Skip to content

How do AI search tools choose sources?

Listen to Story · 6 min listen
Quick Takeaway · Key Summary

A reader asks an AI search feature a question and receives a few paragraphs with links. Ask the same question on another day or in another service and the cited pages may differ. Can a publisher know in advance which source will be chosen?

Not completely. The full selection rules of every system are not public, and results can vary with the product, query and conditions. A linked source does not by itself prove every statement in an answer is correct. Provider documentation can still show publishers where to begin.

What does Google say about its own search?

Google's generative Search guidance describes how its features use underlying search and quality systems to find relevant pages. Indexing, eligibility for a search snippet and the site's inclusion setting for the relevant features matter for eligibility. Eligibility is not a guarantee of selection.

Those statements concern Google's own services. They do not establish that another company evaluates the same page in the same way or chooses the same sources.

One question can lead to several searches

When a listener asks, “Why does internet radio keep cutting out?”, the answer could involve the connection, broadcast software or the listener’s device. Google’s generative Search guidance describes an approach called *query fan-out*: a system can run related searches in parallel to gather additional information. The radio question is an illustration; we have not observed the actual searches Google would generate for it.

This does not mean a publisher needs a separate page for every wording. An article about connection failures can cover the genuinely different possibilities a reader needs; a deeper account of a broadcast chain may justify its own article when it serves a distinct purpose. Google’s description explains a general approach, not which sub-search will retrieve a particular page or whether a response will cite it.

What can a publisher improve?

Start by checking whether an important page is accessible, crawlable and eligible for indexing under the relevant conditions. Then return to the reader's question: does the page answer it clearly? Are important claims sourced? If it includes a genuine interview, method or first-hand experience, can the reader understand what was actually done?

📻
RadioTEDU Live Stream

Listen while exploring this story

Tune into campus radio with non-stop independent music and culture.

Google's guidance says a special AI file or special type of structured data is not required to appear in its generative Search features. A third-party technical score does not directly measure how often real answers cite a site either. Technical access, content value and observed citations are different questions.

How should a small sample be read?

Imagine checking whether RadioTEDU appears for “How does internet radio work?” and its Turkish equivalent. Record the exact prompt, platform, date, language and, where known, country and session conditions. Keep a mention of the RadioTEDU name separate from a link to radiotedu.com, and record links to third-party pages separately.

One response does not represent every user. Answers can change, so record how many times you repeat the query. If you cannot obtain an answer, mark the attempt unmeasured rather than absent.

What can Search Console show?

Suppose a RadioTEDU article starts appearing more often in Google’s AI features. A handful of manually checked prompts cannot tell the whole story. With access, Google’s generative AI performance report lets eligible properties with enough data review link impressions in AI Overviews and AI Mode by page, country, device and date. An impression counts a displayed link; it is neither a click nor a citation in another provider’s service.

Imagine one answer displays links to both RadioTEDU’s signal-chain article and its interruption guide. Google says the chart normally aggregates by property while the page table aggregates by page, so their totals can differ. Report dates use Pacific Time and exclude Search Labs experiments. If an export shows zero, return to the interface: Google notes that unavailable values shown there as “~” or “-” become zeros in the downloaded data.

A missing report can reflect insufficient impressions or a feature setting. According to Google’s feature-control documentation, a child property can inherit its parent’s choice; an authorised user can check the current setting in Search Console. Check your own property’s setting and report availability in your account.

Rather than reducing source selection to one score, consider access, editorial contribution and dated observations separately. That makes the limits of each piece of evidence easier to see.

RadioTEDU monthly

The month in podcasts.

Receive podcast episodes from the latest 30 days and upcoming RadioTEDU ticketed events. No daily mail or unrelated announcements.

This consent covers only the monthly podcast newsletter and ticketed event updates. See our privacy notice.

Live RadioTEDU Pick a station and start listening
0:00 0:00

RadioTEDU · Big5 Türkiye Watch Party

RadioTEDU Live Screening: Big5 Türkiye Watch Party. Thursday 19 November, 18.00, D032. Click for free tickets.