Why ChatGPT Quotes Some Pages and Not Yours

A customer no longer always types your category into Google. Sometimes they type a question into ChatGPT or Perplexity and read a paragraph of answer with three or four small source links underneath. If your business is in that paragraph, you got the enquiry. If it is not, you never knew the conversation happened.

The honest starting position is this: nobody outside OpenAI, Google and Perplexity knows exactly how a source gets chosen. There is no documented ranking system the way there is for ordinary search, and anyone who tells you they have cracked it is guessing with confidence. What we can describe is the plumbing, because most of it is publicly documented, and the plumbing tells you quite a lot about what to do.

How an answer engine actually finds you

These tools do not know about your business from memory. When a model was trained, your Diwali offer did not exist. So when someone asks a question that needs current or specific information, the system runs a search, fetches a handful of pages, reads them, and writes an answer grounded in what it just read. The industry word for this is retrieval — look it up first, then answer.

That means three separate gates have to open before you can be quoted:

1. A crawler has to be allowed to fetch your page. OpenAI publishes three named bots: GPTBot, which collects content for training; OAI-SearchBot, which builds the search index ChatGPT uses; and ChatGPT-User, which fetches a page live when a user's question requires it. Perplexity uses PerplexityBot. These names are in public documentation, and your robots.txt file can allow or block each one by name.
2. The search step has to surface your page for that particular question, which still depends heavily on ordinary search signals — relevance, clarity, links, whether the page exists at a stable address.
3. The model has to find a sentence on your page it can lift without rewriting. This is the part almost nobody works on, and it is the part you control most directly.

Check gate one this afternoon. Open yoursite.in/robots.txt in a browser. If you see User-agent: GPTBot followed by Disallow: /, somebody — a developer, a plugin, a security setting, a hosting provider's default — has told ChatGPT not to read your site. Several hosts and security services now block AI crawlers by default, and plenty of owners have no idea it was switched on. If you want AI visibility, that block has to go.

What makes a sentence quotable

Imagine the model as a junior researcher who has eight seconds per page and must attribute every claim. It is looking for statements that survive being cut out and pasted somewhere else.

A quotable sentence is self-contained. It does not begin with "this", "it" or "as mentioned above". Compare:

We offer this at very competitive rates with fast turnaround across the city.

with

Wwwebtech installs and configures CRM systems for businesses in East Delhi, and a typical first-stage setup covers contact import, pipeline stages and WhatsApp enquiry capture.

The second one can be dropped into an answer unchanged. The first one cannot be used at all, because it means nothing once separated from the paragraph above it.

Beyond that, quotable writing tends to share a few properties:

• It names things. Product names, suburb names, document names, file formats, specific services. Models retrieve on specifics. "We serve Delhi NCR" is weaker than "We work with clients in Laxmi Nagar, Preet Vihar, Mayur Vihar and Noida Sector 62."
• It states numbers that are yours to state. Opening hours, minimum order quantity, warranty length, how many days a repair takes, which payment methods you accept. Not invented market statistics — your own operational facts, which nobody else can publish.
• It answers the question in the first sentence of the section, then explains. A page that builds to its answer over four paragraphs is less likely to be excerpted than one that leads with it.
• It is dated. A visible "updated March 2025" gives a retrieval system a reason to prefer you over an undated page saying something contradictory.
• It is consistent everywhere else. If your address and phone number differ between your website, your Google Business Profile, Justdial and your Instagram bio, every system reading about you has conflicting evidence and tends to hedge or skip.

The second-hand problem

Here is the uncomfortable bit. Answer engines frequently cite directories, listicles, forum threads and review sites rather than the business's own website — partly because those pages compare several options, which is what the user asked for, and partly because a business describing itself is weak evidence.

So being quotable is not only about your pages. It is also about whether a third party page that an answer engine does trust contains an accurate sentence about you. A complete Google Business Profile with correct categories, an accurate Justdial and IndiaMART entry, a trade association listing, a genuine mention in a local news piece or a supplier's partner page — these are the sources a model may summarise. You cannot fabricate them, and you should not try. You can make sure the ones that genuinely exist are correct and current.

This is also why ordinary search work still matters more than most AI visibility pitches admit. If a page cannot be found by a search engine, it cannot be retrieved by a system that uses search to find things.

What not to buy

This is a young market with a lot of confident selling in it. Some of it is sold by agencies like ours, so treat this as a warning about the category, not about other people.

• Guaranteed placement in ChatGPT answers. There is no placement to buy. Answers vary between users, sessions and phrasings of the same question. Anyone guaranteeing a position is describing something that does not exist.
• "AI ranking reports" with scores out of 100. You can meaningfully test whether a brand gets mentioned for a set of questions, and that is worth doing. A composite score with no stated method is a dashboard, not a measurement.
• An llms.txt file sold as a fix. This is a proposed file format for telling language models about your site. It is a reasonable idea, it is cheap to add, and as of now no major answer engine has publicly committed to using it. Paying a meaningful fee for one is paying for a hope.
• Schema markup presented as the whole answer. Structured data — machine-readable tags describing your business — is genuinely useful for search features and costs little. But a model reading your page reads the visible words. Tagging a vague page does not make it quotable.
• Bulk "AI-optimised" article packages. Twenty thin pages answering questions nobody asked gives a retrieval system twenty weak candidates instead of three strong ones.

A test you can run yourself

Write down the eight questions a real customer asks before buying from you. Not keywords — questions, in the words they would use. "Who repairs commercial RO systems in East Delhi?" "What does a GST-compliant invoice need on it?" "Is there a minimum order for custom printed boxes in Delhi?"

Put each one into ChatGPT and into Perplexity. Record whether you were mentioned, who was mentioned instead, and which sources were cited. Do it from a logged-out or private session so your own history does not flatter the result.

You will usually find one of three situations. Either competitors are named and you are not, in which case look at what the cited sources say about them. Or a directory is cited and your entry on it is thin or wrong, in which case fix the entry. Or the engine gives a generic answer with no businesses named at all, which means nobody in your category has written a page that answers it properly — and that is the opening.

Then write the page. One question, one page, the answer in the opening sentence, your own specifics underneath, a date on it, and a clear path to contacting you. That page serves a human reader and a retrieval system equally well, which is the only strategy here that is safe from the next change in how these tools work.

Where to start this week

In order, because the first two cost nothing:

1. Read your own robots.txt and confirm the AI crawlers are not blocked. If you are unsure what you are looking at, ask whoever runs your hosting and technical support to confirm in writing.
2. Run the eight-question test and write down what you find.
3. Make your Google Business Profile and major directory entries agree with your website, to the character.
4. Rewrite your three most important pages so each section opens with a complete, standalone answer.

If you would rather have someone do the audit and the rewriting with you, our AI visibility work starts with exactly the test above, and the findings are yours whether or not you continue. If the deeper problem turns out to be that your site is hard for anything to read — slow, built as images, no real page structure — that is a build conversation first. Either way, tell us the eight questions and we will tell you what we see.

Questions we get asked