How do you appear in ChatGPT?

How does ChatGPT answer a question?

There are two mechanisms, and conflating them produces most of the confusion around this subject. In the first the model answers purely from its weights — from what it learned during training; it visits no site and gives no link. In the second it searches, fetches a handful of pages, summarises what it read, and shows the sources as links.

What triggers a search is usually the question itself: anything asking for current information, anything local, anything asking for a company or a price comparison sends the model outside. "What is physics" needs no search; "can you recommend a corporate identity agency in Istanbul" does. Almost every commercially intended question falls into the second group — and that group is precisely the one that is worth anything to a brand.

The conclusion sets the whole strategy: the target is not to be knowledge inside the model, but to be the best source it finds when it goes outside. The first is not in your control; the second is entirely in your control.

Which crawler does ChatGPT read your page with?

OpenAI documents four separate crawlers, and only two of them decide visibility. OAI-SearchBot indexes the sources shown when ChatGPT searches. ChatGPT-User fetches the page at the moment a user opens a link in an answer. GPTBot is for model training only, and OAI-AdsBot runs safety checks on advertising landing pages.

In practice the common failure is a decision to "close the AI bots" that closes all four. The result is trading absence from training data for absence from answers — a trade that is right for almost no company. Which group does which job, and how to decide, is laid out as a table in the robots.txt guide.

Can getting into the training data be a target?

No — and it needs saying plainly, because part of what gets sold promises exactly this. Which pages enter a training set is decided by the model provider, the process is not transparent, there is no route to apply for inclusion, and nobody knows how many months out the next model is. It has no property you can plan against.

What is controllable is being readable and quotable when the model searches. That has three components, all three measurable: is crawler access open, does the page carry an answer passage that survives being cut out of context, and does the page tell a machine which organisation it belongs to. The rest of this guide is about building those three.

What does a passage ChatGPT quotes look like?

A generative engine does not read a page from start to finish. It splits the page into passages, scores each one against the query on its own, and quotes only the strongest. So a passage that needs its neighbours is worthless: a sentence saying "as described above" loses its meaning the moment it is lifted out.

NOT QUOTED
----------
As we mentioned above, our many years of experience in this field
allow us to offer our clients the very best service. As for pricing,
it may vary depending on the scope of the project.

  The problem: no subject (who is "we"?), no numbers, no question,
  and the first sentence leans on the paragraph before it.

QUOTED
------
An identity project covers logo design, a colour and typography
system, an application manual and file handover. At Galata Media it
runs about six weeks and consists of three presentations: discovery,
direction, application. Price follows the number of applied surfaces.

  Why it works: the subject is named, the scope is enumerated, the
  duration and structure are concrete, and the passage stands alone.

The heading above the passage has a job too: before scoring the passage, the engine matches the heading text against the query. Section headings should therefore be questions, not topics. "Pricing" is a topic and resembles no query; "How long does an identity project take?" is a question, in the words someone actually typed.

Is your brand an entity to the model?

Before an engine can recommend your brand it has to resolve it as an entity: which company this name belongs to, where it is, what it does, which records mention it. If the same organisation's address is written one way on the website, another in the Google business listing and a third on LinkedIn, the engine sees three weak records and trusts none of them.

• Name, address and phone (NAP) must be written identically on every surface — abbreviations, punctuation and order included.
• schema.org Organization markup must stand as one identity across the whole site, and its sameAs field should list verifiable profiles.
• If the organisation belongs to a group or parent company, name it: a second independent record makes the entity markedly easier to resolve.
• Being mentioned in third-party sources outweighs what you write on your own site; an engine knows a brand by what others write about it.
• Founder and team names kept consistent with their own profiles tie the organisation into a network of people, and that network is itself a verification channel.

How do I measure whether I appear in ChatGPT?

Most organisations "measure" this by typing their own brand name once, then relax or panic at the answer. Both reactions are wrong, because that is not a measurement. A brand named in the question appears in the answer; all you measured was the model's reading ability. Real measurement uses questions that never name the brand, and repeats them.

The resulting number is simple: in how many of the twenty questions did your brand appear. On a first measurement that ratio is zero for most organisations, which is not bad news but a baseline. Running the same set under the same rules three months later is the only honest way to show what the work bought.

A step-by-step checklist for ChatGPT visibility

1. 01

   Verify OAI-SearchBot and ChatGPT-User access

   Reading robots.txt is not enough; send a real request with the bot's user agent and look at the code that comes back. A 403 means the problem is not in the file but in the firewall.

2. 02

   Put a 40–75 word answer passage on every page that matters

   The passage must name its subject, use no pronouns, and answer directly the question the page exists to answer. Keeping it near the top of the page helps.

3. 03

   Turn section headings into questions

   Not "Our services" but "What does an identity project cover?". The heading is matched against the query before the passage is scored.

4. 04

   Set up Organization and WebPage markup

   Markup ties the page to an organisation and turns that organisation from a string of text into a resolvable entity. The detail is in the schema markup guide.

5. 05

   Write the frequently asked questions into the page

   Each question a heading, each answer an independent 40–75 word passage. Printing them open rather than hiding them inside an accordion makes the passage boundary clean.

6. 06

   Attach a source to every number

   An unverifiable statistic costs the whole page: an engine does not quote around a claim it cannot confirm, it skips the page.

7. 07

   Take the baseline measurement and date it

   Run the twenty questions under the five rules above and record the result. Without that number, nobody can say what changed three months later.

What does not work?

• A service that "adds your site to ChatGPT". There is no submission form and no directory; the only sellable work is opening access and making the content quotable.
• Keyword repetition. A generative engine matches on whether the passage answers the question, not on word counts; repetition only makes the passage unreadable.
• Hundreds of bulk-generated pages. Passages that copy one another suppress each other, and the single strong passage an engine would have picked disappears.
• Publishing llms.txt and nothing else. The file opens no access, writes no passage and builds no entity — and Google states outright that AI features require no such file.
• Hidden text, or claims visible only to machines. The rule that everything in the markup must also be visible on the page is precisely where those attempts get penalised.