An engine’s picture of your brand comes from two different places with two different update speeds: a training corpus weighted toward third-party writing about you, and live retrieval of whatever page it can fetch right now. Your own marketing pages contribute far less to the first than teams expect.
You do not have to guess which sources carry you. One of them is reported to you for free, and the rest are readable in the answers themselves.
What are the two pipelines?
They are separate systems, and the vendors document them separately. OpenAI’s crawler documentation runs two different bots for them: “OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features,” while GPTBot “is used to crawl content that may be used in training our generative AI foundation models.”
So a page of yours can be reachable to one and not the other, and the consequences differ. Blocked to the search crawler means you cannot be cited today. Absent from the training corpus means the model has no memory of you when nobody fetches anything.
Everything published about you before a cutoff. Third-party sources outnumber your own pages by a wide margin and repeat each other, which weights them further.
Your own pages compete on equal terms here, provided they are indexed and answer the question being asked.
The asymmetry in the left panel is the part teams underestimate. You publish a few dozen pages about yourself. The rest of the internet publishes reviews, comparisons, forum threads, job listings and news about you, and in aggregate that is simply more text.
Why do third-party sources carry more weight?
Volume and repetition, mostly, rather than any judgement about credibility. A fact repeated across many independent pages is better attested than the same fact stated once on your own site, and models absorb attestation.
There is also a structural reason your own pages underperform in the training half. Marketing copy states positioning rather than facts, and positioning has no stable form to absorb. “The best choice for growing teams” cannot be stored the way “founded in 2019” can, which is the storage argument in what is a knowledge graph.
The internet writes more about you than you do, and it writes in a form models retain more easily.
How do you find out which sources actually carry you?
Three methods, in descending order of reliability, and the first is free and first-party.
Microsoft’s AI Performance report is the unusual one. It shows “the total number of citations that are displayed as sources in AI-generated answers,” page-level citation counts, and “the key phrases the AI used when retrieving content.” That is a census of one surface rather than a sample of many, and it costs nothing.
For everything else you sample. Record every cited URL from your prompt runs, group them by domain, and you have your own citation distribution, measured on your own category rather than borrowed from someone else’s aggregate.
What should you do with the answer?
Sort the sources into three piles, because each takes different work.
| Source type | Can you change it? | The work |
|---|---|---|
| Your own pages | Directly | Make the fact plainly stated, indexed, and snippet-eligible. Fastest feedback loop available. |
| Third-party listings and comparisons | Sometimes, by asking | Correct the entry, or get added to it. Slow, unglamorous, and the highest-yield work in this discipline. |
| Community threads and archives | Rarely, and never by editing | Treat as a signal about how you are described, not a surface to manage. Attempting to manage it is usually visible and counterproductive. |
The middle row is where most of the available movement sits. A comparison page that carries you into answers is worth more than a blog post nobody retrieves, and getting listed on one is a request rather than a campaign.
Does the source mix differ between engines?
Yes, and assuming otherwise leads to wasted work. Engines run separate retrieval over separate corpora, so the page that carries you into one answer may be invisible to another.
Google states the divergence for its own two surfaces, noting that “AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.” If two surfaces from the same company differ, two surfaces from different companies certainly do.
The practical consequence is that a source list has to be built per engine rather than pooled. A domain appearing in half your ChatGPT citations and none of your Perplexity citations is not a weak source, it is a source one engine reaches and another does not, and those are different problems with different fixes.
It also means a single-engine audit will mislead you about where to spend. Check at least two before concluding that a particular site is or is not carrying you.
Why do your own pages underperform in retrieval too?
Usually because they are not eligible rather than not good. Google’s documentation on AI features sets the requirement plainly: a page “must be indexed and eligible to be shown in Google Search with a snippet,” and there are “no additional requirements to appear in AI Overviews or AI Mode.”
The second most common reason is form. Anthropic’s citations documentation describes returning “the exact passages that support each claim,” so a fact needs to exist as a passage. A price inside a styled table or an integration shown only as a logo has no passage to return.
Check eligibility first, then check whether the fact exists as a sentence. Those two account for most cases where a company has the right page and still never gets cited from it, and the wider version of that diagnosis is in entity problem vs content problem AI visibility.
How often should you recheck?
Quarterly for the source mix, because it moves slowly, and continuously for the answers themselves. A source that carried you for a year can be rewritten in an afternoon by someone who has never heard of you, and the first sign is usually a drop confined to a few prompts across several engines, which is the shape covered in AI visibility dropped suddenly.
Keep the cited URLs from every run rather than only the counts. The list of who carries you is an asset that takes months to build and one export to preserve, and it is the input to almost every decision about where to spend effort next.