How AI Search Actually Works & How I’d Build for It From Scratch
Share
AIO and GEO get much easier to understand once you stop starting with the acronyms.
Someone asks a question. The system has to work out what that person means, find information that might help, and produce some kind of response. That is the whole problem.
Google, OpenAI and Microsoft are obviously not running identical systems, and nobody outside those companies has the complete decision process that explains why one source gets selected and another does not. I get suspicious fairly quickly when somebody claims to have that formula. But enough of it is public now that we can build a useful working model without pretending it is a disclosed universal architecture.
Here is how I think about it.
A planning model, not an architecture diagram.
The real systems are more complicated than that, they differ by platform, and they contain proprietary mechanisms none of us get to see. But the public pieces are real. Google openly describes retrieval augmented generation and query fan out in its generative Search guidance. OpenAI documents that ChatGPT search can rewrite a question into one or more targeted searches and issue further searches after reviewing the first set of results. Microsoft now gives publishers a limited window into citation activity through cited URLs and sampled grounding query phrases in Bing Webmaster Tools.
That is enough to stop guessing at everything.
Start with what the customer actually means
Search marketing spent years training businesses to think in short phrases. Outdoor kitchen contractor. Inventory management software. Best travel backpack. People do still search that way, but conversational search lets them explain the entire problem instead.
Someone might now ask:
I want to rebuild an outdoor entertaining area, but the property gets heavy afternoon sun and strong seasonal wind. What should I plan before I start talking to contractors?
That question is about considerably more than finding a contractor. Buried inside it are shade, materials, utilities, wind exposure, cooking equipment, drainage, layout, budget and maintenance. Google describes query fan out as a set of concurrent related queries generated to gather more information around the original request, and ChatGPT search can rewrite a question into targeted searches, then issue more specific ones once it sees what comes back.
This changes how I think about keyword research. I still want to know what people search for. I also want to know what they are trying to work out, and those are not always the same thing. A spreadsheet may tell me that outdoor kitchen cost gets searched frequently. Talking to the people who sell and build them may reveal that customers are far more worried about how utilities get routed, which materials survive exposure, and what has to be decided before any concrete gets poured.
That second layer usually produces the better content.
The information has to be findable in the first place
This sounds too obvious to state, and yet a surprising amount of AI search advice is being sold to companies whose most important pages are poorly linked, difficult to crawl or not indexable at all.
I went through the eligibility mechanics in SEO Didn’t Disappear, so briefly: for Google, a page still has to meet the technical requirements for Search, be indexed, and be eligible to appear with a snippet before it can qualify for generative visibility. The Search Console control governing whether a site can appear in and help ground those features completed its worldwide rollout on 31 August 2026, with inclusion set as the default. For OpenAI, allowing OAI-SearchBot is what makes content available for ChatGPT search answers, citations, summaries and snippets.
Before worrying about being cited inside an AI answer, I want to know the important information can be found at all.
Finding a page is not the same as understanding it
A good site makes the relationships fairly obvious. This company offers these services. These are the products. These people work here. This is the kind of customer it serves. These are the areas where it has real experience, and these pages explain those areas in more depth.
A surprising amount of business copy works against that. Take a consulting firm whose homepage says it creates transformative strategies for tomorrow's leaders. That could be an accounting practice, a management consultancy, a leadership coach, a software vendor or somebody running corporate retreats. Now compare it with a firm that says it advises privately held construction and manufacturing companies on ownership transitions, restructuring and financial planning. Less clever. Far more useful.
Structured data helps Google understand standardized information and can make pages eligible for certain Search features, but Google is explicit that no special structured data format is required for generative Search. Schema is helpful plumbing. It does not turn vague information into good information.
Retrieval is where it gets interesting
A system backed by search does not load the entire public web into the context of every answer. It retrieves what looks relevant to the particular request. Google describes this directly as retrieval augmented generation: its core Search ranking systems pull relevant, current pages from the index, and that information then grounds the response.
Bing's terminology offers another clue. Its AI Performance report includes grounding queries, meaning the key phrases used when retrieving content that later appeared as a reference in supported AI answers. Microsoft shows only a sample of overall citation activity, so nobody should mistake that for a complete view of retrieval.
This gives marketers a better objective than writing content for AI. I have never known what writing for AI is supposed to sound like anyway. The practical goal is to publish information useful enough to be available when the right question is being answered.
There is a real limit here, though. Google tells us some of how its retrieval and grounding work. Microsoft exposes some citation and grounding data. OpenAI publishes parts of how its search queries get formed and how publishers can make content available. None of them publishes the complete source selection and evaluation formula. So when somebody produces a chart claiming a certain kind of mention is worth 1.7 times another one inside the LLM algorithm, I want to know where that number came from. That is usually the point where the conversation gets quieter.
Average information has become very cheap
Consider two companies selling outdoor equipment. One publishes 10 tips for your next camping trip. The other publishes We tested four tent fabrics through a full wet season: what stretched, what leaked, and what we would use again.
The second has the chance to show real testing, photographs, measurements, failure points and methodology. Maybe the test is imperfect. Maybe one result surprises them. Even better, because at least there is something to evaluate. The first article could have been produced by almost anybody, including a model, in about ninety seconds.
Google's current generative Search guidance encourages unique information led by people with genuine expertise, and warns against recycling material that already exists or could easily be produced by a generative model.
That is one of the larger practical opportunities AI has handed to good businesses. It raised the floor on generic writing. Everybody can produce average information now, quickly and at no real cost. The companies that actually know something still hold the advantage, provided they bother to publish what they know.
I also pay attention to how information is written inside the page. A generative system may surface a relevant portion of a page rather than treating the page as one indivisible answer, and Google says its systems can understand multiple topics within a page without publishers artificially breaking everything into tiny sections. So a case study should explain what actually happened. A comparison should make clear what is being compared. A technical explanation should not hide the useful answer under three paragraphs of throat clearing. None of that means writing like a machine. Clear writing works fine.
Authority is not how often you call yourself an authority
Imagine a wellness brand claiming it is a leader in its category. The claim appears on the homepage. Then in a blog post. Then in a founder biography. Then in a press release the company wrote about itself. Then across ten social posts. Plenty of mentions, and still one source talking.
Now imagine a company whose products have been independently reviewed, whose founders have been interviewed by credible publications, whose customers have left substantial feedback, and whose expertise gets referenced by other organizations. Different situation entirely. That outside evidence carries weight precisely because the business does not control all of it.
I would be careful about turning that into another exercise in manufacturing signals. Google specifically warns against seeking inauthentic mentions across the web in the hope of influencing generative Search. Do not build fake evidence. Build a company that produces real evidence. Sometimes that takes far longer than marketers would like, which does not make it any less useful.
I use the word corroboration carefully, too. It makes sense to create an information environment where claims about a business are supported somewhere other than the business's own website. But no platform publishes a rule telling us how many reviews, mentions, citations or independent sources are required before an AI system trusts a company. That kind of precision simply is not public.
The answer now sits between the search and the click
Traditional search made the user do most of the synthesis. Open one site. Back out. Open another. Compare. Search again. Generative search performs part of that work inside the answer itself. A product appears in a comparison. A technical article supports a factual statement. A business turns up on a shortlist. A page gets cited whether or not anybody clicks it.
That is why visibility is becoming broader than rankings and organic sessions. Measurement needs the same restraint as everything else, though. Google launched dedicated Generative AI performance reporting in Search Console on 3 June 2026, starting with a subset of UK site owners and completing a worldwide rollout on 31 August. The reports show how often URLs appeared in Google's generative Search and Discover features, which pages appeared, country and device information, and change over time. Useful. They do not reveal every query behind every appearance, and they do not explain why Google selected a particular page.
Bing goes further in a different direction. Its AI Performance report, still in public preview, shows citation counts, cited pages, visibility trends and a sample of grounding query phrases across supported Microsoft AI experiences, and it expanded in June 2026 with intent classification, topic grouping, citation share and comparison tools. Microsoft states plainly that those citation figures do not indicate ranking, authority, page importance or where a citation appeared inside an answer.
So these tools are getting good at a specific set of questions:
- Are we appearing at all?
- Which pages are involved?
- What kinds of queries seem connected to them?
- Is visibility rising or falling over time?
They are not handing over the retrieval, evaluation or source selection formula. Worth remembering the next time somebody turns a dashboard into certainty.
If I were starting with an empty domain
I would not begin with GEO. I would begin with facts. What is this company actually called? Who works there? What does it sell? What can it legitimately claim expertise in? Which products, services and customer problems actually generate revenue? That becomes the truth layer, and everything else gets built on top of it.
Then I would build the website around how customers understand the business, not around an internal org chart and not around a giant list of keywords. A design build company might need strong pages for its major services, project types, process, portfolio and expertise. A product company is different: categories, product pages, comparisons, support information and solid product data probably matter more. Software is different again, with product, use cases, integrations, implementation, security and documentation carrying a large share of the weight. I would build those commercial pages before producing fifty supporting articles.
Once the core exists, start collecting real customer questions. A specialty outdoor contractor probably hears things like: what should be designed before concrete is poured, which countertop materials hold up outside, how should utilities be planned, and which features turn into maintenance problems. There are dozens more. I would not turn every one into a 1,000 word page. Some deserve a paragraph. Some deserve video. A few are worth a serious article.
Then look for the material competitors cannot reproduce by opening the same AI tool. Project lessons. Testing. Original numbers. Case studies. Photographs. Comparisons. Approaches that failed. Real customer situations. Internal expertise. Sometimes the best authority piece is not glamorous at all. It is just useful.
From there, build the outside evidence: reviews, legitimate publications, professional references, partnerships, product reviews, expert commentary, whatever makes sense for the category. Then distribute. One real field test can become a long article, a video, three short clips, a customer email and a genuine addition to the product page. A strong case study can support sales conversations, LinkedIn content and a conference talk. Not every platform needs a fresh invention every morning. Use the knowledge you already paid to create.
What I would not spend time on first
I would not publish an llms.txt file and call the project finished. I would not generate hundreds of near duplicate question pages. I would not rewrite normal sentences into awkward fragments because somebody decided language models need everything in answer blocks. I would not manufacture brand mentions. And I would not bury a simple business under layers of structured data that claim more than the visible site delivers.
For Google Search specifically, the current guidance directly rejects the need for special AI markup, llms.txt optimization and artificial content chunking in its generative Search features.
Most companies have better things to fix first. Make the business clear. Make the important information accessible. Publish something that comes from actual knowledge. Give independent people reasons to validate it. Then watch what search systems genuinely retrieve, cite and surface, using the limited but improving data the platforms now make available.
That is less exciting than promising to rank a brand in ChatGPT. It is also a far more serious way to build one.
Sources: Google's AI features and your website guidance and its Search Console generative AI performance reports announcement; Microsoft's AI Performance public preview and its June 2026 expansion; and OpenAI’s crawler documentation for OAI-SearchBot.