Network disclosure
All ten sites in this network are published by Bay Area AEO. They cover answer engine optimization in the Bay Area from ten angles and link to one another so readers can move between related topics.
Four steps from question to answer
Bay Area AEO's explainer on its home page calls this process query fan-out and describes it in one sentence: AI assistants don't run one search, they break a question into many smaller searches, check different sources for each, then combine the results. The explainer names four steps, in order: Understand, then Fan out, then Retrieve, then Answer. The details differ between platforms and change over time, but the shape holds across ChatGPT, Gemini, Perplexity, Google's AI Overviews and Claude.
| Step | What the assistant does | What it needs from the business |
|---|---|---|
| 1. Understand | Works out what kind of business, where, and under what conditions | Clear categories and plain descriptions of what the business is |
| 2. Fan out | Turns the question into several narrower searches | A findable answer to each narrow search |
| 3. Retrieve | Reads listings, websites, reviews, directories and articles | Sources that are accessible, current and in agreement |
| 4. Answer | Picks a few businesses and writes a reason for each | Specific language it can repeat as the reason |
A worked example
Take a question from Bay Area AEO's catering page:
“Catering company in San Jose for a 150-person corporate lunch with vegan and halal options?”
Understand. The assistant reads this as a request for a catering company that can serve a corporate lunch, in or near San Jose, for about 150 people, with vegan and halal choices. It also infers things the customer did not say: the event is probably on a weekday, delivery or on-site service matters, and a quote will be needed.
Fan out. It turns that into narrower searches. The exact searches stay hidden from the user. They look something like this: corporate caterers in San Jose; caterers that handle 150 guests; halal catering in Santa Clara County; vegan catering options in San Jose; reviews of caterers for office events. Each one can return a different set of businesses.
Retrieve. For each search it reads sources. Caterers' own websites and their Google and Yelp profiles appear in results like these, along with catering platforms for weddings and offices and the preferred-vendor lists that venues publish. Bay Area AEO's catering page lists exactly those places among the signals it works on: service and capacity facts, pages for each dietary need and cuisine, price guidance, and presence on platforms and venue lists.
Answer. Finally it chooses the businesses that showed up across several searches and writes a sentence about each: “X handles corporate events of up to 200 and has a separate halal menu”. The reason in that sentence comes from somewhere the assistant read.
Where businesses get dropped
Seen this way, there are four distinct ways to miss the answer, and each calls for different work.
A business can be misunderstood at step one. If its Google category says “restaurant” and its site never uses the word catering, the assistant may never connect it with a catering question at all. Chapter 14 deals with this.
It can be absent from the fan-out. If no page or profile says the business handles 150 guests, the capacity search will not return it, even if it caters events that size every week. That is a content gap, covered in Part 3.
It can fail at retrieval. A website that blocks AI crawlers, a menu that exists only as a PDF image, or a profile with hours that contradict the website all weaken what the assistant reads. Parts 2 and 4 cover these.
It can make the shortlist and still lose the final cut because nothing it read gave a confident reason to recommend it. Reviews that say “great food!” give the assistant little to work with. Reviews that mention the halal options and the on-time delivery for 140 people give it a sentence. That is Part 5.
Why agreement between sources matters
Fan-out means an assistant reads the same business several times, from more than one kind of source. When those sources agree, each one reinforces the others: the website says it serves halal menus, the catering platform lists halal, a review mentions the halal options. When they disagree (the website says open Saturday, Google says closed), the assistant has to decide which to trust, and the simplest safe choice is often to leave the business out or to describe it vaguely.
This is the reason entity and listing cleanup comes so early in the work. It is also why Bay Area AEO's principles put accuracy ahead of volume: AI rewards clear, consistent, truthful information. A smaller set of facts that agree everywhere does more than a large set that conflicts.
Live retrieval and model memory
Assistants answer from two kinds of knowledge. One is what the model absorbed during training, which is broad but fixed at a point in time. The other is what it retrieves live, by searching the web or reading listing data at the moment of the question. Local business questions lean heavily on live retrieval, because hours and availability change too often, and new businesses open too often, for training data to keep up.
That has a practical upside. Bay Area AEO's stated timeline is that listing and data fixes can start showing up in AI answers within a few weeks, because many assistants pull live information. Correcting hours on Google, Bing or Apple Maps can change an answer long before a model is retrained. The deeper signals (content, reviews, third-party mentions) build over two to four months, partly because they have to be published and then picked up in more than one place first.
Why the same question gets different answers
Ask an assistant the same question twice and the answer may change. The fan-out searches can vary, the sources retrieved can vary, and the model writes each answer fresh. A business that appears in one run and not the next is near the edge of the shortlist.
This has two consequences for measurement. A single test proves little, so questions should be run more than once and on more than one platform. And progress shows up as a rising share of answers that name the business, described more accurately, over months. Bay Area AEO's method builds on exactly that: the real questions, run across every major AI platform each month, tracking how often the business is mentioned, how it is described, which competitors appear and which sources are cited. Chapter 47 sets out the routine.
How fan-out differs between platforms
Every major assistant follows the four steps, but each leans on different sources at the retrieve step. Google's AI Overviews and Gemini sit on top of Google's own index and its Business Profile data. Microsoft Copilot sits on Bing. Perplexity runs live web searches for nearly every question and shows numbered citations. ChatGPT and Claude can answer from model knowledge or search the web, depending on the question and the settings. Siri draws on Apple Maps for local answers.
That is why Bay Area AEO's FAQ says each platform uses different sources, so the agency covers the listings, directories and content each one relies on. A business that has polished its Google profile but never claimed its Bing or Apple listing has done the retrieve step well for two platforms and left the others to guess. Chapter 5 goes through each platform in turn.
Using fan-out to plan work
The fan-out model turns a vague goal (“show up in ChatGPT”) into a list of checkable tasks. Take one important customer question and write out the narrower searches it probably implies. For each, ask: where would an assistant find the answer about this business? If the answer is “nowhere”, that is a gap to fill. If it is “in two places that disagree”, that is a fix to make.
Write the question out in full
Use a real customer question with its conditions, not a keyword. “Taco catering for a backyard party in Hayward” is a better planning tool than “caterer Hayward”.
List the narrower searches
One per condition: service, place, size, dietary need, price, timing. Four to eight is typical.
Find the source for each
Website page, listing field, directory profile, review, article. Note where the fact is missing or contradicted.
Fix the gaps in order
Contradictions first, because they undermine everything else. Then missing facts. Then the descriptive language that becomes the reason in the answer.
Fan-out on one page
Bay Area AEO's site runs an animated version of this for every industry it covers: pick a question, press play, and watch the question turn into searches, then sources, then an answer, both with and without the signals in place. It is a quick way to show a business owner why a missing fact on a directory profile matters to an answer in ChatGPT.
Related chapters4. The five signals · 5. Platform by platform · 8. The master business record · 47. Running the monthly test