Verifiable.
Not necessarily
discoverable.

An answer to “Does this business offer this service?” starts with a named target. An answer to “Which businesses offer this service in this area?” does not. These illustrative question templates test different things; neither is an exact quotation from a collected answer.

BoostGeo studied 100 businesses using a frozen discovery question and a frozen named-service question for each. Each question was run twice on three named surfaces, producing 1,200 preserved observations. The independent business sample is 100—not 1,200.

The useful distinction in this pilot is between discovery, citation and support for a specific fact. A successful named-business answer should not be treated as evidence that the business will also appear when its name is absent from the question.

• Unbranded target mention: the preserved answer names the frozen target business. Mention does not automatically mean endorsement, recommendation or a high position.
• Unbranded official citation: an actually used link to the target’s official site or a documented official alias supports a nearby target-related proposition. A search-result inventory or unrelated link does not qualify.
• Named-service fact support: the answer and retained evidence support the specific advertised service proposition asked about. This is not a score for whole-answer accuracy, professional qualifications, treatment effectiveness or business quality.
• Named-business official citation: at least one official-site citation supports a nearby target-related proposition. A qualified answer may have a valid citation without establishing the primary service proposition. “Not established” does not mean “false.”

The study used a purposive quota design: 20 national/high-prominence businesses, 20 regional/multi-location businesses and 60 local-target businesses. These groups are shown before the combined cohort. Their questions, industries and geographies differ, so differences are not controlled estimates of the effect of prominence.

The local-target group spans Philadelphia, San Diego, Denver, Charlotte, Nashville and Portland, Oregon. A local-target assignment describes the study’s question context; it does not guarantee the company operates in only one market. Later identity and service-area concerns remain disclosed, not silently repaired by replacing businesses.

Across all three tested surfaces, support for the named service proposition was more frequent than an unbranded mention of the target. The two columns describe different tasks—not an accuracy contest between question types or a causal benefit of naming a business.

The surfaces used different models and collection dates. These figures are observations of the recorded configurations, not a ranking of providers, a test of every current model, or an estimate for all users.

Agreement counts businesses whose two runs have matching non-null labels. Matching absence and matching “not established” count as agreement; two unknown labels do not. Agreement is not proof of correctness or stability across weeks.

Treat “Can it check a fact about my business?” and “Does my business appear without being named?” as separate questions. Record both outcomes instead of allowing a strong result on one to stand in for the other.

Keep the question wording, location, surface, model label, date and answer when testing. Check that a cited page belongs to the intended organization and supports the claim beside it. Repeat the same question to observe variation; do not keep only the answer you prefer.

This pilot does not demonstrate that a particular website edit, markup format or BoostGeo feature increases citations. Website Readiness measurements and observed AI-answer visibility should remain separate evidence.

The original collection-time website measurements were unavailable for this analysis. On September 9, the frozen 60 local targets received a separately dated assessment using unchanged SIGIL/λ-12 in an isolated local Worker with local Chromium and paid probes disabled. These are not recovered September 6–7 page snapshots or production-edge measurements.

The batch returned 54 numeric scores. Four assessments were unable to retrieve representative evidence; two URLs were not attempted after prior non-retryable tool safety blocks. No missing value became zero. These outcomes describe collection limitations, not judgments that the businesses are unsafe or invalid.

Saved-evidence screening retained 47 businesses for the main exploratory comparison. Seven measured cases had identity, changed-page, location or purpose-classification concerns. Their original scores remain in an all-numeric sensitivity view. The screening was AI-assisted and recorded before executing the join, but after earlier outcome review—not blinded or preregistered before outcomes were known.

The screened group’s unbranded official-site citation rates do not show a consistent higher-score/more-citations pattern across all three surfaces. That is not proof that readiness has no effect, that low scores help, or that website diagnostics have no value.

Removing limited-confidence scores leaves 44 screened businesses, including only nine in the highest score band. The higher-versus-lower citation advantage seen for the API and Perplexity in the main view no longer appears in this sensitivity view. Small cells limit interpretation; this is not a population estimate.

All seven held cases fall in the middle band, so including them changes that group without changing the low/high groups. This sensitivity preserves questionable measurements for inspection; it does not certify their identity, location or classification.

Other outcomes are not identical to citation outcomes: in the main screened comparison, Perplexity target mentions are 8/34 (23.5%) in the lowest band and 11/22 (50.0%) in the highest. Named-service fact support is lower in the highest than lowest band on all three surfaces. Neither pattern identifies the effect of improving a website.

No intervention was applied, no confounding adjustment was made for industry, geography or prominence, and the scores were measured after the answers. This supplement cannot substantiate a promise that raising a Readiness score increases citations—or the opposite claim that readiness has nothing to do with citation. A causal question needs a separate prospective design with contemporaneous measurements, repeated outcomes and an appropriate comparison group.

The cohort and question register were frozen before the full answer collection; the operational dry run is excluded. Each target received an unbranded discovery question without supplied target-site content and a named-business question about one service. The protocol called for new temporary conversations for consumer repetitions, without prior turns, memory or custom instructions.

Consumer ChatGPT Search: 400 records dated September 6, 2026 UTC, signed-in Pro, labeled “GPT-6, medium thinking.” Collection combined 351 human-operated captures and 49 browser-controlled captures through the consumer interface. They are disclosed as different operator modes, not different independent samples.

OpenAI API with web search: 400 records dated September 7 UTC, model identifier gpt-5.4-mini-2026-03-17. Perplexity Sonar: 400 records dated September 7 UTC, model label sonar. All preserved records report en-US and search invoked. Recorded configuration is not proof of identical search behavior or a verified physical search location.

Review was AI-assisted. All 600 named-service facts have individual review coverage. Discovery citation assessment combines 158 individual assessments with 442 deterministic checks showing no official link in the preserved answer. Discovery mention/role labels received targeted corrections, not exhaustive second individual review.

Some review passes exposed first decisions or selection context. This is not independent human double review or fully blind review. Source checks performed on September 9 are supplemental, not collection-time snapshots. Consumer transcripts have no provider finish metadata; preserved nonempty text alone does not independently prove complete interface capture.

This selected 100-business cohort is not a representative sample of businesses worldwide, all ChatGPT users, or any national business population. Repetitions increase observations, not independent businesses. Industry cells are not used for winner/loser rankings.

Six citation-support outcomes remain unknown: one unbranded and five named-business observations. They remain in the displayed denominators. A failed source check is not evidence that a business lacks the service.

One consumer named-business capture exactly duplicates a discovery response and is flagged as a possible prompt/paste mismatch, not counted as a supported primary fact. It remains in the main frozen dataset. Excluding it changes ChatGPT named-service fact support from 185/200 (92.5%) to 185/199 (93.0%); the original response is preserved and no replacement was invented.

Additional sensitivities cover a frozen name/domain mismatch, temporal branding changes, a parent/affiliate boundary, strict entity matching and unresolved citations. They are retained in the evidence package; the principal sensitivities and limitations are reported here. They do not turn this sample into a controlled experiment.

An official-site citation establishes at most a narrow nearby proposition under this rubric. It does not certify every sentence, an organization’s independence, professional credentials, treatment effectiveness or the quality of a recommendation.

BoostGeo designed and conducted this pilot and sells website-analysis and AI-visibility software. That commercial interest should be considered when interpreting the report. The study is not evidence that using BoostGeo improves citations, traffic or business outcomes.

This report publishes aggregate findings, chart data and methodology without named-business scores, rankings, raw answers or source URLs. Raw captures, frozen registers, hashes and reviewer decisions are retained internally for audit; they are not included in public downloads to avoid identifiable business-outcome joins and unnecessary reproduction of third-party content.

Published September 10, 2026. BoostGeo is the editorial and corrections owner. Collection and AI-assisted evidence review are described above; this is not a claim of peer review or independent human double coding. Version 1 retains the questionable capture in the main denominator and reports the exclusion sensitivity. Material corrections will be dated here.