Contents
- ChatGPT doesn't rank pages, it composes an answer and cites a few sources: the logic isn't a search engine's, and neither are the levers.
- Two paths lead to a citation: live web search, which depends on your pages, and knowledge acquired in training, which depends on what's said about you elsewhere.
- You don't influence an individual answer, you influence the raw material: short, self-contained, sourced and dated passages.
- Blocking AI engine crawlers in robots.txt guarantees no citation through the web-search path. It's a trade-off, not a default setting.
- A single observation proves nothing: two sessions on the same question give different answers. Only a repeated, monthly measurement produces a usable signal.
How ChatGPT picks its sources
Two mechanisms coexist in a single interface, and they don't call for the same actions. Understanding which one is at work on a given question is the first useful reflex.
- Your pages are cited by name
- Fast effect, sometimes within days
- Worked with SEO levers
- Depends on crawler access to your site
- Three to five sources kept per answer
- Strong implicit recommendation
- Doesn't depend on your site
- Slow effect, old corpus
- A wrong fact is hard to correct
- No citation, so no traffic
This second path explains a situation that puzzles most clients: a company recommended by ChatGPT without its site ever appearing in the sources. Conversely, a very well-ranked site can be cited on specific questions without ever being recommended on the general questions of its market.
What makes a passage citable
Nothing guarantees a citation, and you should distrust anyone who claims otherwise. But the passages actually reused share four observable traits.
- Short and self-contained
A one-sentence definition, a formula, a figure with its source. A model reuses a passage, never a whole page: a paragraph that only makes sense with the three before it won't be extracted.
- Verifiable
A figure with no source has little chance of being reused, because the model has no way to back it up. Naming the source in the sentence beats relegating it to a footnote.
- Tied to a stated question
A heading written as a question, followed immediately by its answer, is markedly more reused than an allusive heading. It's the most profitable format change, and it costs nothing.
- Dateable
On evolving topics, content carrying a visible update date is favored. A page with no date on a moving topic is treated as uncertain.
The four actions that have an effect
In order of observed effect, from the most direct to the slowest.
Write the passages you want cited
Identify the questions a prospect would actually ask, and answer them in a self-contained paragraph, with figures where possible. It's exactly the logic of a definition entry: a term, an answer, an example. Our glossary of marketing metrics is built on this principle.
Exist somewhere other than your own site
Industry comparisons, serious professional directories, guest articles, specialist community discussions. The model sees these as independent sources, which often weighs more than a page on your own site. It's the main lever of the second path, the recommendation one.
Let the crawlers reach the site
Check your robots.txt file. Blocking the AI engine crawlers, GPTBot and OAI-SearchBot for OpenAI, guarantees no citation through the web-search path. It's a legitimate trade-off for some publishers, but it has to be deliberate: OpenAI's documentation on its crawlers spells out which does what.
Correct what's wrong
If ChatGPT says something inaccurate about you, there's no correction form. The only path is to publish the correct information where the model goes looking, and wait. Count several months on an established topic.
What doesn't work
| What you hear | How it actually is |
|---|---|
| Pay to appear in the answers | There's no advertising space equivalent to search ads in generated answers. |
| Install an llms.txt file | Emerging standard, no demonstrated effect to date. A low-cost bet, not a solution. See our dedicated article. |
| Stuff the page with keywords | Counterproductive: selection is about the clarity of a passage, not its lexical density. |
| Publish more articles | Volume doesn't make up for the absence of self-contained passages. Ten well-structured pages beat a hundred diluted ones. |
| Guarantee a citation | Answers vary from one session to the next on the same question. No guarantee holds. |
On the llms.txt file in particular, the debate deserves better than a blunt verdict: we detailed it in our article on llms.txt and its real usefulness.
Measure rather than assume
You don't know your visibility in ChatGPT without measuring it, and intuition is a particularly poor guide here: the first question you ask is always the one whose answer you already know.
- Build a set of real questions
About fifteen questions as a prospect would ask them, not as a marketer would phrase them. Mix brand questions, category questions and comparison questions.
- Ask each question several times
Three passes minimum, in separate sessions. The variability between two answers to the same question is the main trap of the exercise.
- Record what's said, not just whether you appear
Is the brand cited? In what terms? Which competitors appear in your place? That's often the most instructive answer.
- Repeat every month
A single reading describes a moment. It's the series that shows whether the editorial work is producing an effect.
The full method, applied across the different answer engines, is detailed in our article on brand visibility in AI answer engines.
Automate the measurement with the GEO Tracker
Done by hand, the method above takes two to three hours a month per client. For an agency tracking fifteen accounts, it isn't sustainable, which is why almost no one does it.
The Adwise GEO Tracker runs exactly this protocol, automatically. It queries ChatGPT, Perplexity, Gemini, Claude and Mistral on the set of questions defined for a Workspace, repeats the queries, and produces a report: the engines that cite you, the ones that don't, the competitors that appear in your place, local presence, and the change from one audit to the next.
The point for an agency isn't technical curiosity. It's being able to answer with a figure and a series when a client asks whether they're talked about in ChatGPT, instead of a screenshot taken that morning. The tools are detailed on the site.
Frequently asked questions
The Adwise team connects Google Ads, Meta Ads, LinkedIn Ads, GA4 and Search Console so agencies and freelancers can run their campaigns from a single place.
