect shrinks when comparison targets are concrete rather than abstract. See Svenson 1981 — 93% of US drivers rated themselves above the median for skill and The concrete-target moderation — the design lever against overplacement. The widget exploits the moderation directly via R3 — Where comparison is unavoidable, use a concrete NAMED competitor target.
• Dunning-Kruger: Kruger & Dunning (1999, JPSP) — bottom-quartile performers (12th percentile actual) estimated themselves at the 62nd percentile; the unskilled lack the metacognition to recognize incompetence. Serious critiques exist: Krueger & Mueller (2002) and others argue the pattern is substantially a statistical artifact (regression to the mean + better-than-average effect); Nuhfer et al. found little tendency toward inflated self-assessment for most people. See Kruger & Dunning 1999 — low performers cannot self-assess skill (with caveats). Honest concession: we should not lean hard on a strong DK mechanism; the safer, better-established claim is simply that low performers cannot accurately self-assess skill.

Net direction and magnitude for the widget: The bias is upward (owners overstate competitive position and digital strength), it runs primarily through overplacement, and it is strongest in easy markets. This is not a small effect — entrepreneurial samples show large majorities (81%) rating their own odds far above realistic survival rates.

RQ4 — Proxy questions (observable correlates of unmeasurable metrics)

The strategy: replace each metric the owner cannot assess with an observable behavior or count that correlates with it.

• Competitive density → count of businesses/ads appearing above the owner in a live search for their service + city. This is a counting task, not a judgment (CAN — "Search your main service + city. How many businesses appear above you?" (live counting task)); competition-neglect research implies owners under-weight competitors when asked to judge but can count them when directed to look. Directional-Speculative on the precise correlation; the conversion principle itself is Verified via BARS / item-specific literature.
• Digital maturity → website age + update recency + presence/claim status of Google Business Profile (CAN — "How old is your website?" (concrete date-based fact), CAN — "When was the site last updated?" (recent behavioral event), CAN — "Do you have a Google Business Profile? Is it claimed?" (binary state)). These are concrete dates/binaries. Industry-consensus that these correlate with digital maturity; treat the strength of correlation as Single-source/Directional.
• Reputation signal strength → review count (in bands), average rating, and recency of most recent review (CAN — "How many Google/Yelp reviews do you have?" (counting from own profile), CAN — "When was your most recent review?" (recent, distinctive, checkable)). BrightLocal's Local Consumer Review Survey 2024 (representative panel of ~1,000 US adults) found that ~75% of consumers read reviews "always" or "regularly," just 3% "never" do, and "88% of consumers would use a business that replies to all of its reviews, compared to just 47% who would use a business that doesn't respond to reviews at all" — confirming that review count, rating, recency, AND response behavior are the signals consumers act on. Single-source / vendor — see quarantine; the underlying consumer-behavior direction is consensus.
• Ad presence/intensity → "do you run ads, on which platforms, roughly what monthly budget band" (CAN — "Do you currently run paid ads?" (behavioral, binary, current)). Behavioral and current. Verified as answerable; budget bands less reliable than yes/no.

Caveat: proxies are correlates, not equivalents. The widget should treat them as inputs to a tier, never as precise measurements, and should be transparent that it is estimating.

RQ5 — SMB digital-marketing literacy

The evidence here is largely vendor-originated and must be quarantined (see Part 3), but the direction is consistent across multiple independent vendor surveys, which raises confidence to a directional consensus:

• The Fractl SEO Knowledge Gap survey (977 people, 394 business owners) found "Only 13% of consumer respondents and 26% of business owners believed they had a 'good' or 'expert-level' understanding of why Google shows certain results"; on the term "organic search," "half of the respondents incorrectly guessed... with another 23.1% saying they didn't know"; and on "backlinks," "Nearly 39% of respondents didn't get the answer right, while another 28.6% said they didn't know." See Fractl SEO knowledge-gap survey — only 26% of business owners claim "good" or "expert" SEO understanding. Single-source, vendor — directional only.
• A separate survey of 500 US small-business owners (reported via Smart Insights/Numiko) found over one-third had no understanding of SEO and over half only "basic" understanding. Single-source, vendor.
• The Manifest (2020) reported only ~30% of small businesses use SEO at all. Single-source, vendor. See Other vendor surveys — Smart Insights, The Manifest — converge on low SEO literacy.

What owners reliably understand: the existence of keywords and that Google ranks results; whether they personally run ads; that reviews matter. What they systematically misunderstand: how rankings are determined, backlinks/authority, "optimization," and their own relative ranking. Directional consensus across independent vendor surveys; flagged commercial incentive. This directly validates the widget's premise: ask about observable behavior, not SEO self-assessment.

RQ6 — Questionnaire design to reduce bias

• Item-specific (construct-specific) beats agree/disagree. Saris, Revilla, Krosnick & Shaeffer (2010, Survey Research Methods 4:61–79), using split-ballot multitrait-multimethod (SB-MTMM) experiments in the European Social Survey, found item-specific response options yielded much higher data quality than agree/disagree scales — item-specific quality coefficients (q² = reliability × validity) around 0.74–0.89 versus agree/disagree around 0.18–0.51, with the gap driven overwhelmingly by validity (reduced method/acquiescence bias), not reliability. In their illustrative decomposition the quality gap of 0.44 broke down as only a 0.05 reliability difference but a 0.54 validity difference (item-specific validity ≈ 1.00 vs. agree/disagree ≈ 0.46). They conclude "responses to A/D rating scale questions indeed had much lower quality than responses to comparable questions offering IS response options." See Saris et al. 2010 — item-specific response options far outperform agree/disagree on data quality.
• But the benefit is contested for criterion validity. Lelkes & Weiss (2015, Research & Politics, "Much ado about acquiescence"), using two ANES waves (n ≈ 5,900), found construct-specific questions were no better than agree/disagree on test-retest reliability (both averaged polychoric r = 0.63) or criterion validity (differences of -0.02 to -0.05, confidence intervals overlapping), even among acquiescence-prone respondents (low verbal ability, high agreeableness, face-to-face mode). They conclude construct-specific format "should not be considered a canonical solution." See Lelkes & Weiss 2015 — skeptical counter: item-specific no better on criterion validity. Important skeptical counterpoint. The reconciliation: Saris measured CONVERGENT validity via MTMM, Lelkes & Weiss measured CRITERION validity — different facets. Honest concession: rewriting items helps measurement quality, but it is not magic; the bigger win is converting judgments to observations entirely.
• Behaviorally Anchored Rating Scales (BARS) (Smith & Kendall, 1963, Journal of Applied Psychology): anchoring each scale point to a concrete, observable behavior reduces measurement bias (halo, leniency, recency) relative to abstract scales — "sometimes (but not always)," and the benefit depends on rigorous development. See BARS (Smith & Kendall 1963) — anchoring scale points to concrete observable behaviors. Verified — with the honest "not always" caveat from the primary literature.
• Vague quantifiers leak norms and vary by respondent. Schwarz et al. (1985) and the vague-quantifier literature show "often/sometimes/rarely" mean different frequencies to different people. Open numeric counts or behaviorally realistic bands are preferable where the count is knowable. Mixed evidence: some studies (e.g., Schneider & Stone 2016 on quality-of-life scales) find vague quantifiers can sometimes match open numeric formats when exact counts are hard to retrieve — so use counts when the owner can read them off a screen, bands when they must estimate. See Schneider & Stone 2016 — caveat: vague quantifiers can sometimes match open numeric formats.
• The general principle: concrete/behavioral items outperform abstract/evaluative ones for both accuracy and bias resistance. This is the throughline connecting BARS, item-specific scales, the concrete-target moderation of the better-than-average effect, and the logged-vs-self-report gap.

---

PART 3 — QUARANTINED VENDOR SOURCES (commercial incentive flagged)

These were used only as directional signal, never as the backbone of any claim. Each has a commercial incentive to portray SMBs as under-served by SEO/marketing and therefore in need of paid services or tools. The aggregated quarantine entry is Caveats — vendor quarantine and survivorship bias in the capture-layer evidence; the per-source incentives below are preserved here for transparency.

• Fractl (fractl.com) — content-marketing/SEO agency. Its "SEO knowledge gap" survey markets its own SEO expertise. Incentive: sell SEO services; motivated to show owners are ignorant of SEO.
• Search Engine Land (owned by Semrush, an SEO-tool vendor) — reporting on the Fractl survey; the page itself advertises Semrush. Incentive: sell SEO software.
• Smart Insights / Numiko — digital agency reporting an SEO-understanding survey; promotes an "SEO Toolkit." Incentive: sell digital-marketing services/training.
• The Manifest — B2B vendor-listing/lead-gen site; "30% use SEO" stat promotes the case for hiring SEO providers it lists. Incentive: lead generation for agencies.
• BrightLocal (brightlocal.com) — local-SEO software vendor. Its Local Consumer Review Survey is methodologically transparent and widely cited, but BrightLocal sells review/local-SEO tools. Incentive: demonstrate that reviews/local SEO drive revenue. Used here only for the direction that review count/rating/recency matter — a direction independently plausible and consensus.
• SeoProfy, RivalMind, Firstep, TechBullion, surfsigma, Marketing LTB — SEO agencies/aggregators citing recycled statistics, often without primary sourcing. Incentive: sell services; statistics frequently uncited or daisy-chained. Treated as non-authoritative.

Note: Grokipedia and Wikipedia were used only to locate primary sources (Svenson 1981, Alicke 1995, the acquiescence literature), not as authorities themselves.

---

PART 4 — QUESTIONNAIRE DESIGN FOR THE WIDGET

Build entirely from observable/countable inputs

Every widget question should pass this test: could the owner answer it correctly by looking at something or counting, without making a quality judgment about themselves? If not, redesign it.

The core technique: convert judgments into observation/counting tasks

This is the single most important design move, supported by (a) the item-specific > agree/disagree finding, (b) the concrete-target moderation of the better-than-average effect, (c) BARS, and (d) the competition-neglect literature (owners under-weight competitors when judging but can count them when directed). Codified as R1 — Convert every judgment into an observation or counting task.

Before → After examples (digital presence & competitive density):

❌ Judgment question (unreliable)	✅ Observation/counting task (reliable)
"How strong is your SEO?"	"Search your main service + your city in Google. Counting only the regular (non-ad) results, how many businesses appear above you? (I'm not listed / 0 / 1–3 / 4–10 / more than 10 / I don't appear)" — CAN — "Search your main service + city. How many businesses appear above you?" (live counting task)
"How competitive is your market?"	"In that same search, how many paid ads appear before the first regular result? (0 / 1–2 / 3–4 / 5+)"
"How good is your online reputation?"	"Open your Google Business Profile. How many reviews do you have? (none / 1–9 / 10–49 / 50–199 / 200+)" + "What's your average star rating?" + "When was your most recent review? (this week / this month / 1–3 months ago / longer / never)" — CAN — "How many Google/Yelp reviews do you have?" (counting from own profile)
"How modern/optimized is your website?"	"When did you last add or change content on your website? (this month / this quarter / this year / over a year ago / I don't have a website)" + "How old is your website? (under 1 yr / 1–3 / 3+ / none)" — CAN — "When was the site last updated?" (recent behavioral event)
"Rate your digital marketing maturity 1–10"	A set of binaries: "Do you run paid ads? Do you have a claimed Google Business Profile? Do you collect reviews systematically? Do you update your site at least monthly?"
"Do you compare well to competitors?" (abstract)	"Pick your single biggest competitor. Look at their Google profile. Do they have more reviews than you, about the same, or fewer?" (concrete, named target — shrinks the better-than-average effect; see R3 — Where comparison is unavoidable, use a concrete NAMED competitor target)

Additional design rules (R1-R7)

1. R1 — Convert every judgment into an observation or counting task — Convert every judgment into an observation or counting task.
2. R2 — Never use agree/disagree statements in the widget — Never use agree/disagree statements. Use item-specific response options or counts (Saris et al. 2010). Where the construct is evaluative, prefer an observation proxy entirely.
3. R3 — Where comparison is unavoidable, use a concrete NAMED competitor target — Where comparison is unavoidable, anchor to a concrete, named competitor target.
4. R4 — Ask about recent, distinctive events — not vague aggregates — Ask about recent, distinctive events, not vague lifetime aggregates (recall literature).
5. R5 — Output tiers, not precise scores — Output tiers, not precise scores. Self-report error is irreducible; the meta-analytic gap between self-reported and logged behavior alone justifies coarse tiers over false precision.
6. R6 — In easy-looking markets, override owner optimism with observed counts — Treat the easy-market case as the highest-risk zone: when inputs suggest a low-competition vertical, weight observed counts heavily and discount any owner optimism (overplacement peaks there).
7. R7 — When the owner can read a number off a screen, have them do so — Prefer open counts or behaviorally realistic bands to vague quantifiers; when the owner can read a number off a screen, have them do so rather than estimate from memory (Junco 2013 shows recalled counts overstate by ~5×).

---

RECOMMENDATIONS (staged, with thresholds)

Stage 1 — Build the v1 questionnaire from CAN-answer items only.

• Use exclusively the question types in the Part 1 "CAN" table and the "After" column of the Part 4 conversion table.
• Ship zero self-rating items and zero agree/disagree items in v1.
• Output 3–4 relative-difficulty tiers, never a numeric score.
• Threshold to revisit: if user testing shows >15–20% drop-off on the live-search counting task, simplify it (e.g., pre-fill the search) rather than reverting to a self-rating.

Stage 2 — Validate the proxies against ground truth (this is the thinnest evidence link).

• Once the widget has run on a sample of businesses, compare self-reported inputs (review count, website age, competitor count) against scraped/API ground truth (Google Business Profile API, live SERP, WHOIS/site age).
• Quantify the self-report error for each input and recalibrate the tier thresholds.
• Threshold: if any single input's self-reported value diverges from ground truth by more than ~one band for >25% of users, either drop that input or replace the self-report with a direct measurement where the widget can fetch it.

Stage 3 — Use measurement to backstop the highest-risk inputs.

• For competitive density and review signals — the two highest-stakes, most-biased constructs — fetch the data directly (live SERP, profile API) instead of, or alongside, asking. Reserve self-report for what genuinely cannot be fetched (e.g., whether they run ads, internal update cadence).
• Decision rule: prefer measurement wherever it is technically cheap; reserve self-report for behavioral facts only the owner knows.

Stage 4 — Guard against self-selection in interpretation.

• Because widget users self-select (worried owners) and exclude failed firms, do not generalize tier distributions to "all SMBs." Report tiers as relative within the user pool, not as population estimates.

---

CAVEATS & HONEST GAPS

• Self-selection / survivorship bias: Nearly all SMB and entrepreneurship survey evidence (Cooper-Woo-Dunkelberg, the SEO-literacy surveys) samples active, surviving businesses and willing respondents. Failed firms and non-responders are excluded, which understates the prevalence of fatal overconfidence and may overstate average digital sophistication. Widget users also self-select, so their baseline differs from the population.
• The proxy-to-metric correlations are the thinnest-sourced link. That website age/update frequency map to "digital maturity," or that competitor-count maps to "keyword difficulty," is plausible and directionally supported but not precisely quantified in independent academic literature. Present these as estimates and validate against real outcomes (Stage 2).
• Dunning-Kruger is contested. Lean on the robust, simple claim (low performers can't self-assess skill) rather than the strong metacognitive-deficit mechanism, which faces credible statistical-artifact critiques (Krueger & Mueller 2002).
• The item-specific advantage is facet-dependent. It is strong for convergent validity / method bias (Saris 2010) but not clearly better for criterion validity (Lelkes & Weiss 2015). The defensible conclusion is "convert judgments to observations," not merely "reword the scale."
• Most SMB digital-literacy evidence is vendor-originated. The direction (owners poorly understand SEO/rankings) is consistent across multiple independent commercial surveys, but no high-quality independent academic measurement was located; treat magnitudes as directional.
• Geographic scope: consumer-review and SEO-literacy data are US-centric (some UK). Canadian-specific SMB self-report literacy data was not separately located; the cognitive/survey-methodology findings (Tourangeau, Schwarz, Moore, Saris, etc.) are not geographically bounded and apply to both US and Canadian owners.

See Caveats — vendor quarantine and survivorship bias in the capture-layer evidence for the consolidated caveats entry.