ChatGPT search happens when somebody asks a chatbot a question and reads an answer instead of a list of links. A few sources sit underneath that answer, and those are the pages a reader might still click. This article explains how a business becomes one of them, and what to fix on its own site first.
None of the work is exotic. Most of it is ordinary publishing done with more care than usual. Some of it is technical, and a little of it is just being easy to find elsewhere on the web.
What ChatGPT search actually does
Ask a chatbot something it already knows and it simply answers. Ask it something current, local or unusual, and it goes looking. That second behaviour is ChatGPT search.
In broad terms the tool searches the web, opens a few of the results, then writes a summary. Therefore there are two gates rather than one. First the page has to turn up in that search. Then it has to be worth quoting.
Most advice online only deals with the first gate. The second gate is where the effort pays off, because a page can rank well and still say nothing a chatbot can lift.
Where the answers come from
Two very different sources feed a reply, and they behave nothing alike.
What the model already learned
Part of any answer comes from training, which is a snapshot of a great deal of text taken before the model shipped. You cannot edit that snapshot. Asking it for a correction does nothing either. If your company changed its name last month, it has not heard.
What the tool reads live
The rest comes from a live fetch. Here the chatbot picks a handful of addresses, opens them and uses what it finds. Consequently this half is the half you can change this week, and it is the half this article deals with.
How ChatGPT search chooses a source
Nobody outside the company can see the selection rules, so treat anyone who claims otherwise as a salesperson. Still, you can reason from the job the tool is doing. A source has to clear a few plain hurdles.
- Reachable. If a crawler cannot open the page, nothing else matters.
- Relevant. The page has to answer that exact question, not the general subject.
- Liftable. A short passage that stands on its own is easy to quote.
- Corroborated. A fact repeated across several sites is easier to state with confidence.
- Current. Where the answer moves over time, an out of date page is a poor source.
None of that is new thinking. It is ordinary editorial quality, judged by a reader who never skims.
The kind of page a chatbot quotes
Start from what the tool needs. It wants a short passage it can lift, on the exact question somebody asked, without having to interpret the page around it.
A page called "Our services" offers a list of nouns. A page called "How long does a boiler service take in winter" offers a sentence that answers a question, so there is something to quote. Consequently the unit of work is the question, not the topic.
Put the answer in the first two sentences of the section. Add the detail after that. A human reader loses nothing, while a machine finds the answer exactly where it expects one.
Writing for ChatGPT search without writing badly
There is a fear that writing for machines means writing worse. In practice the opposite is true. Every habit that helps here also helps a tired person reading on a phone.
- Answer first, then explain, in every section.
- Keep paragraphs down to two or three sentences.
- Use headings that read like questions somebody would actually ask.
- Say the plain thing rather than the clever thing.
- Give the specifics: places, materials, timings and conditions.
Vague copy is the real problem. "We deliver bespoke solutions" gives a machine nothing to repeat, so nothing comes back. Our companion piece on writing content for AI search goes further into that habit.
One caution for regulated trades. If you work in law, medicine, dentistry or finance, your regulator or professional body sets rules about what a public page may claim, and a chatbot repeating that page does not soften them. Check the current guidance from your own body before you publish anything that reads as a promise about results.
The technical side of ChatGPT search
Three technical things decide whether a chatbot can open your page at all.
First, crawler access. Some sites block unfamiliar visitors by default, or a security product does it on their behalf. Check the robots file and the firewall rules before blaming the writing.
Second, rendering. If the words only appear once JavaScript has run, some fetchers see an empty page. Plain HTML from the server is the safer choice.
Third, speed. A fetch that times out is a fetch that failed, so ordinary hosting hygiene covers most of this.
Structured data is worth adding as well, though treat it as a label rather than a switch. Nobody outside these companies can say what weight it carries. Still, it states what a page holds in a form nothing has to interpret, and it costs little. We cover the types worth adding in structured data that machines can read.
Being one business, everywhere
A chatbot builds its picture of your company from every mention of it, not from your homepage alone. When those mentions disagree, there is no single version of your business to repeat. That hands the tool a choice you do not control.
So keep the name, the address and the description identical on your own site, in directory listings and on your profiles. Then name the services you actually offer, in plain words, on pages you control. That is entity work at a practical level, and entity SEO covers the thinking behind it.
When the answer about you is wrong
Sooner or later a chatbot will say something about your business that is out of date, or plainly wrong. An old price, a branch you closed, a service you dropped, a name you changed. Reporting it inside the chat window may go nowhere, so treat the correction as your own job.
Work on the sources instead. State the correct fact on your own site, in a full sentence, on a page that exists to answer that question. Then fix the third party listings that still carry the old version, since a live fetch can reach those just as easily.
The trained half will not update because you asked it to. Yet the live half reads current pages, so what you publish is the only real lever you hold over what ChatGPT search repeats. Note the wrong claim, then check it again in a month.
Measuring ChatGPT search without guessing
You cannot open a rank tracker and read off position four. The answer changes between people and between days, so measurement here means sampling rather than counting.
Three habits are worth building, and none of them costs anything.
- Ask your ten real customer questions yourself, on a fixed day each month.
- Record which sites the answer names, including yours when it appears.
- Watch referral traffic from chatbot domains inside your analytics.
Write the results in the same place every time. A pattern only shows up over months, and memory is not evidence. Our piece on measuring LLM visibility sets out a routine you can copy.
A worked example, with round numbers chosen to make the arithmetic easy rather than to describe a typical result. Take ten questions and ask each of them three times, in a fresh session every time. That gives thirty answers. Count the ones that name your site and you have a score out of thirty for the month. Ask the same ten next month, and the two numbers compare properly, because the method did not change.
ChatGPT search next to ordinary search
The two behave differently, and the difference explains most of the advice above.
| Ordinary search | ChatGPT search |
|---|---|
| Often short typed phrases | Often longer, spoken style questions |
| A page of links to choose from | One answer with a few sources |
| Ranking wins the click | A quote wins the mention |
| Position is stable enough to track | Answers vary between sessions |
| The click is the outcome | The mention is often the outcome |
Neither one replaces the other yet. Your own analytics report the split for your site, and until it shifts, treat this as an extra surface rather than a migration.
What is not worth doing
Some of the advice going around wastes effort, and one or two pieces of it do harm.
- Paying a directory that promises a place in AI answers. No directory decides what gets quoted.
- Stuffing the same phrase into every heading, which reads badly and changes nothing.
- Publishing thin pages in bulk, which dilutes the ones that were working.
- Hiding text meant only for machines. That has carried a penalty risk for a long time.
Also resist rewriting the whole site at once. Change ten pages, wait a month, then look at what moved.
Six things, in this order
If you do nothing else, do these six things in this order.
- Confirm that crawlers can open your pages at all.
- Pick the ten questions your customers genuinely ask.
- Give each question its own page or its own clear section.
- Put the answer in the opening lines of that section.
- Make your business details identical wherever they appear.
- Sample the answers monthly and keep the record.
That list is deliberately dull. ChatGPT search rewards the same things a careful editor would ask for, so it is one surface inside the broader practice we set out in generative engine optimization, rather than a separate trade of its own.
Frequently asked questions
Is ChatGPT search the same thing as Google?
No. Google returns a list of pages and lets you choose between them. A chatbot reads a few pages for you and writes one answer, naming only a small number of sources underneath. The overlap is real, because both start with a search of the web. With Google you compete for a position on the page. Here you compete to be one of the sources worth quoting, which rewards a clear answer far more than a clever landing page.
How do I tell if ChatGPT search sends me traffic?
Check the referral report in your analytics for chatbot domains, then compare it against your own sampling notes. Referral data undercounts, because some people read an answer and type your name into a browser later. So treat the referral number as a floor rather than a total. Direct visits rising at the same time as your mentions is the pattern worth watching, and it only becomes visible over several months.
Do I need new pages or can I fix the ones I have?
Fix what you have first. Most sites already own a page for every service they sell, and the fault is usually the writing rather than the absence of a page. Add a clear answer near the top, break the wall of text into headed sections, and name the specifics a customer would ask about. New pages make sense only when a real question has nowhere sensible to live.
How long before I appear in an AI answer?
Nobody can put a date on it, and anyone who does is guessing. The live half of an answer can only use a page once a crawler has fetched it, so that half moves at the speed of crawling rather than on a schedule. The trained half moves far more slowly, since it shifts only when a new model ships. In practice, judge ChatGPT search over a season rather than a fortnight, and ask the same questions each time so the comparison means something.
Does structured data help with ChatGPT search?
It is worth adding, though not as a shortcut. Nobody outside these companies can say what weight it carries. Structured data labels what a page contains, so a machine does not have to infer it from the layout. That matters most for things with a fixed shape, such as opening hours, reviews, staff and events. It will not rescue vague copy. Think of it as making a clear page easier to read rather than making a poor page acceptable.
Should I block AI crawlers from my site?
Only if you have a reason you can state out loud. These tools usually run more than one crawler, and the one gathering training text is not always the one that fetches a page to answer a live question. So check which agents your robots file names, because a blanket block takes you out of the answers as well. Most small businesses gain nothing from blocking. Publishers with a paid archive often decide the other way, and both choices are defensible.
Which businesses benefit most from ChatGPT search?
Businesses whose customers ask complicated questions before buying. Trades, professional services, healthcare, software and anything with a comparison stage all fit that description. If people buy on habit alone and do no research, the effect will be small. The other strong case is any business with a narrow speciality, because a chatbot needs an obvious answer to a narrow question and fewer sources compete for it.
Can I pay to appear in ChatGPT search results?
No, and be careful with anyone selling it. A directory does not decide which sources a chatbot names, so an offer to place you in that list is either advertising in a different product or a listing hoping you will not check. Advertising may well appear inside these tools over time, which is a separate thing from being quoted as a source. Spend the money on the pages instead, since that is what ChatGPT search actually rewards.