How AI assistants find and choose their sources

AI assistants answer from two places: what the model learned in training, and pages or listings it looks up while answering. Nobody outside the AI companies knows the exact selection rules, but the outline is clear enough to act on.

Two broad modes

From training. A model learns general patterns and facts from large amounts of text before release. This knowledge has a cut-off and cannot reflect last week's change to your prices.

From live retrieval. Many assistants can search the web or fetch pages while answering, then summarize what they find. This is where a current, readable website helps most.

What retrieval tends to reward

We have no insider knowledge, but these things plainly help. Pages a crawler can fetch and read. Clear, consistent facts. Content that answers the question directly. A verified listing that shows who you are. A reputation across independent sources.

A small example

A customer asks an assistant for a local roofer. If it searches live, it may read a business website, a Google Business Profile and a few directories. If your site blocks crawlers, your listing is incomplete and your phone number differs across sites, there is little reliable to quote. If all three are clear and consistent, there is more for it to use. Nobody can predict the exact answer, but you can remove the obvious obstacles.

What you can and cannot influence

• You can: allow crawlers, serve readable pages, add structured data, keep facts consistent, publish helpful answers, maintain listings.
• You cannot: control the wording of an answer, force a recommendation, or guarantee inclusion.

Why wrong answers happen

Assistants can state false or outdated things with total confidence. This is called hallucination. When your own information is clear and consistent, it is easier to quote correctly. See AI hallucinations and what to do about them.

What to do next

Start with what is AEO and run the monthly visibility test to see how you currently appear.

What a crawler can and cannot see

Whether an assistant looks things up live or not, it can only use what it can read. Crawlers fetch the raw page. Text in plain HTML is easy for them. Text that only appears after scripts run, text baked into images and content behind a login or pop-up can be missed. Many crawlers also follow the permissions in your robots.txt file, so a rule written for another purpose can shut them out without anyone noticing. Our guide to robots.txt for AI crawlers lists the common ones.

A practical test: open your page, turn scripts off or view the page source, and see whether your name, services, area and phone number are still there as text. If they are not, that is the first thing to fix. For the full picture of the discipline, start with what is AEO.

What happens when sources disagree

An assistant that reads several sources about your business may find that they disagree: one directory shows an old phone number, your site says you serve five towns while your listing says three, an old article mentions a service you dropped. It has to decide what to say. Sometimes it picks one, sometimes it blends them, and sometimes it hedges. You cannot see which, which is why contradictions are risky.

The remedy is dull but effective. Choose one set of correct facts, update the sources you control, and ask directories you do not control to correct theirs. If an assistant has already said something wrong, record it, fix the underlying sources and check again later. See AI hallucinations and what to do about them.

What makes a source look trustworthy

No one outside the AI companies can give you the exact formula, so be skeptical of anyone who claims to. The general signals are the ones that already help people trust a business. A site with real contact details and a named owner. Content that is specific rather than vague. A verified listing. Reviews written by real customers over time, and sensible replies to them. Mentions on other respected sites. Consistency between all of the above.

You can build most of these honestly. You can also damage them quickly with fake reviews or manufactured mentions, which tend to be spotted and are not worth the risk. See automated review requests done right and what to do when a bad review lands.

A worked example (hypothetical)

Picture a small pest control company. A customer asks an assistant, "Who handles termite inspections near me, and do they offer free quotes?" The assistant searches and reads three things: the company's website, its Google Business Profile and a local directory.

The website says inspections are available but does not mention quotes. The profile lists weekday hours only. The directory has an old phone number and no mention of termites at all. With that material, the assistant might name the company with the wrong number, or skip it because nothing clearly answers the question.

The owner fixes it in an afternoon. The service page now says plainly that termite inspections are offered and how quotes work. The profile hours and services match the site. The directory entry is corrected. A short FAQ answers the free-quote question truthfully. The owner cannot know whether the assistant will now recommend the company, but the next answer has accurate, consistent material to draw on. This example is invented to show the reasoning.

Common mistakes owners make

• Assuming the assistant knows you. A model's training has a cut-off, and small businesses are easy to miss.
• Hiding facts in images, PDFs or scripts. Put them in plain text.
• Leaving old listings alone. Out-of-date directories keep feeding old facts into answers.
• Overreacting to a single answer. Answers vary by wording and day. Look for patterns.
• Trying to trick the system. Hidden text and planted instructions are poor practice. See good vs bad AI practices.
• Never checking. Without a regular test you only find out about problems when a customer mentions them.

A quick checklist

• Key facts appear as plain text on your site.
• Robots.txt does not block crawlers you want.
• Your Google Business Profile is verified and matches your site. See Google Business Profile and AI answers.
• Directory listings use the same name, phone and area.
• Your main pages answer real questions directly, with structured data that matches.
• You re-test a fixed set of questions each month. See how to measure your AI visibility.

What to do this week

• Ask two or three assistants about your business and your main service. Write down what each says.
• Mark every fact that is wrong, missing or out of date, and where it probably came from.
• Fix the sources you control first, then request corrections from directories.
• Put a reminder on your calendar to repeat the test next month.

If you would like help tracing where a wrong answer came from, use the contact page, or text or call (904) 293-3040.

Common questions

Want help putting this into practice? See our AEO, SEO and UGC service or request a free read. You can also text or call (904) 293-3040.

This article is general information for small-business owners, not legal, financial, medical or professional advice. Rules and platform policies change; confirm anything that matters with a qualified professional.

Keep reading

• AEO

  What Is AEO? Answer Engine Optimization Explained

  3 min read
• AEO

  robots.txt for AI Crawlers: GPTBot, ClaudeBot and More

  2 min read
• AEO

  How to Measure Your AI Visibility (Without a Score)

  2 min read