How ChatGPT and Google AI pick their sources (and why they sometimes get you wrong)

Ask ChatGPT and Google’s AI the same buying question and you will often get two different lists of companies, each with citations. Neither is random. Both engines run a pipeline that can be described in five steps, and once you can see the pipeline, the question “why did it pick them and not me” usually answers itself.

This article describes the mechanism as the engine companies have documented it and as it is observable from the outside. It is written for business owners, not engineers; the goal is to know which parts of your own web presence the pipeline actually touches.

The five steps every answer engine runs

Step	What happens	What it touches on your side
1. Retrieve	The question is rewritten into one or more search queries and run against a web index	Whether your pages are indexed and rank for those queries
2. Read	The top results are fetched and parsed, usually five to ten pages	Whether the page’s facts are in HTML text, not JavaScript or PDF
3. Select	Passages that most directly answer the question are extracted and scored	Whether a sentence on your page states the answer plainly
4. Synthesize	One answer is written from the selected passages, resolving conflicts between them	Whether your facts agree with other sources about you
5. Cite	The pages behind the selected passages are attached as sources	Whether the passage came from your page or from a page about you

Everything a business can do about AI visibility maps onto one of those five rows. Nothing else in the process is reachable from outside.

What each engine actually searches

Engine	Index it searches	Its crawler tokens	Notes
ChatGPT (search on)	Bing, plus OpenAI’s own fetches	OAI-SearchBot, ChatGPT-User	Bing Webmaster Tools and IndexNow are how you get in
Google AI Overviews	Google	Googlebot	Built from core ranking; fans the question out into sub-queries
Google AI Mode	Google	Googlebot	Broader fan-out, more sub-queries per question
Gemini app	Google (Search grounding)	Googlebot	Same index, more conversational selection
Microsoft Copilot	Bing	Bingbot	Same pool as ChatGPT search
Perplexity	Its own index	PerplexityBot	Also fetches live for many answers

The practical consequence: there are two indexes that matter in North America, Google’s and Bing’s. A site that is well indexed by one and ignored by the other is visible to half the engines. Verifying in both webmaster consoles and turning on IndexNow closes that gap at no cost.

Step one: retrieval decides the pool

An engine cannot cite a page it never retrieved. Retrieval is ordinary search: the same ranking systems that produce the ten links, run on queries the engine wrote for itself. Google’s AI products call the rewriting “query fan-out”: a single buyer question becomes several narrower searches, and the top results of each form the reading pool.

This is why generic pages rarely get cited. A page that ranks for a broad phrase but for none of the narrower sub-questions is not in the pool when the engine looks. It is also why a page can be cited for a question it never targeted: it happened to rank for one of the sub-queries.

Step two: reading decides what the engine can see

The engine reads pages the way a fast, literal reader does. Text that is present in the HTML is read. Text that appears only after JavaScript runs may not be. Facts inside images or PDF brochures are usually invisible. Pages that return an error, a challenge screen or a login wall to the engine’s crawler contribute nothing.

Structured data is read alongside the text. A LocalBusiness or Organization block that states the name, address, service area and services gives the engine a second, unambiguous copy of the facts.

Step three: selection decides who gets quoted

Selection works at the level of passages, not pages. The engine is looking for the sentence or table that most directly answers the sub-question. Several things make a passage win:

Directness. “Standard lead time is two to four weeks; rush orders in five business days” answers a lead-time question. “We pride ourselves on fast turnaround” does not.
Specificity. Numbers, ranges, place names, model names, certifications. Specific claims are quotable; vague ones are not.
Position. The answer near the top of the page, under a heading that matches the question, is found more reliably than the same sentence buried in paragraph nine.
Recency. Dated pages with recent updates are preferred for questions where things change, such as prices and availability.
Format. Tables and short lists are lifted almost verbatim. Long prose is summarized, and the summary may drop you.

Step four: synthesis decides whose facts survive

When the selected passages disagree, the engine has to pick. It tends to side with the version that appears in more places and in more authoritative places. This is where consistency stops being a tidiness issue and becomes a visibility issue: if your site, your Google Business Profile and a directory give three different service areas, the engine may drop the claim entirely or pick the version you like least.

Synthesis is also where third-party pages earn their weight. A claim on your own site is one voice. The same claim on an association member page or in a news story is a second voice, and for a system resolving conflicts, two voices beat one.

Step five: citation decides who gets the credit

The sources shown are the pages whose passages made it into the answer. If the engine took the fact about you from a directory page, the directory is cited, not you. That is still a mention, and often still a lead, but it means the buyer’s click goes elsewhere. The way to be the cited source is to be the page that states the fact best, and to be in the pool when the engine retrieves.

Why answers are sometimes confidently wrong

“Why does the AI lie about my business” is one of the most common questions we hear, and the honest answer is that it rarely lies; it inherits. Four failure modes account for nearly all of it:

Thin pool. The sub-question retrieved only one or two pages, and they were wrong or outdated. The engine had nothing better.
Stale sources. A page the engine trusts describes your old address, your discontinued service or your price from two years ago.
Entity merge. Another business shares part of your name, and the engine blended the two. Common for two-word names, franchise names and any name that is also a common word.
Wrong genre. A forum thread, a satirical post or a competitor’s comparison page ranked for the sub-question and was read as fact.

None of these is fixed by complaining to the engine. They are fixed by making the correct fact the easiest one to retrieve, read, select and corroborate: update the stale page, state the full legal name and location together, list every spelling of your name in structured data and llms.txt, and get a second source to say the same thing.

Memory versus search

Everything above describes an engine that searched before answering. When no search happens, the model answers from what it learned in training. For most businesses that memory is empty, and no amount of on-site work changes it quickly; training data is refreshed on the scale of model generations, and what gets in is what has been said about you widely and for a long time. It is worth knowing which mode produced an answer before drawing conclusions from it. A search-backed answer reflects your web presence this month; a memory answer reflects your reputation over years.

How to become a source, in one list

Be indexed by Google and Bing, and allow the AI crawlers in robots.txt and at the CDN.
For each buyer question, have one page that ranks for its sub-questions and states the answer in the first screen, in text.
Prefer numbers, ranges, places and tables over adjectives.
Make name, address, phone, category and service area identical across your site, structured data, Google Business Profile, Bing Places and directories.
Get at least one page you do not control to describe you accurately.
Give the engine enough to tell you apart from anyone with a similar name.

Where to start

If you want to see which sources the engines currently use for your buyer questions, and whether you are among them, we measure it: twenty real questions across Gemini and ChatGPT, every source recorded, a PDF back within two business days. It is free.

Questions people ask

How does ChatGPT decide which websites to cite?

With search enabled, ChatGPT rewrites the question into search queries, retrieves results from Bing plus pages fetched by OpenAI’s own crawler, reads the top results, and cites the pages whose passages it used in the answer. Pages that rank for the query and state the answer in plain text are far more likely to be cited than pages that merely mention the topic.

Why does Google’s AI Overview sometimes give wrong information?

Because it synthesizes from whatever ranks for the question and its sub-questions. If those pages are thin, outdated, contradictory, satirical or about a different entity with a similar name, the summary inherits the error. The fix, for a business, is to make the correct fact easy to find and consistent everywhere the engine looks.

Do AI engines prefer big websites?

They prefer pages that rank and that answer directly. Large publishers often win generic questions because they rank for them. For specific, place-bound questions, a small business page that states the answer plainly is routinely cited ahead of a large site that only mentions the topic.

Does ChatGPT use Google?

No. ChatGPT’s web search draws on Bing’s index and OpenAI’s own crawling. Google’s AI Overviews, AI Mode and the Gemini app draw on Google’s index. To be visible in both, a site needs to be indexed by both Google and Bing.

( WHERE YOU STAND )

Does AI recommend you?

We put twenty real buyer questions about your category and market to Gemini and ChatGPT and send you a PDF showing who they cite instead of you, plus fourteen technical checks. Free, no call required.

Request the free audit