How an assistant decides which businesses to name

People ask me what the ranking factors are for ChatGPT. The honest answer is that no such list has been published, and anybody who hands you one has made it up. What can be done is to describe the machinery well enough to work out which parts of it are within reach. That turns out to be most of the parts that matter.

What follows is a working model. It is not a leaked diagram. It is what the observable behaviour of these systems is consistent with, and it has held up well enough to plan work around.

The answer is retrieved before it is written

When somebody asks an assistant for a plumber in Peckham, it is not recalling plumbers from training data. For anything local, current, or specific, it goes and fetches. It issues one or more searches, pulls back a set of pages, and composes an answer grounded in what it has just read. OpenAI documents ChatGPT rewriting a prompt into one or more targeted search queries, sometimes adding a general location inferred from the IP address,¹ and Microsoft says Copilot is grounded in web search results when someone is looking for information.²

That single fact reorganises the problem. The question is no longer whether a model has heard of your business. It is whether your business appears in the set of pages it pulls back, and whether those pages say something clear enough to summarise.

What it is reading, and what it is not

Your own site is one source among several, and for a local question it is often not the most influential one. The set usually includes your Google Business Profile or its equivalent, the directories and review sites that cover your sector, local press, trade bodies, and threads where people have discussed businesses like yours.

The providers say a little about this. Google's AI Mode draws on the Knowledge Graph and information about the real world as well as web pages,³ and Google offers developers a Gemini tool that answers from Maps data on more than 250 million places.⁴ OpenAI says ChatGPT works with third-party providers for local information such as nearby business listings, and does not name them.¹ So listings data sits alongside web pages in what gets read, and the sources differ from one assistant to the next, which is one more reason to check each one separately.

The practical consequence is uncomfortable. A business can have an excellent website and still be absent from the answer, because the sources being read do not mention it. Rebuilding the site does not fix that. Getting into the sources does.

There is a second, more technical version of the same problem. Most of the crawlers these assistants use fetch the HTML and do not run JavaScript. A December 2024 log analysis by Vercel and MERJ found that the OpenAI, Anthropic and Perplexity crawlers it measured did not render it, while Google's Gemini and Applebot did.⁵ If your pages assemble their content in the browser, a crawler that does not render sees an empty shell where your copy should be. It is worth checking what your site returns before any script runs, because that is the version some of these systems are reading.

Agreement between sources does the heavy lifting

An assistant naming three businesses is making a small bet each time, with no way to verify anything it says. So it leans on the cheapest proxy for confidence available: whether independent sources agree.

A business described identically on its own site, its profile, two directories and a local paper is a safe pick. A business with three phone numbers in circulation, two trading names, and a service list that changes depending on where you read it is a risk, and there is no shortage of alternatives. The assistant has no incentive to work out which of your addresses is current. It can name somebody else.

This part is my reasoning, with no provider documentation behind it. The nearest documented equivalent is Google's own local guidance: businesses with complete and accurate information are more likely to show up in local results, and those results are ranked on relevance, distance and prominence.⁶ Given that Google's AI features draw on its own information about places, I would expect the same discipline to carry over.

The questions are longer, and that changes what wins

Nobody types "emergency plumber SE15" at an assistant. They describe the situation: a boiler making a noise, on a Sunday, in a flat with a shared meter, and can anyone come out today. That is more detail than a search box ever collected, and it means the answer can be far more specific about who fits.

Specificity is where the opportunity sits. A business whose pages say it covers emergency call-outs at weekends, across a named set of postcodes, with a typical response time, matches that question directly. A business whose pages say it offers a friendly and reliable service matches nothing, because every competitor has written the same sentence.

The best controlled evidence points the same way, though it is not about local search. In the GEO paper (Aggarwal and others, KDD 2024), adding statistics, quotations or citations to a source page raised its visibility in generated answers by 30 to 40% on the authors' main measure, while keyword stuffing gave little to no improvement.⁷ The tests used GPT-3.5 and Perplexity on a benchmark the authors built, so treat it as a direction to test.

What a quotable page looks like

• The service named the way a customer would say it out loud, in a heading, near the top.
• Where you cover, as actual place names and postcodes, not "and surrounding areas".
• What it costs, or what it starts at, or the range. A number you are willing to stand behind beats silence.
• How long it takes, and how quickly you can get there.
• Who it is for and who it is not for. Saying who you do not serve is a strong signal, and almost nobody does it.

Where this leaves the work

Mostly where it already was. Be findable, be described consistently, be specific, and be present in the places your sector is discussed. I have written separately about what this changes and what it does not, because the strategic answer is shorter than people expect.

What nobody can tell you

How the sources are weighted. Whether a citation from one directory counts for more than another. How often the index behind a given assistant refreshes. Whether the answer you saw yesterday is the answer today, because these systems change without announcements and without version numbers you can read.

So treat anybody selling certainty here with suspicion, and that includes me. What is defensible is the direction of travel: consistent, specific, well-cited businesses do better than vague ones, and that has held through every change in search I have worked through. This is more of the work you were already meant to be doing, and the price of skipping it has gone up.