Written by Tristan Michel
No public formula explains how every AI product chooses every source. Some answers use live web search, some rely on information already available to the model, and some combine several inputs. A local business does not need the secret recipe to investigate a bad answer, but it does need to know what the visible evidence can and cannot prove.
First, separate three kinds of input
An answer can reflect patterns learned during model training, information found through a search or retrieval step, and context supplied in the conversation. The mix can change from one question to the next.
A question about a long-established trade may be answered without a visible web search. A request for today's opening hours or nearby emergency service is more likely to need current information. If the customer has shared a location, uploaded a document, or mentioned a preference earlier in the conversation, that context can also shape the result.
These routes matter because only some leave visible footprints. You may see links from a web search, but you cannot inspect all training material or every internal decision that shaped the wording.
What a citation actually tells you
A citation connects a page with part of the answer shown in that run. OpenAI says ChatGPT responses that use search may include inline citations and a Sources panel. Anthropic documents citations for Claude web search, while Gemini and Perplexity present source information through their own interfaces.
Open the page before drawing a conclusion. A source may support the industry background while another page supplies the business hours. A directory can be cited beside a recommendation even when the company's own website contains the richer service detail. Sometimes the page does not clearly support the sentence beside it.
A citation does not prove that the page caused the recommendation, that it was the only source consulted, or that the same source will appear next time. It is a lead. Treat it like one.
Why a useful page may be chosen
Platforms do not publish a universal checklist for source selection, so claims about exact weights are speculation. Still, the pages that help answer a question tend to have recognizable qualities. They address the need directly, contain specific and current information, make the subject easy to identify, and provide evidence a reader can check.
Consider a request for a licensed electrician who installs EV chargers in Shediac. A generic homepage saying ‘powering your future’ contributes very little. A service page that names EV charger installation, explains panel requirements, identifies the service area, and shows verifiable licensing information is easier for a customer to use. An up-to-date provincial licence register may support the qualification. A local directory with an old phone number may create conflict.
This does not guarantee that any of those pages will be retrieved or cited. It explains why improving the information for a real buyer is a safer investment than adding phrases meant to impress a model.
Search products can split one question into several searches
A generated answer may need more than one piece of evidence. OpenAI says ChatGPT Search can rewrite a question into targeted searches and send follow-up searches. Google describes a related process called query fan-out for its AI search features. A request for a commercial roofer may lead to separate searches about location, flat-roof experience, certifications, and the type of building involved.
That helps explain why a page that ranks for a broad phrase may not appear in a detailed answer. The AI product may be trying to resolve narrower parts of the request. It also explains why creating a separate page for every imaginable wording is a poor response. A strong page can answer the connected questions coherently.
How to review the sources behind an answer
Use a question tied to a genuine buying decision and turn on web search when the product offers that choice. Save the exact wording, date, location conditions, full answer, and every source the interface exposes. Then read the sources as a customer would.
- Which sentence or claim does the page appear to support?
- Is the information current and specific to the question?
- Does the page identify the business, service, and place without ambiguity?
- Are important claims backed by evidence outside the company's own marketing?
- Do conflicting hours, addresses, staff names, or services appear elsewhere?
- Does the source recur when the same useful question is tested again?
Fix the fact, not the citation pattern
If the answer shows the wrong hours and cites an old directory, correct the directory or ask its publisher to do so. If the cited page is accurate but your service is unclear, improve the relevant page on your site. If the source supports a competitor who is genuinely a better fit, copying its headings will not change that.
Do not create fake discussions, buy low-quality mentions, or flood directories in an attempt to imitate a source list. Those actions make the public record less trustworthy and may leave customers sorting through more contradictions.
Repeat the original question after material corrections have had time to be discovered. A new source, a better description, or a more accurate answer is useful movement. It still does not reveal a permanent ranking formula.