Almost every business owner I talk to asks some version of the same question: how does ChatGPT decide who to recommend? From the outside it can feel like a coin flip. AI search follows a series of steps, and most of them can be explained in plain English. Here they are, in order.
- It reads your question and decides whether it needs to look anything up.
- It turns your one question into several smaller searches.
- It pulls pages from a search index, a giant saved catalog of the web.
- It reads those pages and keeps the passages that answer each piece of your question.
- It checks those passages against each other and against what the rest of the internet says.
- It writes one answer in its own words and names a few sources.
Step 1: The AI decides whether it needs to look anything up
Tools like ChatGPT and Gemini are built on what's called a large language model: software trained by reading an enormous amount of text, then learning to predict what words should come next. Everything it learned in training is a snapshot. It has a cutoff date, and it doesn't update when your hours change or you add a new service.
So for a lot of questions, the AI goes and checks the live web before it answers. Questions about local businesses, prices, hours, anything "near me," and anything recent usually trigger that live lookup. A question like "what is a root canal" might not, because the training snapshot already covers it.
This matters for you because the live lookup is where your website is in play. The training snapshot is old, fuzzy, and out of your hands. When the AI searches in real time, it's reading what's on your site today.
Step 2: The AI turns your question into several smaller searches
A real person's question has a lot packed into it. So before searching, the AI breaks it apart. OpenAI says ChatGPT typically rewrites a question into one or more targeted searches; ask for "good restaurants near me" from San Francisco and it may search "top restaurants San Francisco" (OpenAI Help Center, ChatGPT search). Google says its AI Mode breaks a question into subtopics and runs many searches at the same time on your behalf (Google I/O, May 2025). The industry name for this is query fan-out, and I wrote a full explainer on query fan-out if you want the deeper version.
In Coast Guard search and rescue, you never search a big area all at once. You split it into smaller sections and clear each one. AI search handles a big question the same way.
Here's a question a parent might type: "My son is autistic and scared of clippers. Where can I take him for a haircut in Clearwater, FL?" Behind the scenes, that can become searches about sensory-friendly barbershops in Clearwater, barbers with experience cutting autistic kids' hair, and haircuts done with scissors only. When I tested that exact question in ChatGPT, it recommended Sapia's Barbershop, a client whose site has a page written for that parent.
One question in. Several searches out. One answer back.
Step 3: The AI pulls pages from a search index
A search index is a giant saved catalog of web pages. Search companies send out automated programs called crawlers to visit websites, read them, and save what they find. When the AI runs its searches in Step 2, it's searching that catalog. If your page was never saved, the AI can't find it.
Different tools use different catalogs:
- Google AI Overviews and AI Mode use Google's regular index. Google says a page has to be indexed and eligible to show in Google Search with a preview snippet, and there are no extra technical requirements beyond that (Google Search Central, "AI features and your website").
- ChatGPT uses OpenAI's own crawler, called OAI-SearchBot, and sometimes partners with other search providers. OpenAI says sites need to allow OAI-SearchBot to be included in ChatGPT search answers (OpenAI Help Center).
- Perplexity runs its own crawler too, called PerplexityBot.
One technical detail explains a lot of invisible websites. A study of live crawler traffic by Vercel and MERJ found that none of the major AI crawlers ran JavaScript, including OpenAI's and Anthropic's (Vercel, December 2024). JavaScript is code that builds or changes a page after it loads. Google has a second step that runs it. The AI crawlers mostly read only the raw page the server sends first.
That means if your words only appear after JavaScript runs, which is common with some page builders and app builders, an AI crawler can arrive and find a nearly empty page. That's exactly what was happening to BrightWater Pool Co. before I rebuilt their site. It looked complete to visitors and was a blank shell to AI crawlers. It's a big reason I hand code every site so the words are right there in the page the server sends, and I keep each page lightweight so there's nothing extra for a crawler to download.
Step 4: The AI keeps the passages that answer each piece
Once it has a stack of pages, the AI reads them looking for specific passages that answer each of its smaller searches. A page can cover several topics. Each answer just needs its own clear section.
Ranking first on Google helps less here than most people expect. Moz analyzed nearly 40,000 queries and found 88% of the pages cited in Google's AI Mode were not in the organic top 10 for that same search (Moz, via The Next Web, March 2026). In Ahrefs' updated study, only 38% of pages cited in AI Overviews also ranked in Google's top 10, down from 76% in its July 2025 study (Ahrefs, via Search Engine Journal, 2026). The AI is pulling from whatever answered its smaller searches best, and that's often a different list.
My own research points the same way. I tested 28 independent acupuncture practices across Tampa Bay. Only 10 had pages built around specific conditions like fertility, back pain, or migraines, and 9 of those 10 were cited by AI. Among the other 18, fewer than half were cited at all (Answer Visible Lab Report #1, 2026).
Chopping pages into tiny fragments is unnecessary, and Google's own search team has said so (more on that here). Clear headings and a direct answer near the top of each section are enough.
Step 5: The AI checks the passages against each other
Before it names a business, the AI looks at whether its sources agree. If your website, your Google Business Profile, and your Yelp listing all say the same thing about where you are and what you do, the AI can say it with confidence. If they disagree, it has a reason to hedge or pick someone else.
It also pulls from places beyond business websites. Moz found user-generated platforms like Reddit, YouTube, Facebook, and LinkedIn make up a meaningful share of AI Mode citations, and most AI Mode answers cite 10 or more different web addresses (Moz, via The Next Web, March 2026). Reviews, local news mentions, and directory listings all count as the rest of the internet vouching for you.
None of these companies publish exactly how much each source counts. What I can tell you from testing is that mismatched details are one of the most common reasons a solid business gets skipped.
Step 6: The AI writes one answer and names a few sources
Last, the AI writes an answer in its own words using the passages it kept, and links a few of the sources it relied on. This is where a lot of the "it's random" feeling comes from. The writing step has some built-in variety, so the same question can come back worded differently, or with a different business mentioned. Your location, whether you're logged in, and what you've asked before can all change it too.
That's why a single screenshot doesn't tell you much. When I track a client's visibility, I run the same set of real customer questions every month and look at the pattern across all of them.
And no, a business can't pay its way into the answer itself. OpenAI began testing ads in ChatGPT on February 9, 2026, for Free and Go users in the US, and says the ads are labeled "Sponsored," shown separately from the answer, and don't influence what ChatGPT says (OpenAI, February 2026). Google also shows ads in AI Overviews and AI Mode, labeled as sponsored. Ads exist. They sit next to the answer, clearly marked.
What you can control at each step
| Step | What the AI is doing | What you control |
|---|---|---|
| 1. Decide to look it up | Choosing between its old training snapshot and a live search | Keep your website current. The live search reads what's there today. |
| 2. Split the question | Turning one question into several smaller searches | Answer the follow-up questions your customers really ask: cost, timing, specific situations, comparisons. |
| 3. Pull from the index | Searching a saved catalog built by crawlers | Let the crawlers in, make sure your words are in the page without needing JavaScript, and keep pages fast and lightweight. |
| 4. Keep the best passages | Reading for sections that answer one piece clearly | One clear question per section, with the answer near the top. |
| 5. Cross-check sources | Comparing your site with listings, reviews, and mentions | Same name, address, phone, and services everywhere you're listed. |
| 6. Write the answer | Summarizing and citing a few sources | Track the same questions over time so you see the pattern, not one lucky or unlucky screenshot. |
Google also made this easier to measure. In June 2026, Search Console added a report showing how often your pages appear in AI Overviews and AI Mode, though for now it shows appearances only, without clicks or the searches behind them.
What's still unknown
The steps above are well documented. The exact math inside each one is private to each company, and it changes often. Anyone who promises you a guaranteed spot in ChatGPT's answers is guessing. What you can do is remove every reason the AI would skip you at each step, then watch the results with real tracking. That's the whole job, and it's a lot less mysterious than it looks from the outside.
Summary
- AI search runs in six steps: decide whether to look it up, split the question, pull pages from an index, keep the best passages, cross-check sources, and write the answer.
- ChatGPT rewrites questions into targeted searches, and Google's AI Mode breaks them into subtopics and searches them all at once (OpenAI Help Center; Google, May 2025).
- Major AI crawlers don't run JavaScript, so words that only appear after JavaScript loads can be invisible to them (Vercel, December 2024).
- Ranking first on Google matters less than it used to: 88% of AI Mode citations weren't in the organic top 10 for the same search (Moz, March 2026).
- Ads in ChatGPT and Google's AI answers are labeled and kept separate from the answer itself (OpenAI, February 2026).
Sources: OpenAI Help Center, "ChatGPT search" (query rewriting, third-party search providers, OAI-SearchBot). Google, I/O announcement, May 2025 (AI Mode and query fan-out). Google Search Central, "AI features and your website" (eligibility requirements). Vercel and MERJ, "The rise of the AI crawler," December 2024 (JavaScript rendering). Moz study of about 40,000 queries, via The Next Web, March 2026 (AI Mode citations vs. organic top 10). Ahrefs, via Search Engine Journal, 2026 (AI Overview citations vs. top 10). OpenAI, February 9, 2026, via eWeek (ChatGPT ads test). Perplexity, PerplexityBot crawler documentation. Google Search Console generative AI performance report, June 2026, via PinMeTo. Answer Visible Lab Report #1, 2026.