How to get cited by ChatGPT: what actually moves the needle

The uncomfortable finding#

Almost everything written about getting cited by ChatGPT is advice about your own website. Tidy your headings. Add an FAQ. Write in short, quotable sentences. All reasonable, all marginal — because that is mostly not where the model is looking.

An analysis of 21,311 brand mentions across ChatGPT, Claude and Perplexity found that 85% of them originated on external domains, and that brands were about 6.5× more likely to be mentioned via a third-party source than via anything they published themselves. [1]

Read that as a budget instruction. If five-sixths of the mentions come from places you don’t control, then five-sixths of the effort belongs there too — and almost nobody allocates it that way, because on-site work is the part you can finish this week.

Where the answers actually come from#

When an assistant is asked “what’s the best X for Y”, it reaches for sources that look like consensus rather than like marketing. In practice that means:

• Reddit and forums. Unpolished, opinionated, dense with real comparisons. Heavily represented in answers about tools and purchases.
• Review directories. G2, Capterra, AlternativeTo, Product Hunt, SaaSHub and the category-specific ones. Structured, comparative, and trivially parseable.
• Wikipedia and knowledge-graph entries. Wikidata, Crunchbase. These anchor whether you exist as an entity at all.
• Trade press and niche authorities. One paragraph in the publication your industry actually reads outperforms a year of your own blog.
• YouTube. Transcripts are text, and comparison videos are exactly the shape of a buying question.

Note what these have in common: someone else wrote them. That is not an obstacle to route around, it is the mechanism. A model treats a third party describing you as evidence and treats you describing you as a claim.

What to do, in priority order#

1. Get listed where your category gets compared. Claim the directory profiles that come up for your space, and make sure the description matches the language buyers use rather than your internal positioning. This is slow, unglamorous, and the highest-leverage work available.
2. Be present in real discussions. Not astroturfing — answering questions in the places your buyers ask them, under a real name, including when the honest answer is that you’re not the right fit. Threads outlive campaigns.
3. Publish original data. The single strongest citation magnet: if you publish a number that exists nowhere else, an engine that wants to use it has to attribute it. A survey of 200 customers is a citation asset for years.
4. Add schema to the pages you want quoted. Pages carrying structured data see roughly 2.8× the citation rate. Fully in your control, done in an afternoon.
5. Refresh, don’t just publish. Content updated within the last 30 days gets around 3.2× more citations than stale pages. Update the five pages that matter before writing a sixth.
6. Let the crawlers in. Check your robots.txt actually permits GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and Google-Extended. Blocking them is a common own-goal, usually inherited from a security-hardening ticket nobody revisited.

Think in topics, not keywords#

A study of 50,000 brands in ChatGPT found visibility behaves as a topic-level property rather than a keyword-level one. [2] The question is not “do I rank for this phrase” but “when the assistant answers anything across the subject my business lives in, does my name keep coming up?”

That changes what good coverage looks like. One page targeting one phrase is a keyword habit. A cluster of pages that between them cover the whole topic — definitions, comparisons, how-tos, the honest trade-offs — is what earns topic-level presence. That is also the structure that survives a model update, because it isn’t betting on one phrasing.

How to tell whether any of it worked#

You cannot see this in Google Analytics. There is no impression to log and often no click to attribute, which is precisely why the funnel goes unmanaged.

What you can do is ask the engines the questions your buyers ask, on a schedule, and record whether you were named, where in the answer, and which sources the answer leaned on. That last column is the actionable one — it is a list of pages to go get mentioned on. The five numbers worth tracking goes through it properly.

Sources, verified August 2026: Reditus on the 21,311-mention analysis, Semrush on topic-level authority across 50,000 brands, Goodie, LLM Pulse and Apollo Digital on citation strategy. Multipliers are the figures reported by those analyses; they describe correlations across sites, not a guarantee for yours.

One thing not to bother with first#

Publishing an llms.txt is the current default advice, and it is much weaker than it sounds — Perplexity reads it, and no other major provider has confirmed they do. We wrote up the full evidence, including why we still generate one. Do it fifth, not first.