How do answer engines choose their sources?
Mostly they do not answer from memory. They run a search, retrieve a handful of pages, and write the answer out of what those pages say — so the question is not “what does the model know about me” but “which pages does it fetch, and are any of them mine”.
That distinction matters more than anything else on this page. A model’s training data is fixed months before you read this and you cannot edit it. Retrieval happens at the moment somebody asks, against the live web, and it is the part you can affect this week.
The pattern almost all of them share
- The question is rewritten into one or more searches. “Who tracks how often AI assistants cite my website?” becomes a handful of queries, not one.
- A search index returns candidates. This is a conventional web index — the engine’s own, or a partner’s.
- A small number of pages are actually fetched and read. In our Perplexity run that number was ten for seven of eight questions, and twenty for the broadest one.
- The answer is generated from those pages, with the sources listed. What is not in the fetched pages does not make it into the answer, no matter how true it is.
The consequence. Getting into an AI answer is mostly a retrieval problem, not a reputation problem. If no page of yours is in the ten that get fetched, nothing about your brand reaches the model at the moment it matters.
What we measured, and on which engine
We hold an API key for Perplexity, and only Perplexity. Everything in this section is from our own stored run of eight questions on 19 September 2026, analysed in full on what Perplexity actually cites:
- Ten sources per answer for seven of eight questions; twenty for the broadest.
- Zero of the 90 cited URLs were a homepage, in either run. Every one was an interior page.
- 74 of 90 cited URLs in run A contained a word from the question in the path, on a strict literal rule; 75 of 90 in run B.
- 65 of run A’s 74 cited domains were cited exactly once; 66 of run B’s 75.
- Vendors’ own narrow product pages were cited as evidence about the product.
That is a complete census of a small sample: one engine, one category, one day. It is not a law of AI search.
What is publicly documented about the others
We have measured none of these. What follows is what the companies themselves publish, which is a much weaker kind of evidence than a stored run, and it changes often. Check the vendor’s own current documentation before you rely on any of it.
| Engine | API for measuring it | What we know |
|---|---|---|
| Perplexity | Yes — this is the one we use | Measured by us. See the analysis linked above. |
| ChatGPT search | No citation-reporting API | OpenAI documents separate crawlers for training, for search indexing and for a user-triggered fetch. Our adapter is built but not keyed, so we have measured nothing. |
| Claude | No citation-reporting API | Adapter built, not keyed. Not measured. |
| Google AI Overviews | None at all | No public API exists. Reviewed by hand for Scale clients. Anyone selling you automated AI Overviews tracking is scraping a results page. |
| Microsoft Copilot | None, but Bing Webmaster Tools reports it | Bing’s own AI performance reporting is first-party, free, and the best Copilot signal available. It tells you about your site, not your competitors. |
| Gemini | Not built by us | We have built nothing for it and measured nothing. Stated so you do not have to guess. |
What nobody outside those companies can tell you
There is a genre of confident blog post about “the ranking factors of AI search”. Treat all of it, including anything we might be tempted to write, with suspicion. These are the things that are genuinely unknown outside the vendors:
- How candidates are ranked before fetching. We see which ten pages were cited. We do not see the hundred that were considered, or why these ten won.
- Whether the model’s prior knowledge is weighted against the retrieved pages, and by how much.
- Whether a citation reflects a passage the model actually used or a source it merely fetched. The two are not the same and the difference is invisible from outside.
- How much personalisation, location and session history change the set. Our figures come from clean API calls with no user context, which is not what your customer is doing.
What follows for you, in order
- Be fetchable. A page that is blocked, JavaScript-only or behind a login cannot be retrieved at all. Start at AI crawlers and robots.txt.
- Have one page per question. Retrieval matches a question to a page. A page covering nine topics matches none of them well.
- Answer in the first two sentences. The generation step copies from what it reads. Give it something short, specific and self-contained to copy.
- Put the facts in text on your own domain. Prices, coverage, limits. A number in an image or a PDF cannot be quoted.
- Measure the questions, not the keywords. Retrieval is triggered by a question. Your prompt set should be questions. See choosing the questions to track.
Where this leaves training data
Nothing above touches the model’s weights, and that is deliberate. You cannot edit a trained model, and no vendor — us included — can get a claim about you into one. When an assistant says something outdated about your company from memory rather than from a source, the fix is to make the correct page so retrievable that the search step overrides the memory. That is a different problem with a different playbook: when AI says something wrong about your brand.
Common questions
Do AI answer engines search the web or answer from training data?
Most of the consumer answer engines search. Perplexity, ChatGPT search, Claude with search, Google AI Overviews and Copilot all run a live retrieval step, fetch a handful of pages and generate the answer from what those pages say. A model will fall back on training data when no search runs or nothing useful comes back, which is where stale claims about your company come from.
How many pages does an answer engine actually read?
Far fewer than a search results page shows. In our own Perplexity run on 19 September 2026, seven of eight answers cited exactly ten sources and the broadest question cited twenty. We have not measured the other engines, so we cannot tell you their numbers.
Can I influence what an AI model was trained on?
No, and be wary of anyone who says they can. Training data is fixed long before the model reaches you and cannot be edited by a vendor or by you. What you can influence is retrieval: which pages get fetched at the moment someone asks. That is where the work is, and it is also why results can move in weeks rather than never.
Which engines can Bungad measure automatically?
Perplexity, and only Perplexity. Our ChatGPT and Claude adapters are built but not yet keyed. Gemini is not built. Google AI Overviews and Microsoft Copilot have no public API at all and are reviewed by hand for Scale clients. We publish this in every place it is relevant because it is the first thing a buyer should check about any vendor in this category, including us.
Find out which pages get fetched for you
We ask Perplexity three of your buyers’ questions and show you every source it retrieved — and whether any of them are yours.
3 questions on Perplexity, free. A domain checked in the last week is served from that stored run rather than asked again, and the daily free allowance resets at midnight UTC.