Getting cited by ChatGPT and AI search

Two routes in, one of them open

A model can know your company two ways. It read about you while it was being trained, or it went and looked while it was answering the question.

The first route is closed to you. The weights of a shipped model are frozen, the crawl that produced them happened months ago, and the next one is not something you can schedule or influence.

If your company came out of stealth this year, most models have never encountered your name at all, and no amount of publishing changes that until somebody trains a new one.

The second route is live. When a buyer asks which agent frameworks handle long-running tasks, the system runs searches, pulls a set of pages, and writes the answer out of what it just read.

That set gets decided in the moment, on the open web, from whatever is reachable and relevant. It can include something you published last week.

Everything worth doing sits in the second route. The rest of this page is about that.

What the crawler can actually reach

Answer engines send named crawlers. OpenAI runs GPTBot and OAI-SearchBot, Anthropic runs ClaudeBot, and Google separates its AI access under Google-Extended. Each of them obeys robots.txt, which means a line written two years ago by someone worried about scraping can remove you from every answer in the category.

Check the file first. It takes a minute and it is the only item on this page that can be silently costing you everything.

• Blocked crawlers. A disallow on GPTBot or OAI-SearchBot ends the conversation. There is a real argument for blocking training crawlers and no argument at all for blocking the search ones, which is the distinction most robots.txt files miss.
• JavaScript-rendered content. Several of these crawlers fetch HTML and do not run your app. If the page is empty until React fills it in, the crawler sees an empty page.
• Gated pages. A comparison guide behind an email form is invisible. So is anything behind a login, a cookie wall or an interstitial.
• PDFs and slide decks. Retrieval on these is inconsistent. If a claim matters, it belongs in HTML as well.

The X problem, stated plainly

This is the part that matters most to anyone who has just run a launch, and we would rather say it than let a client discover it.

Posts on X are largely unreachable by these systems. The platform restricts automated access to its content, so the thread that did two million views is mostly invisible to whatever writes the answer when your buyer asks the question three weeks later.

A launch can be the biggest thing on the platform for a day and leave close to no trace inside an answer engine directly.

What it leaves is everything downstream of it. A launch that lands produces newsletter write-ups, blog posts, forum threads, video breakdowns, roundups, and a run of other people describing your product in their own words on pages that crawl perfectly well. Those pages are what gets retrieved.

So the mechanism runs one step removed. The coverage a launch causes is the thing that gets quoted, which means the citation footprint you end up with is roughly proportional to how many people felt the need to write something after seeing it.

Reach is the input and the write-ups are the output, and the gap between them is a few days.

What gets quoted

Retrieval works on passages rather than pages. The system pulls a chunk of text, decides whether it answers the question, and either quotes it or moves on. A page can be relevant end to end and contribute nothing if none of its sentences stand up alone.

The working test: could a sentence be lifted out of your page, dropped into an answer with no editing, and still be true and attributable?

Every sentence in the right column names the subject, states one fact, and survives being read with no surrounding context. That last property is the whole thing. The system has no idea what paragraph came before.

This is also why the question-and-answer format works so well here, and why it is worth putting the answer in the first two sentences under each heading rather than building to it. A page that opens each section with its conclusion gives the retriever something to take on every pass.

Your site is one vote

Answer engines synthesise across several sources at once. When six of them describe your company one way and your homepage describes it another, the six win, and the answer your buyer reads is the one the six agreed on.

This is the part teams underestimate. You can write the perfect positioning page and still be described as something else entirely, because the model is averaging over everyone who has ever written about you. Correcting a description that has settled is considerably harder than establishing the right one while the category is still forming.

Which makes narrative consistency an infrastructure decision rather than a copy decision. If every creator, article and founder post lands on the same sentence about what the company does, the model has one stable entity to reason about.

If eleven people describe you eleven ways, you get described vaguely, or you get described as your competitor's category. Narrative design is where that sentence gets decided, and a launch is the moment where the largest number of people write it down at once.

What paid placements are now worth

Worth addressing directly, because the calculation on syndicated and sponsored articles changed and most teams are still pricing them the old way.

A paid placement on a news domain used to buy a backlink, a logo for the press page, and some search authority. That was a modest return and the category knew it.

What it buys now is an additional independent-looking source in the retrieval pool: a crawlable page, on a domain with authority, stating in the third person that your company does a specific thing. That is precisely the shape of source an answer engine pulls when somebody asks the category question.

So the honest position is a split one.

• Strong for machines. A retrieval system weighs the domain and the phrasing rather than the commercial arrangement behind it. Several placements saying the same sentence about you contribute to the consensus described above, and they do it faster than earned coverage.
• Weak for humans. Any experienced buyer recognises the format on sight: no byline, the company name repeated as anchor text, a superlative the publication would never write itself. A placement asserting that you are the best in your category is owned media wearing a news domain's clothes, and it gets discounted accordingly.

Which gives a straightforward buying rule. Price them as citation-footprint infrastructure rather than as persuasion, keep the copy factual enough to survive a reader who knows what they are looking at, and label them as sponsored wherever the publication's policy and your own advertising obligations require it.

A placement written as a plain description of what your product does is worth more in both directions than one written as a press release with a superlative in the headline.

They are a supplement rather than a substitute. Twenty people writing about you because a launch gave them a reason will always beat six pages you bought, because the first group chose their own words and the second group used yours.

If you invented the category

New-category companies have a specific version of this problem. Nobody is asking for your term because nobody knows it exists, and no source defines it because you are the only one using it. You get retrieved, if at all, for the category you are replacing.

Two moves, and the order matters.

1. Describe yourself in the old category's language on at least one crawlable page. A buyer asking about the thing you are replacing is your buyer. A page that says plainly how you compare to it is the page that gets pulled into that answer. This costs some pride and it is the fastest route to being in the set at all.
2. Get other people using your term. Nothing you publish about your own category counts as evidence. The model needs to see the word in other people's writing, used naturally, by people with no obvious stake. That is a distribution problem with a familiar shape.

What to fix on your own pages

The site half of this is cheap and finite. It will not carry the whole job, and skipping it means the coverage you earn points at a page that cannot be quoted.

• One page per question a buyer actually asks, with the answer in the opening two sentences. Not one page per keyword.
• Answers in visible text. Structured data helps machines parse what is there. It does not substitute for the words being on the page, and a question that exists only in your schema is a question nobody can quote.
• Server-rendered HTML, for the reason in the section above.
• Published and updated dates, and a named author. Recency shows up in what gets selected, and an undated page is a page of unknown age.
• The same one-line description of the company everywhere. Site, about page, LinkedIn, docs, press kit. Word for word, so nobody paraphrases you into someone else's category.
• Comparison and pricing content ungated. These are the two things a buyer asks an answer engine about most, and the two things companies hide most often.

On llms.txt: it is a proposal for a plain-text file describing your site to models, adoption across the engines is thin, and it takes ten minutes. Add it if you like. Do not expect it to do any of the work above.

How to tell whether it is working

There is no Search Console for this yet, and anyone selling you a confident number is selling you a modelled estimate. Two reads are real.

Referral traffic. Analytics will show sessions arriving from chatgpt.com and the other answer engines. Volumes are small next to organic search and the intent is unusually high, because the visitor has already read a summary and is clicking through to verify something specific.

Asking the questions yourself. Write down the ten questions your buyers actually ask, run them across the engines that matter to you, and log which companies get named.

Do it monthly. It is manual, it is the only direct read available, and it is more informative than any dashboard because you see the sentence the model uses to describe you.

One caution that catches everybody: answers vary between runs. The same question asked twice retrieves different sources and produces different names. Ask each question ten times before you conclude anything, and track the share of runs you appear in rather than whether you appeared.

What you will usually find on the first pass is that the companies getting named are the ones people have written about most, which is the same finding as every other channel. Answer engines are downstream of what the internet says about you. Changing what gets cited means changing how many people write it down.

Questions we get asked

Can we get our company into ChatGPT's training data?

Not on any timeline you can plan around. The weights of a shipped model are frozen and the next crawl is not something you can schedule. Every practical move sits in the retrieval path instead, where the system runs a live search while answering and can pull a page you published last week.

Do X posts get cited by answer engines?

Mostly no. X restricts automated access to its content, so a thread that reached two million views is largely invisible to the systems writing these answers. What the launch leaves behind is the reachable part: newsletter write-ups, blog posts, forum threads and video breakdowns published by other people on crawlable pages.

How do we know whether an answer engine is citing us?

Two reads, and neither is a dashboard. Analytics will show referral traffic from chatgpt.com and the other answer engines, in small volumes at high intent.

Beyond that you ask the questions your buyers ask, across several engines, and log who gets named. Answers vary run to run, so ask each question ten times before believing the result.