How search and AI answers work

How AI answers see you

What an answer engine is

ChatGPT, Gemini, Perplexity and Google's AI Overviews (the summary Google now prints above the blue links) all do the same thing. Instead of handing you ten links, they write you a paragraph. Somewhere in that paragraph a business gets named, and beside it there is usually a small citation linking back to a page. GEO is the shorthand for the work of being the business that gets named.

Where the answer comes from

Two places, and only one of them is yours to influence.

The first is training. The model read an enormous amount of text months before you asked it anything. You cannot edit that, and it may be out of date about you, or wrong about you.

The second is a live look at the web. When a question needs current or local facts, say who should I use for this, near me, the engine runs a search, fetches a handful of pages, reads them, and writes its answer out of what it just read. Those fetched pages are what the citations point at. That step happens after your site changed, which makes it the one that can reflect your business as it is today.

If the robots are blocked, you cannot be quoted

Those fetches are made by named robots, and the names matter, because they do different jobs. Three groups:

• The ones that decide whether you can be found and shown: OAI-SearchBot, Claude-SearchBot, PerplexityBot and Googlebot. These are the search fetches. Turn one of these away and that engine has nothing of yours to quote.
• The ones that are a preference about model training: GPTBot, ClaudeBot and Google-Extended. Blocking these is a decision about whether your words help train a model. It does not decide whether you turn up in an answer, and if you blocked them on purpose, leave them blocked.
• The ones a person sets off: ChatGPT-User, Claude-User, Perplexity-User. These fetch a page because somebody in a chat asked the assistant to open that link. Blocking one stops that single fetch; it does not decide anything else.

A website can turn any of them away in two places. One is robots.txt, a plain text file at the root of a site listing which robots may read what. The other is the network layer: a security or bot-protection service answering the robot with an error before it ever reaches a page. Plenty of sites are turning these robots away without anyone having decided to; it is often a default somebody left switched on.

When a search fetch fails there is nothing to quote and nothing to link to. A model may still mention your name from what it remembers, but it cannot cite you, and it cannot see a single thing you have changed since. Nothing else in this section helps while that is true.

One honest caveat about how we check it. We ask for your homepage from our own servers, using each robot's name. A security service that verifies a robot by its address can refuse us and let the real one through, so a refusal we see is evidence rather than proof. What robots.txt says is settled either way: it is the instruction they all obey, whoever is asking.

There is more than one index

Engines do not all read the same list of pages, and which list any given assistant draws on is not something anyone outside that company can see from the outside. The point is not which name is on the index this quarter. It is that a site can sit comfortably in Google's index and be missing from another one entirely, and an engine drawing on the list you are missing from has no route to you at all. If you have only ever checked Google, you have checked one of them.

What a business can actually influence

• Say plainly who you are, where you work and what you do, in the visible words on the page, near the top. An engine writing a sentence about you can only use sentences it can find. A homepage that says only "Quality Services" has not said what the work is or what town it is in, and neither has one that says "We make great things".
• Structured data. A small block of machine-readable code inside the page, written in a shared vocabulary called schema.org, that states the business name, address, phone and hours as data rather than as decoration. It takes the guesswork out of which words on the page are the name. It is tidiness, not a lever: see the last section on this page.
• The same name, address and phone everywhere. On your own pages, on your Business Profile, on any listing that carries you. When those disagree, every engine and directory reading them has to guess which one is you, and some of them guess wrong.
• A complete Business Profile, where you have one. Hours, address, categories and the basics filled in. Reviews matter here too, as a trust signal and a factor in the map, and we will not tell you they are what makes an assistant say your name.
• Being crawlable at all. See above. It is the precondition for the rest.

None of these is a lever that makes an engine name you. They are the difference between being reachable and quotable, and being invisible to the step that does the naming.

How Google sees you

What we mean by Google's own results

The list of links, and the map with a few businesses on it. Not the AI summary above them; that is the section before this one. This is still where most people land, and it is the ordinary, unglamorous half of the job.

Titles and descriptions

Every page carries a title, the headline Google usually prints in the result, and a description, the grey sentence underneath it. Two honest caveats. Google rewrites titles when it decides its own wording fits the search better, so yours is a strong suggestion rather than a guarantee. And the description has never been a ranking factor; Google has said so plainly. It is still worth writing, because it is often the last sentence a person reads before deciding whether to click.

A title reading "Services - Services" costs you on both counts at once: it tells Google nothing specific, and it tells a person nothing at all. Plenty of sites carry one, usually because a theme filled it in and nobody looked again.

One page per thing you sell, in the place you sell it

Search is matching. A person types what they want and where they are. A homepage listing nine things you do and naming no town is a weak match for all nine of them. A page that is about one of them, in one town, can be a strong match for one. This is not a trick. A page about the thing is simply the honest answer to a question about the thing. It also gives you somewhere to put the detail, the pricing and the photographs that would clutter a homepage.

Speed, on a phone

Most people arrive on a phone, and often not on wifi. Google does use page experience signals, load speed among them, and describes them as a small factor. The larger effect is much simpler than ranking: people leave. A page that takes six seconds has lost a share of its visitors before it has said anything, and that happens whether or not it ever moves your position.

Links, and why spam anchors hurt

Links from other websites have been part of how Google ranks pages from the beginning. A link is roughly a vote. Which is exactly why they have been bought and manufactured for just as long, and why Google's spam policies name link schemes as a violation.

The visible words of a link, called the anchor text, are part of what is read. When most of the links pointing at a site use unrelated commercial anchor text, meaning wording that has nothing to do with the business, that pattern was built by somebody else, usually with no idea the owner would ever see it. The realistic outcome is that those links are discounted rather than counted. Google's own guidance is that the disavow tool is for a manual action, which almost nobody has, so we watch this and tell you what it looks like rather than selling you a cleanup.

Nobody can promise a ranking

Google's own guidance on hiring an SEO says to walk away from anyone who guarantees rankings, and that is the right advice. No one outside Google controls the ranking. It is decided by a system nobody outside the company can see, against competitors who are also working at it, and it changes.

What can be promised is the work: the pages written, the settings corrected, the blocks removed, and a straight account afterwards of what changed and what did not.

What we do not sell, and why

Every one of these is on somebody's list of things to buy. They are not on ours, and this is the whole reason for each.

• An llms.txt file. It is a proposed file for telling AI assistants about a site. No engine has said it reads one, and we have never seen it change anything, so charging for it would be charging for a hope.
• Structured data as a way into AI answers. Schema is worth having: it states your name, address and hours as data instead of as decoration. What it is not is a lever that gets a business named in an answer, and anyone selling it as one is selling tidiness at the price of a result.
• A rank position in an AI engine. There is no ranking to hold. An assistant writes a paragraph and names whoever it names, differently the next time you ask, so a number in that space would be made up.
• A disavow file. Google's own guidance is that the disavow tool is for sites with a manual action against them, which almost nobody has. If your link profile has junk in it we will show you and tell you plainly that no action is recommended.
• Credit for a change we cannot prove we caused. Answers and results move on their own, all the time. Without holding something back and comparing, nobody can say a fix caused a change, so we report what changed and stop there.
• Bing as a route into ChatGPT. Being listed on Bing is ordinary search hygiene and worth doing on its own account. Which index any assistant actually draws on is not something we can see from outside, so we will not sell you one as the door into another.

If somebody quotes you for one of these, that is not automatically a scam. It is a good moment to ask them what result they expect and how they would know it happened.