How AI search finds local businesses: websites, third-party sources and MCAC

1. Research questions: appearance is not the whole outcome

The familiar search journey—results, click, website evaluation and conversion—is one possible route rather than a complete account of information discovery. Generated answers can provide descriptions, phone numbers or directions before someone opens an official site. A brand may be named through a review, cited without being named, or named with inaccurate details.

The motivating question is practical: if a long-established local restaurant can be recommended from maps, reviews and food writing, what additional function does its own website provide? This article examines four questions separately.

“No confirmed independent domain” is a research classification: failure to confirm one does not prove that a website does not exist. The literature review is a question-led narrative synthesis, not a systematic review with exhaustive database searches, independent double screening or meta-analysis. No preregistration of this pilot was found.

2. AI answers, zero-click searches and referral visits use different denominators

An answer can satisfy a question without sending a visit, while other answers direct people to a site. Search results with an AI summary and a conversation started inside an assistant are also different entry points. The following historical studies should not be joined into one conversion funnel.

Pew’s 15% and 8% differ by seven percentage points; the relative difference is not a controlled causal effect of an overview. Zero-click can also include another search, an abandoned task or leaving the browser. Fewer clicks do not on their own prove better information satisfaction.

Comparisons across the 2024 and 2026 zero-click reports are further limited by different panel providers and methods. One later article retrospectively cites a different 2024 number; LS-Labs preserves the original report’s 58.5% US figure rather than manufacturing a single consistent time series.

Wang and colleagues’ 2026 preprint adds a different kind of evidence: a preregistered field experiment with 1,100 participants and seven-day exposure to three browser-interface conditions. Removing AI features increased publisher referrals; an AI-Mode-only experience lowered clicking and some experience and trust measures. Its forced interface and recruited sample still differ from ordinary long-term use.

The strongest counterpoint to a simple traffic-loss story is that AI referrals can grow from a small base and arriving visitors may differ in intent. Google’s statement that some overview-origin visitors spend longer is a platform observation, not proof of additional Taiwan restaurant bookings. Visits, calls, reservations and sales should be measured separately.

3. More machine activity does not mean all traffic is AI shopping

The original report records a Cloudflare Radar global HTML-request snapshot at 14:30 UTC on 8 June 2026: Bot 57.2%, Human 42.8%, alongside Matthew Prince’s 3 June statement. These are classifications of requests in that view, not people and not a measure of all internet activity attributable to AI.

HUMAN Security’s 2026 industry report separately describes 7,851% year-on-year growth in agent and agentic-browser traffic, concentrated in areas including retail, media and travel. Its observed environment and classifications belong to that provider. Neither statistic establishes completed purchases or demonstrates that human visits will disappear.

Agents may retrieve information, operate tools or assist a transaction. That makes reliable specifications and actionable endpoints relevant, while adoption, permission and execution remain separate questions. The commerce readiness guide examines those limits in more detail.

4. A digital entity can be described by more than its own URL

In this article, “digital entity” means the business represented across names, addresses, phones, menus, reviews, media and managed platform pages. Google Business Profile documentation lists public websites, licensed third-party data, user submissions, merchant submissions and other location interactions among its information sources.

That makes it plausible for a restaurant without a confirmed independent site to leave a substantial public footprint. It does not establish which exact data source a particular Sonar answer used. Mentioning a restaurant is not evidence that the model necessarily consulted its Google Business Profile.

Retrieval-augmented generation, as introduced by Lewis and colleagues, combines parametric knowledge with retrievable information. Contemporary commercial products are not one disclosed universal pipeline. Google describes query fan-out; ChatGPT Search can rewrite queries. A fixed number of retrieved pages or selected citations should not be asserted as a cross-platform rule.

The website as a maintainable digital anchor

A digital anchor is an official information node a business can update and others can inspect. Its value is control over publication and a stable address, not control over an AI system’s decision to fetch, use, cite or accurately reproduce it.

A stale first-party page may be less useful than a current third-party source. Source ownership and factual reliability must be checked separately. For a small walk-in business whose main information is an address and phone number, maintained platform listings may meet much of the need. A website’s extra value depends on the functions and upkeep the business actually requires.

Why a mention count is insufficient

Semrush’s 2026 “ghost citations” study reports 3,981 domain-appearance events across 115 prompts, 14 countries and four AI interfaces: 61.7% had a source link without a brand mention, 25.1% a mention without a link, and 13.2% both. The denominator is domain appearances in that external study, not all answers or restaurant results in this pilot.

Source lists, adjacent citation markers, claim support and external truth are also distinct. An API returning a URL does not prove that every statement used it, nor that its passage supports the answer. That distinction is central to interpreting the following results.

5. MCAC separates four checks

MCAC is LS-Labs’ operational framework: Mention, Citation, Accuracy and Conversion-readiness. The historical records called the fourth dimension Actionability. The public label was changed to align the acronym; the pilot still measures actionable information, not conversion rates. This is not a validated universal scale or a statistical demonstration that its dimensions are independent.

The GEO guide illustrates these four checks with one restaurant answer. Read it alongside this study’s sample and methods to distinguish observations from improvement hypotheses that still need testing.

Each dimension points to a different investigation: missing topic relevance, dependence on third-party information, a conflicting fact or an absent next step. Compressing them into one GEO score can hide which problem needs attention. See the measurement methods guide for denominators and missing-value reporting.

6. Six Taipei restaurants, four task types

The pilot purposively selected Taipei beef-noodle restaurants and classified three with confirmed independent official domains and three without a confirmed independent domain. It did not match groups on reputation, review volume, media exposure, prices or branch count, and it did not verify legal domain ownership through merchant administration.

During collection, Yong Kang Beef Noodles was reclassified after an independent official domain was found. Its three Fact, Action and Comparison records were excluded before final group analysis and Fu Hong was introduced. This was a documented collection-stage change, not an untouched preregistered sample. The excluded records remain preserved.

The 18 brand-by-run opportunities in three Discovery answers share those answers and are not 18 independent participants. Fact and Action have only one answer per restaurant. The prompt context is Chinese and Taipei; there is no evidence of explicit latitude/longitude configuration.

Retained timestamps span 16 September 2026, 14:57:32–15:07:02 UTC, or 22:57:32–23:07:02 in Taipei. Scripts recorded time at different stages, so this is the range of saved timestamps rather than exact start and finish times for every request.

The recorded model label is sonar, without an immutable version. Scripts requested temperature 0.2 and max_tokens 900. They requested a top-level medium search-context setting, while all 21 usage records reported low. LS-Labs discloses that discrepancy rather than assuming medium took effect. The saved data are selected API fields, not complete HTTP archives; they lack finish_reason, headers and backend query traces. Cache behaviour and exact retrieval paths are unknown.

7. Results: retained mentions, reference agreement and URLs

Discovery answers have a material completeness limitation

The retained text in all three Discovery answers includes Chef Hung, Tian Xia San Jue, Niu Dian, Lin Dong Fang and Liu Shan Dong, but not the registered Fu Hong name or alias. Chef Hung appears as a brand rather than specifically the Jianbei branch. A different restaurant name was not automatically treated as Fu Hong.

All three endings contain unfinished passages; the first stops within a recommendation table. Without finish_reason, the cause cannot be established. Positive appearances can be reported in retained text, but a missing name means only “not observed in the retained part.” Under a strict complete-answer-only rule, no Discovery answer here has confirmed normal completion from which to estimate a stable mention probability.

The three Discovery citation arrays contain the same 20 URLs in the same order. That is evidence of short-window repetition, not proof of caching or stability over time.

Reference agreement is not comprehensive truth verification

Each of six Fact answers supplies an address and phone matching the adopted reference table. Some references use Maps search links without a contemporaneous page snapshot or direct merchant confirmation, and may overlap with the model’s own information sources. Prices, menus, dietary restrictions and opening times were not comprehensively verified.

Tian Xia San Jue’s Fact answer lacked inline citation numbers despite an API source list. For Chef Hung, an official-source marker supported the address line while the phone line used other sources. “The source list contains an official domain” must not become “every correct fact came from the official site.”

A phone number can reach the business directly

All Action answers provide a phone number; the confirmed-domain group also lists registered official URLs. Whether the phone was answered, the restaurant was open, a reservation could be made or a purchase occurred was not tested. The late-night “today” wording did not supply the model with a local clock time.

The original ownership label could confuse information provenance with action destination. A phone copied from a food blog can still call the restaurant directly, while an official website can redirect to an external booking platform. An Oddle subdomain labelled brand-owned by the model was not independently established as wholly first-party. The article retains the observed URL difference rather than claiming the other group had no direct action route.

8. Recounting citations with matched numerators and denominators

Overall URL uniqueness is computed across all records, not by adding task-level unique counts. The six official URLs are Chef Hung’s homepage and /pages/h; Niu Dian’s /niu-dian/; and Noodle3’s homepage, /about and /news. The other 200 unique URLs are not automatically all unofficial or user-generated: the official classification here covers only the three registered sample domains.

The two valid shares are 6/206 = 2.91% of literal unique URLs and 10/408 = 2.45% of source-list entries. The historical figure’s 2/206 = 0.97% mixed the Fact-only numerator with an all-task denominator and is not used. The original five-tier source classification lacks reproducible per-URL coding rules, so the former UGC percentage is not retained as a research result.

Literal URL counting does not merge redirects, language variants or related paths. The 112 hostnames remove leading www. only; they are not 112 independent publishers or legal organisations. Source lists are also not model weights, trust scores or verified sentence-level contributions.

9. Reversing restaurant order exposed a measurement failure

Reversing the order moved the missing information to the other group. The same businesses could supply address and phone in single-business Fact questions, yet those fields became missing in second position. Reading only the original order could have mislabelled a prompt-related difference as a website benefit.

This is position-related missingness, not proof that the mechanism is solely language-model positional bias. Query rewriting, retrieval coverage, context allocation, generation and sequential timing may all contribute; the saved records do not isolate the stages. The six paired answers are retained as a measurement finding, not as six negative business-performance observations.

Even first-position completeness does not establish consistency of every detail. Some answers expressed conflicting phone or closing-time information. Those conflicts are not resolved by the separate 12/12 Fact reference-agreement result. Report completeness and correctness separately, and do not silently turn UNKNOWN into a wrong real-world value.

10. What the study supports for website decisions

The clearest finding is functional separation. Public third-party information can support a business mention. An official domain can provide a maintained first-party reference and another destination. Neither result establishes the average commercial return on building a website.

The strongest counterargument is a small walk-in business adequately served by maintained profiles, a phone and maps. Its website’s marginal value may be limited unless it adds useful information or operations. More complex multi-branch brands may need controlled service conditions, reservations, ordering, CRM and consistent updates. Those are decision hypotheses to validate against costs and outcomes, not estimated effects from six restaurants.

Uncontrolled reputation, reviews and media coverage can influence both whether a restaurant maintains a site and whether a search system mentions it. Equal address-and-phone results in one small sample do not prove websites have no effect on accuracy. Conversely, posting a correction on an official site does not establish that AI answers will update faster; adoption delay was not tested.

Use the framework to choose the next investigation. Missing mentions, third-party source dependence, a wrong branch fact and an unusable action endpoint require different responses. For website referral outcomes, use the GA4 guide; for a current operational report, distinguish tool capabilities from what this historical pilot actually measured.

11. Limitations, data availability and disclosure

This is one city, one restaurant category, one API label and one short collection window. The sample is purposive, unmatched and changed during collection. Discovery completion is uncertain; Fact and Action observations are single runs. No consumer-interface, cross-model, longitudinal or content-intervention replication is claimed. These limitations prevent causal, population-level and revenue inference.

The workspace preserves sample definitions, reference tables, 24 extracted API records, exclusions, analysis and a manifest. The 17 September reanalysis verified record hashes and recomputed counts, including the URL register. That confirms consistency of the retained package, not the external truth of every provider value or timestamp. Count reproducibility is not reproducibility of the exact model answers.

The public article supplies methods, sample scope and results but not a downloadable independently recomputable raw dataset. Local evidence paths are not public archives. Data-access questions can use the LS-Labs contact route; do not treat a prose description as open-data publication.

The author participates in LS-Labs research and tool development, which may influence questions, tool choice and interpretation. The available records do not support a complete external funding declaration, so the absence of such a declaration is not represented as absence of interests. AI assisted literature search, code-based recounting, evidence comparison and language editing; that does not establish full manual verification or external peer review.

An academic preprint of this study was published on Zenodo on 30 October 2026, with English and Traditional Chinese PDFs. Cite as: LI-Sheng (李昇). (2026). Retrieval Mechanisms of Local Businesses in the Generative Search Era: Third-Party Footprints, First-Party Anchors, and the MCAC Empirical Evaluation Framework [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23773956

References

• Aggarwal et al. (2024), GEO: Generative Engine Optimization.
• Chapekis and Lieb (2025), Pew AI-summary click study.
• Fishkin (2024), SparkToro / Datos zero-click study; Fishkin (2026), SparkToro / Similarweb analysis.
• Gao et al. (2023), Enabling Large Language Models to Generate Text with Citations; Liu et al. (2023), Evaluating Verifiability in Generative Search Engines.
• Google, AI features and your website, Maps grounding, Local business structured data and Business Profile information sources.
• HUMAN Security (2026), State of AI Traffic & Cyberthreat Benchmarks; Matthew Prince’s June statement and the dated Cloudflare Radar snapshot described above.
• Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.
• Loktionova (2026), Semrush ghost citations study; McDonald et al. (2025), Seer AI search landscape study.
• Murphy (2026), Local Consumer Review Survey; TWNIC (2025), Taiwan Internet Report. These surveys concern different markets and are not this pilot’s outcome data.
• OpenAI, ChatGPT Search announcement; Perplexity, Sonar API documentation.
• Schulte, Bleeker and Kaufmann (2026), Don’t Measure Once; Shi, Zhu and Gu (2026), Answering Without Referring; Wang et al. (2026), AI in Search Reduces Publisher Referrals Without Improving User Experience. Preprints retain their status.

External studies retain their own methods and observation dates. The restaurant counts belong only to the LS-Labs September 2026 pilot; no updated merchant operating information is provided as a recommendation to visit.