How to Get Cited by ChatGPT

Citation is three gates, not one

Getting cited means passing three separate gates, and almost all published advice addresses the middle one. Retrieval decides whether your page is in the set being considered. Selection decides whether your passage is the one used. Attribution decides whether your name travels with it.

These fail differently and they fail invisibly. A site that is never retrieved and a site that is retrieved and passed over produce the same outcome, which is a paragraph naming somebody else. Nothing distinguishes them from outside, and the fixes are not related.

Retrieval is largely mechanical. Can the page be fetched, is it about the subject in a way that is legible, is there enough on the subject that reaching for you is reasonable. Selection is editorial. Is there a passage that answers cleanly and stands on its own. Attribution is partly out of your hands.

Work in that order. Selection work on a site that is never retrieved changes nothing at all, and it is where most effort goes because it is the part that feels like writing.

Gate one: being in the set at all

Three things decide whether you are considered, and the first is depressingly simple.

The page has to be fetchable by the right client. Answer crawlers are named clients and they are governed separately from search crawlers. A site can be entirely open to search and closed to answers, and the closure is usually accidental, arriving with a security product or a hosting default rather than a decision.

The page has to return its content without a browser. Fetch it as a plain document and read what comes back. If the substance appears only after scripts run, it may not be available to whatever is collecting it. This catches a whole category of modern site build, and the businesses affected have usually never checked.

There has to be enough about the subject. Retrieval is a judgment about whether reaching for you makes sense, and one page on a topic is thin. Fifteen connected pages answering the surrounding questions changes that judgment in a way one excellent page does not.

This third point is where the real cost sits and where the temptation to substitute breadth for depth is strongest. Forty subjects covered once produces forty thin cases for retrieval. One subject covered forty times produces a strong one.

Gate two: being the passage worth using

Once retrieved, your page competes with everything else retrieved, and the competition is at the level of the passage rather than the page.

State the answer early. The first two sentences of a page should answer the question in its title. Commercial writing habitually delays the statement until context has been laid, which serves a reader going in order and hurts anything assembling a short reply.

Make claims that survive extraction. A sentence beginning with "this" or "it", a comparison with an unstated other half, a figure with no unit. All of them read perfectly in place and become meaningless once lifted. Writing self-contained sentences is a habit rather than a technique and it is learnable in an afternoon.

Put the substance in prose. Specifications inside tabs, terms inside accordions, guidance inside a modal. Keep a plain summary in text and let the interface hold the detail.

Be specific enough to be worth quoting. A sentence that could appear on any competitor's site adds nothing to an answer. A sentence containing a real constraint, a real number you can stand behind, or a real trade-off is worth reaching for. Generic writing is not passed over for being badly written. It is passed over for being interchangeable.

Gate three: being named for it

The third gate is the one you control least, and being honest about that is more useful than pretending otherwise.

An assistant can use what your page established without naming you. It can synthesise a statement from several sources where no single one is credited. It can name a source that itself learned the thing from you. None of that is within your control and none of it is unusual.

What appears to help is being the identifiable origin of a specific claim rather than one restatement among many. If a number, a distinction or a framing traces to you and is phrased in a way that is recognizably yours, being named with it becomes more likely than if you are the fourteenth site to describe a widely known idea.

This argues for original work over comprehensive summary, which is an uncomfortable conclusion because summary is cheaper and ranks perfectly well. A thorough overview of a settled subject is a good page and a poor citation candidate, because there is nothing in it that needs sourcing to you specifically.

It also argues for consistency of phrasing. If you have a distinctive way of framing something, use the same framing everywhere rather than varying it for freshness. Variation is good writing advice and works against being recognized as the origin.

What does not work

Several tactics circulate and are worth dismissing explicitly, because each one costs real effort.

Writing to the assistant rather than to a person. Pages stuffed with question headings and one-line answers read as thin to a human and are not preferred by anything else. The qualities that make a passage quotable are also the qualities that make it clear, and there is no separate machine dialect to write in.

Publishing at volume. More pages on more subjects makes each subject thinner. This is the opposite of what retrieval rewards and it is the most commonly commissioned response.

Chasing the model rather than the subject. Assistants change. A page written to exploit a current behavior is a page that ages badly, and the effort would have compounded if spent on depth.

Asking to be included. There is no submission process. There is no directory to be added to that changes this. Anybody offering placement is selling something other than what they claim.

How to tell whether any of it worked

This work is unusually hard to measure and the failure to set up measurement first is why so many programs end in disagreement rather than a result.

A person who reads your name in an answer and then searches for you arrives as direct traffic. The assistant leaves no fingerprint. So the measurement has to be done by asking rather than by watching your analytics, and that has to start before the work does.

Fix a set of between six and ten questions phrased the way a buyer would phrase them. Ask each one several times. Record what came back, including which sources were named when the assistant names them. Do this before you change anything, because you cannot reconstruct a baseline afterwards.

Then repeat monthly. What you are watching is whether your name appears more often across a fixed set, not whether a single answer changed. Single answers change on their own, and reading a single change as a result is the fastest way to draw a false conclusion.

Assistants that show their sources are worth more than others for this, because you can see which passage was taken and whether it was the one you rewrote. That is the closest thing to feedback available in this work.

The honest trade-off

The uncomfortable part of this is that the two halves reward different behavior and both are correct.

Retrieval rewards depth, which means concentration. Picking one subject and covering it thoroughly, at the cost of the other subjects you could have covered. That is a strategic commitment and it is genuinely risky, because if you pick a subject nobody asks about you have spent a quarter well and gained nothing.

Selection rewards clarity and specificity, which sometimes costs you rhetorical room. A page that states its conclusion in the first two sentences gives up the build. For some subjects that is a real loss and worth accepting anyway.

What neither rewards is the middle path most sites take, which is moderate coverage of many subjects written in a house style that circles before committing. That path is comfortable, it is what most content calendars produce, and it is close to invisible here.

There is a further cost worth naming honestly. Everything described on this page improves your standing in a channel you cannot measure directly, on a timescale of months, with no guarantee attached. That is a genuinely difficult thing to fund against alternatives that report a number weekly. The argument for doing it anyway is that the work is largely the same work that makes a site good for people, which means the downside case is a clearer site rather than a wasted quarter. Very little else in marketing has that property, and it is the strongest thing that can honestly be said for it.

What to actually write, in practice

The advice above is easier to agree with than to act on, so it is worth being concrete about what a page that clears both gates looks like.

It opens with the answer stated flatly in one sentence. Not a summary of what the page will cover, and not a scene-setting observation about how the industry is changing. The claim, in plain words, before anything else.

It then earns that claim over several sections, each of which handles one idea and is titled with what it establishes rather than with a hook. A reader skimming the headings should be able to reconstruct the argument without reading the paragraphs.

It contains at least one thing nobody else can say. A constraint you have hit, a distinction you draw that others do not, a trade-off you are willing to name. This is what makes a passage worth quoting rather than interchangeable, and it is the part that cannot be outsourced because it comes from having done the work.

It concedes something real. A page with no trade-off in it reads as promotion, and promotional passages are poor material for an answer that is trying to be useful. Naming the case against your own position is the most reliable way to sound like a source rather than a seller.

And it stops. Padding to hit a length adds sentences that dilute the ones worth lifting. Our page on writing for LLMs goes further into the sentence-level habits.

Why the assistant you test on matters

Testing on one assistant and generalising is a common and expensive error, because they differ in ways that change what you should do.

Some show their sources and some do not. The ones that do are considerably more useful during the work, because you can see which passage was taken and from where. That is direct feedback on whether a rewrite landed, and it is the only feedback of that kind available anywhere in this discipline.

They also differ in how much they reach for live material against how much they answer from what they already hold. A change you make will surface quickly in one and slowly in another, and reading the slow one as evidence that nothing worked is a mistake that gets made routinely.

The practical approach is to use a source-showing assistant as your instrument during the work, because it tells you what happened, and to check the others periodically to confirm the effect is general rather than particular. Our pages on monitoring in ChatGPT and monitoring in Perplexity cover how each behaves.

What you should not do is optimize for the quirks of whichever one you happen to test on. Those quirks change, and the underlying requirements of being reachable, being quotable and being the origin of something do not.

Bottom line

Fix retrieval before selection. Confirm the right clients can fetch your pages, confirm the content is in the document without scripts, and commit to depth on one subject rather than breadth across many.

Then make the passages liftable. Answer in the first two sentences, write sentences that stand alone when removed, and keep substance in prose rather than behind interface.

Take a baseline before any of it. Being cited cannot be measured in your analytics, so it has to be measured by asking, and a reading taken after the work has started is not a baseline.