Do AI Assistants Prefer Business Profiles, Websites or Directories for Local Recommendations?
A preregistered protocol for comparing local recommendation sources across ChatGPT Search, Google AI Mode and Maps, Bing or Copilot, and Perplexity.
Research question: When people ask AI-assisted search products for a local recommendation, how often do the visible sources come from first-party business websites, Google Business Profiles or map entities, directories, review platforms, publishers, and other aggregators?
SearchEngineAnswer has not run the controlled sample described here. No universal source preference or ranking factor is claimed. Publishing the protocol first fixes the query set, coding rules, environments, and exclusions before results are visible.
Define what source preference means
A product can name a business without showing a source, link to a first-party site, display a map entity, cite a review page, or ground an answer in several sources. Count these as separate observations.
| Event | Question | Do not infer |
|---|---|---|
| Entity named | Was the business recommended? | Which source caused inclusion |
| Destination link | Where can the user click? | Grounding source |
| Visible citation | Which source is attributed? | Stable ranking |
| Map or profile card | Which platform entity appears? | Website irrelevance |
| Claim support | Does the source support the reason? | Business quality |
Build a balanced query set
Select several markets and local categories with different decision patterns: restaurants, urgent services, regulated professionals, retail inventory, attractions, and home services. Use fixed query templates for “best,” “near me,” open-now, attribute-specific, budget, accessibility, and comparison tasks.
Predeclare locations, language, radius, date, time, device, account state, personalization, and whether precise location is enabled. Keep brand queries separate from discovery queries. Include a negative-control attribute that the local market does not support to test whether products abstain or invent.
Run each platform as a separate instrument
Test ChatGPT Search, Google AI Mode and Maps, Bing or Copilot, and Perplexity only where the products and region are available. Save the exact product surface, version clues, prompt, response, citations, links, map cards, follow-up questions, and screenshots. Do not merge two Google surfaces into one observation.
Repeat each query on a schedule and randomize platform order where practical. A single run cannot distinguish a stable pattern from response variability. Repeated runs still describe the sample, not a permanent index.
Classify source owners and claims
- Resolve every visible URL through redirects to its final owner.
- Classify first-party business site, platform profile, directory, review platform, publisher, government, social network, or other aggregator.
- Separate the page that supports a claim from the page offered as a destination.
- Record whether hours, price, rating, availability, and attributes agree across sources.
- Mark claims with no visible support and sources that do not entail the recommendation reason.
Use two reviewers for ambiguous source ownership and preserve disagreements. A directory owned by a marketplace or a profile embedded in a search product may need a documented compound code.
Report without inventing ranking factors
Report named-entity frequency, destination-link share, visible-citation share, source-owner distribution, support accuracy, disagreement rate, and run-to-run variability. Break results out by product, surface, market, category, and query type.
Do not claim that a source type caused a recommendation. The study observes outputs; it does not expose internal ranking systems. A source can correlate with recommendations because it contains useful facts, has broad coverage, is accessible to the product, or co-occurs with stronger business signals.
Use the local SEO evidence guide and the citation-ready passage test when reviewing claim support.
Ask a question or join the discussion