# How to get cited by ChatGPT

The mechanism behind citations in generated answers, and the work that actually moves it. Written for people who have already read the listicle version.

Source: https://martenfield.com/guides/how-to-get-cited-by-chatgpt/
Published: 2026-07-31

---
The listicle version of this article says: write good content, add schema, get
mentioned on other sites. That is not wrong, and it is not specific enough to
act on.

Here is the mechanism, and then the work that follows from it.

## What is actually happening when you get cited

When a model answers a question that needs current information, it does not
recall your page from training. It runs a retrieval step: a search, against an
index, using a query the model wrote itself from your question. It reads some of
what comes back, and it writes an answer citing the pages it used.

Four separate things have to go right, and they fail independently.

**The retrieval query has to be one you can win.** The model rewrites "which
lease abstraction tool should we use" into one or more search queries of its own
choosing. You are not optimising for the user's question, you are optimising for
the queries the model generates from it, which are usually more literal and more
category-shaped than the original.

**Your page has to be in the index behind that retrieval.** For most engines
that means a conventional search index. Classic ranking did not stop mattering,
it changed what it pays out.

**The crawler has to be able to fetch you when it looks.** This is where a
surprising number of companies fail, silently.

**The content has to be quotable.** A model writing a two-hundred-word answer
needs a specific, extractable claim with something that looks like evidence
attached. Marketing prose gives it nothing to lift.

## Start with the failure that costs nothing to fix

Check whether the AI crawlers can reach you at all.

The bots that matter are separate from Googlebot and are frequently blocked by
default. `GPTBot` and `OAI-SearchBot` for OpenAI, `ClaudeBot` and
`Claude-SearchBot` for Anthropic, `PerplexityBot`, `Google-Extended`, and
`Applebot-Extended`.

Two things block them routinely. The first is a `robots.txt` written before
these agents existed, often with a blanket disallow inherited from a template.
The second is a bot-management product. Several CDNs and WAFs ship managed rules
that block AI crawlers by default, and a growing number treat blocking them as a
selling point.

The failure mode is nasty because it is invisible from your side. Your own
testing looks fine. Analytics show nothing wrong. The crawler simply gets a 403
and you never appear.

Check it from outside with the actual user agent, not by reading your
configuration. Then check your CDN's bot settings separately, because
`robots.txt` permitting a crawler does not mean your edge is letting it through.

## Then make yourself worth quoting

A model composing an answer is choosing sentences it can lift with a citation
attached. What gets lifted has a shape.

**Specific figures with a stated basis.** A lease abstraction platform stating a
median tenant improvement allowance, the number of leases it was drawn from, and
the quarter it covers is quotable. "We help you understand TI allowances" is not.

We have deliberately not written a worked example with numbers in it here.
Every page on this site is also served as plain markdown and concatenated into
`llms-full.txt`, so an invented figure inside quotation marks stops being
obviously hypothetical the moment something reads the plain-text version. On a
site whose subject is sentences models lift out of context, that is not a risk
worth taking to illustrate a point this paragraph already makes.

**A stated method.** Sourced, dated, with the sample size named. This is the
difference between a page that reads as evidence and a page that reads as a
claim.

**Direct answers near the top.** The question restated, then answered in one or
two sentences, then the supporting detail. Not a five-paragraph preamble.

**Dates in the visible text.** Not only in the schema. When the data was
collected, and when the page was last updated. Retrieval systems and readers
both discount undated material, and the model frequently repeats the date in its
answer if you gave it one.

**Real tables.** Marked up as tables, not as styled divs, and not rendered by
JavaScript after load.

## The part most people skip

Fixing your own site is necessary and usually insufficient.

Run the queries you care about and write down every source that gets cited.
That list is your source map, and it is almost never your competitors' websites.
It is review sites, trade press, industry associations, forums, and
occasionally one unusually thorough blog post from 2023.

Your competitors are not beating you because their homepage is better. They are
being cited because they appear in the sources the model retrieves and you do
not.

So the work is: claim and correct your profiles on the review sites in your
source map, contribute data to the industry reports that get cited, get your
methodology referenced by the trade press that covers your category, and answer
questions where practitioners actually ask them. Slow, unglamorous, and it is
where the movement comes from.

## Publish the data you already have

The highest-leverage asset most companies own is a dataset they have never
published, sitting in their production database.

Lease comps. Cost benchmarks. Absorption figures. Permit timelines by
jurisdiction. Whatever the operational core of your product produces. Your
buyers ask models questions about exactly this, and the models answer from
whatever public source they can find, which is currently somebody else.

Publishing an aggregated, anonymised, dated slice of it, with a method note, is
the single strongest move available to most companies in this position. It is
also the one competitors cannot copy, because they do not have your data.

## What to measure

Pick twenty to fifty queries your buyers actually ask. Run them across the
engines you care about, three times each, logged out, and record: which
companies were named, in what order, and which URLs were cited.

Repeat monthly. The number that matters is the share of those answers in which
you are cited, and the shape of the source map behind them.

Do not measure whether you rank for "your brand name". You will always win that
and it will tell you nothing.

## The honest caveat

Nobody has a validated causal model here. The engines do not publish retrieval
behaviour, they change it without notice, and a change in your numbers may be a
model update rather than anything you did.

What you can do is remove the blockers, publish material worth quoting, get into
the sources that are already being retrieved, and measure consistently enough to
tell the difference between a trend and a fluctuation. That is the whole
discipline. Anyone selling you more certainty than that is selling.

## Questions

### How long does it take to see a change?

Technical fixes can show up within days to a few weeks once the crawler returns. Source presence work takes longer, typically six to twelve weeks, because a new profile or placement has to be crawled and then has to become one of the pages the retrieval layer prefers. Anyone promising movement in a fortnight is describing the technical fixes only.

### Does publishing more content help?

Volume on its own does not. What helps is publishing something that is genuinely the best available source for a specific question, with a stated method and checkable figures. One page like that outperforms thirty pages of category commentary, because retrieval is picking a source to quote rather than counting how much you have written.

### Can I pay to be cited?

Not directly, and be suspicious of anyone implying otherwise. What you can do is legitimately appear in the third-party sources the models already retrieve for your category, which is slower and is the actual work.