Duplicate local shop results: 90% corroboration favors entity-preserving merges

Leadgen Economy puts duplicate-related waste from leads at $12,500–$75,000 under its $25–$50 CPL scenario. That is an advertising-cost estimate, not a local-shop match rate. The distinction matters: repeated shop cards may reflect faulty matching, multiple inventory records, or several legitimate locations that should remain discoverable in local search.

Local discovery needs entity resolution before ranking. Normalize exact identifiers, then use fuzzy names, company information, and addresses to nominate possible matches—not automatic merges. Microsoft Dynamics presents possible duplicates for comparison, and Lucidity separates identifying a plausible match from choosing a safe action.

Preserving entities protects the candidate pool: ranking redistributes visibility among available results, but cannot restore a location removed upstream. Corroboration should support only the matches it actually verifies, especially where an unverified chain-wide merge could hide a legitimate shop. A CRM scenario with 30% duplicates illustrates frequency, not whether records identify one location or several; it cannot by itself justify collapsing cards or an entire chain.

Nominatim, Names, Addresses

A Nominatim name–address–coordinate match is place evidence, not a merchant-identity verdict. From a recommender-systems perspective, I keep three objects separate: a result card is a retrieval observation; a shop-location is the entity being discovered; and a canonical shop record is the identity authority for retrieval and ranking. According to Lucidity, citing Salesforce documentation, a matching rule compares records; the duplicate rule governs the action. Matching is not merge authorization.

Build the exact-match key from normalized trading or legal name, street number, shop-unit identifier, postcode, and geocoded coordinates. According to Octeth’s Email Marketing Glossary, leading or trailing whitespace and hidden characters can create apparent duplicates, so trim whitespace before comparison. Retain branch and unit information rather than stripping distinguishing tokens, and preserve coordinate precision and the rounding policy. At the equator, one ten-thousandth of a degree spans approximately 1.1 meters of latitude and 0.6 meters of longitude. That distance makes geocoder variation and premature rounding identity-sensitive, not cosmetic noise.

Label E a collision rate, not a verified-duplicate rate: it measures rows sharing the complete normalized key, not proven merchant duplicates. Different operators can occupy one building, while genuine duplicates can differ in spelling or coordinates. The Real Estate Data Deduplication Guide treats “123 Main St” and “123 Main Street, Unit 4B” as the same property, but supplies no method for validating the unit suffix. For shops, a building-level record must not silently inherit a specific unit.

For C, an operator-controlled website and an independently maintained telephone record are suitable sources only when both support the actual location. A shared brand, registered office, or customer-service number establishes organizational linkage, not storefront identity. Nominatim corroborates a place; it cannot supply both independent merchant-location sources.

Separate branches with different walk-in addresses remain separate, as do different operators occupying distinct mall or market units—even with identical names and building addresses. Two reviewers should label these cases independently, with adjudication of disagreements, before computing S. Preserve the physical-location evidence behind each label, not merely the resulting decision.

Use one noncompensatory batch gate: merge only when all three required results hold.

Implementation close: freeze normalization, coordinate precision, reviewer instructions, and evidence links before scoring. Favorable aggregate evidence cannot cancel a confirmed split; otherwise, keep confirmed distinct and unresolved shop pairs separate.

9 Billion Places and One Listing

Foursquare’s Places API dataset documentation is a coverage reference, not a local-shop error-rate benchmark. A larger catalog enlarges candidate search; it does not validate a proposed shop-identity match. The relevant evidence has to be classified before it enters a matching pipeline.

For reproducibility, I would archive the exact Foursquare documentation version consulted and preserve its access date; a bare, mutable product link does not establish what was actually reviewed. The catalog totals belong in a coverage note, never in the denominator of a claimed false-merge rate.

Google’s 2025 guideline is a policy boundary, not a mathematical identity law. Its service-area-business eligibility language must remain attached to the claim: the guideline constrains eligible business representations on Google, not equivalent observations throughout a search index. A name shared by legitimate branches, separate operators in market units, or a relocated business therefore supplies no merge instruction.

The BrightLocal survey concerns consumer-review trust, not how often duplicate shops occur or whether records identify the correct location. A favorable consumer-review experience supplies no evidence that two shop-location records should become one.

According to the OpenStreetMap Foundation’s 2024 report, the contributor and edit figures characterize a changing, collectively maintained base. For local-place matching, a dated extraction and source-date disclosure make changes auditable; without them, an attribute update can silently alter a match between runs.

Use a provenance ledger, not a popularity score: preserve documentation versions, platform eligibility, consumer-research dates, and extraction dates alongside candidate evidence. These sources do not settle the article’s batch-level identity question. They supplement, rather than replace, its existing evidence and audit; a confirmed distinct shop-location pair falsifies the batch-level merge decision. Catalog reach, review trust, and repeated names cannot substitute for identity evidence.

Corroboration and No Confirmed Splits

Entity-preserving merging wins; producing the smallest card count does not. A canonical record must represent one shop-location, not an entire brand: repeated names are not proof of duplicate listings. My conservative operating policy accepts a candidate batch only when E > 0%, C passes the corroboration gate, and S records no confirmed distinct-location pairs. Any confirmed distinct shop-location vetoes the batch. These are my operating gates, not thresholds established by the cited catalog or policy documents. Otherwise, confirmed distinct and unresolved pairs stay separate.

Here, E requires an exact-key collision, while C measures the share of candidate links supported by independent evidence—not additional fields copied from the same source. In the matrix, “gate met” means the corresponding E or C threshold is satisfied. S records confirmed distinct-location pairs after adjudication. Merge requires all three gates; a valid split makes the otherwise qualifying candidate batch Split batch.

The S denominator must consist of adjudicated candidate pairs drawn from the production candidate pool, with every disposition and its supporting location evidence retained. A result with no confirmed distinct-location pairs is this guide’s required safety gate—not proof that name similarity, catalog scale, or a one-listing policy has demonstrated match accuracy. According to Lucidity’s review of Microsoft Dynamics 365, exact email and phone rules and context-aware similar-name rules can present possible duplicates for comparison; its Salesforce review distinguishes record matching from the duplicate jobs that decide handling. Neither establishes local-shop accuracy. The supplied source set leaves the three-rate framework, similarity cutoff, and merge/split rule undefined.

Before changing the index, freeze the queries and compare their first 10 positions under candidate identity resolution and ordinary separate-record retrieval. Keep the result cap, features, and ranking configuration fixed so identity resolution is the experimental difference. For every verified shop-location, record its query, baseline position, candidate position, and whether it appears, disappears, or moves. A legitimate shop-location disappearing is a ranking consequence even when the total card count falls: a cleaner-looking result set is not automatically a better one.

Next action: Construct the production-pool audit with retained adjudication evidence, then run the frozen-query comparison. Withhold the index change unless the whole batch passes; keep every confirmed distinct and unresolved pair separate.

What the Data Doesn't Tell You

The missing evidence is pair-level, not another aggregate return-on-investment statistic. According to Leadgen Economy, the supplied figures concern lead costs; according to BatchData, its improvement claim concerns return on investment; and according to Octeth, the deduplication discussion concerns email lists. None supplies a named local-search provider, sampled local-shop candidate pairs, and independently verified shop-location identity labels. This plan therefore cannot claim that a provider already possesses effective merge precision for local-shop records. Require that primary pair-level audit before making a provider-wide claim.

A stipulated clean sample under its sampling assumptions does not promise zero future errors. A persistent error can remain invisible when every sampled pair comes from easy, frequently edited records. Include difficult, rarely updated candidates, document the sampling frame, and report uncertainty for the population the sample actually represents.

Corroboration also needs a lineage test. A telephone directory can copy an operator website, and several providers can preserve the same old address. Those are repeated transmissions of an assertion, not independent verification. Trace field-level provenance and originating source families before granting independent-corroboration credit. Several matching web pages can otherwise create the appearance of agreement while leaving the underlying location claim unsupported.

Expect results to vary with address complexity and operator type. A mall can contain separate business units at one address, while a small shop can cross a municipal boundary without changing its name. Report dense retail centers, suburban strips, rural settlements, and business complexes separately rather than extrapolating one market’s audit. Legitimate branches, separate operators in market units, and relocated businesses can share a name without sharing a shop-location.

A deliberate never-merge baseline is a useful stress test, not the discovery solution. It avoids false merges by refusing the risky operation while leaving duplicate observations and fragmented shop evidence in the index. Precision alone therefore cannot establish that local-shop discovery is solved. One false merge can conceal a legitimate merchant; false splits can scatter that merchant’s evidence across competing cards and fragment ranking presence. Compare discovery quality with the baseline instead of treating fewer cards as better data.

Keep the article’s entity-preserving requirement: merge only a batch with a positive exact-collision rate, corroboration meeting the required gate, and no confirmed distinct location in the stipulated audit. Otherwise, keep confirmed-distinct and unresolved shop pairs separate. Any confirmed distinct location falsifies the entire batch-level merge decision; supporting evidence elsewhere cannot rescue it. These audit results apply only to the audited, independently supported batch—not to unsampled records or every locality.

Portland

Portland’s defensible label is “planned audit,” not a completed merge test. Freeze a versioned Overture Maps Place-theme extract within a 10-kilometer straight-line radius of Portland, Maine, centered at 43.6591, -70.2568. Archive the release identifier, extraction date, licence, unmodified records, and a content hash. The boundary specifies coverage; it supplies no collision, corroboration, or split results.

Precommit 50 shopping queries—10 each for coffee, hardware, pharmacies, bookshops, and repair shops—and replay them against that frozen catalogue, retaining at most 20 returned shop records per query. Publish every query string, eligibility filter, ordering function and version, tie-break rule, and actual returned-row count. This is a catalogue experiment, not a live search-engine or market-share study. A repeated name creates a candidate collision, not evidence of a shared location.

Use a published, reproducible sampling frame to select actual candidate pairs from the retained ledger. Define pair identifiers, exact-collision, same-address, and same-brand strata, selection order, and the random seed before selection; retain overlapping category memberships. Report observed counts, including unresolved pairs. An empty stratum stays empty: never fabricate records or backfill convenient examples to reach a target.

Two reviewers independently check premises, retaining each evidence item separately: the operator’s own contact information and an independent business or premises source. Capture access dates and address or unit evidence; adjudicate disagreements without dropping unresolved pairs. Corroboration requires agreement on the same shop-location. Following Lucidity, keep plausible-match retrieval separate from the decision to canonicalize.

The finished guide must insert arithmetic from returned rows and completed reviews, not substitute sampling targets. These three required observed rates are currently unreported:

The proposed rule permits a batch-level merge only when E > 0%, C passes the corroboration gate, and S records no confirmed distinct-location pairs. Until those counts exist, authorize no batch merge: keep confirmed-distinct and unresolved pairs separate and leave other candidates unmerged pending review. Every completed pair ledger needs a disposition for every audited pair. A single confirmed valid split falsifies batch authorization, even with a perfect similarity score. Legitimate branches, separate operators within a market unit, and relocations can share a name without sharing a location. According to mbrenndoerfer.com, character-shingle Jaccard similarity supports near-duplicate candidate retrieval, not premises verification.

Publish the pair ledger, both reviewer decisions, evidence links, adjudication record, catalogue release, and per-query first-ten retrieval comparison against a named, versioned reference ranking over the same extract. Record membership and rank changes without extrapolating beyond that comparison. Keep this exposure review distinct from full returned-row counts. The next action is to publish the release and query specification, then populate the table from observed counts. Without those artifacts, retain the “planned audit” label; do not present an illustrative percentage as research.

Five Concrete Rules for Merge-or-Split Decisions

A clean duplicate-removal dashboard is not an identity audit: it can show fewer cards while a legitimate shop-location has disappeared. Release a batch only when the exact-collision, corroboration, and safety-audit gates all pass. The batch-level audit is an acceptance criterion, not a per-pair guarantee; a confirmed distinct shop-location is a hard stop.

1. Make independence auditable. Score corroboration per sampled pair, then deduplicate evidence by source family before counting it toward C. A copied directory entry and its upstream source express one assertion, not independent confirmations. A shared telephone number or higher name similarity can trigger review, but neither repairs missing evidence. In a hypothetical “Riverside Market” case, a shared line with another operator does not corroborate one shop-location; insufficient support means split, not “same brand, same place.”

2. Make a confirmed conflict a circuit breaker. Adjudicate the underlying shop-location, not merely whether record fields disagree. A geocoding discrepancy can reflect address normalization, while a relocation can preserve business history without preserving location identity. A confirmed distinct pair blocks the entire batch; do not dilute it with an easy-match majority. Retain unresolved pairs as separate rather than promoting them to “probably same.”

3. Interpret zero as “not measured,” not “clean.” Tie E to the qualifying automatic-merge cases actually measured for this batch. A zero E result means no qualifying collision was found—not that the local index is duplicate-free. Retain the collision definition, candidate-batch identifier, and not-detected status. This prevents missing evidence from being silently reclassified as a clean index.

4. Test coverage downstream. Validate the first 10 retrieved positions for each evaluated query against a frozen set of verified shop-location IDs, not raw card totals. Compare the release with the pre-merge baseline for legitimate-operator coverage. A lower duplicate count is a regression if a

Frequently Asked Questions

What does the exact-collision rate E measure, and can E > 0% authorize a merge?

E is the share of result rows sharing the complete normalized key with another row in the same result set, and E > 0% generates candidates rather than proving merchant duplication.

At the equator, what distance does 0.0001 degrees represent in latitude and longitude, and why does it matter?

At the equator, 0.0001 degrees of latitude spans approximately 1.1 meters and 0.0001 degrees of longitude spans approximately 0.6 meters, making geocoder variation and premature rounding identity-sensitive rather than cosmetic noise.

What evidence satisfies the independent-corroboration gate C?

Passing C requires at least two independent sources about the actual shop-location, while a shared brand, registered office, or customer-service number establishes only organizational linkage rather than storefront identity.

Can “123 Main St” and “123 Main Street, Unit 4B” be merged simply because a deduplication guide treats them as the same property?

No; although the Real Estate Data Deduplication Guide treats the addresses as the same property, it supplies no method for validating the unit suffix, so a building-level shop record must not silently inherit a specific unit.

Can strong aggregate evidence override one confirmed pair of separate shop-locations?

No; any confirmed distinct-location pair falsifies the batch-level merge decision, and favorable aggregate evidence cannot cancel a confirmed split.

Do Foursquare’s reported totals of more than 1.9 billion places and more than 50 million businesses measure local-shop duplication or false-merge performance?

No; those totals describe corpus coverage rather than a measured local-shop duplicate or false-merge rate and belong in a coverage note, not the denominator of a claimed false-merge rate.

Quick answers

Also worth reading: Local shop search ranking 2026: 38% to 24% cap not a penalty: Local shop search ranking 2026: · Restaurant search rankings: 3 policies, with caps only conditionally admissible: Restaurant search rankings: 3 policies,