Short answer: when ChatGPT cites a source, it has usually just searched the web, fetched a handful of pages, and picked the ones that most directly answer the question it was given. It is closer to a research assistant with a browser than to a ranking algorithm.
Longer answer below, with an important caveat up front.
OpenAI does not publish selection criteria, and the behaviour changes between model versions. Everything here is observed behaviour and documented crawler policy, not a leaked algorithm. Anyone showing you a weighted list of ChatGPT ranking factors has made it up.
Two different mechanisms, often confused
ChatGPT answers from two places, and the difference decides whether you can influence it.
From memory, with no citation
Ask it something general and it answers from what it absorbed during training. There is no source list because there is no live fetch. If it says something wrong about your product here, you cannot correct it directly — you can only hope the next training run sees better information.
From a live search, with citations
Ask something current, specific, or obviously factual and it searches, reads, and links. This is where the citations come from, and it is the half that responds to your work.
What appears to get a page picked
It can be fetched
This is the gate everything else sits behind. If OAI-SearchBot is disallowed in your robots.txt, or the content only renders after JavaScript that a fetcher does not run, or it is behind a login, the page cannot be quoted. A surprising share of “we are invisible in AI search” turns out to be this.
It answers the question asked, not an adjacent one
A page titled “our approach to content automation” loses to one titled “does this plugin publish automatically”. The second matches the question. Match the question people ask, in their words.
The answer is near the top and in plain prose
Fetched pages get read, but not lovingly. A claim buried in the ninth paragraph, or expressed only in a graphic, is a claim that does not make it into the summary.
It is not contradicted by your other pages
If two of your pages disagree, the model may state whichever it read, or hedge, or cite a competitor who was consistent. Self-consistency is an underrated ranking factor because it is not a ranking factor at all — it is just not giving the model a reason to distrust you.
It looks like it was written by someone who knows
Specific numbers, named limitations, honest trade-offs. Pages that admit what a thing does not do read as more reliable, and get quoted in exactly the comparisons where that matters.
What does not seem to matter much
- Keyword density. There is no evidence of term-frequency weighting in the way classic SEO tools model it.
- Domain age. New pages get cited when they answer the question better.
- Word count. A tight 600-word answer beats a padded 3,000-word one. Padding actively hurts, because the quotable sentence is harder to find.
- Backlinks, directly. They plausibly feed the underlying search index, so they matter indirectly, but a link-building campaign is not a GEO strategy.
How to check your own position
There is no rank tracker for this, so do it by hand and write the answers down.
- List twenty questions a buyer would type before choosing something like your product.
- Ask each one in a fresh chat, with search enabled.
- Record: were you mentioned, were you cited with a link, and was what it said correct.
- Repeat monthly. The third column is the one that will surprise you.
That last column matters more than the first. A confident, wrong statement about your pricing does more damage than not being mentioned, and the fix is usually a page of yours that was ambiguous.
Common questions
Does ChatGPT use Google’s index?
It has its own crawlers and its own search infrastructure, and the arrangements behind commercial search products have changed more than once. Treat “which index” as an implementation detail you cannot depend on, and make sure you are crawlable by the named OpenAI agents either way.
Why does it cite a competitor instead of us when we have the better page?
Usually one of three things: their page answers the literal question and yours answers a nearby one, your claim is buried below the fold, or your own pages contradict each other and yours was the less confident source. Check in that order.
Can I ask ChatGPT to correct something wrong about my company?
Not persistently. A correction inside one conversation does not carry to anyone else’s. The durable fix is to publish the correct information clearly on your own site, where a retrieval pass will find it, and to get third-party pages that repeat the error updated.
Do citations send real traffic?
Some, and less than a search result in the same position would. Treat a citation as brand visibility with a modest click-through rather than as a traffic channel, and judge it accordingly.
Should I block OpenAI’s crawlers?
Only if you would rather not appear in its answers, because the two go together. It is a commercial decision. If your content is your product and you sell access to it, blocking is defensible; if you sell something else and your content is marketing, blocking removes you from the shelf.
Where to go next
GPTBot vs ClaudeBot vs PerplexityBot covers which agents to allow and what each one is for. How to make a website more visible in AI search is the checklist.
Did this answer it? Tell us if not and we will fix the article.