How AI search engines choose sources is the question behind almost every client complaint about visibility. When someone asks ChatGPT or Perplexity a question and it answers with three links, the engine did not pull them out of a hat. The engine builds a shortlist of candidate pages, then picks the ones it can trust and, just as importantly, the ones it can lift a clean answer from. If your page is easy to trust and easy to quote, it gets picked. If it isn’t, it gets skipped, even when it’s the better page. In our experience much of what makes a page easy to quote is the same work that makes it usable with a screen reader — real headings, text that exists as text rather than baked into an image, link labels that say where they go. We look at that overlap in web accessibility and AI search.
Here’s the part most “get cited by AI” advice skips over: a lot of these engines don’t crawl the open web from scratch every time. They lean on the index Google already built. That’s why your Google ranking still matters for AI visibility, and it’s not a theory anymore — there’s a live court case that hangs on exactly this point. So the honest answer to how AI search engines choose sources has two halves: where they get their candidates, and how they choose winners from that pile. Take them in order.
Where the candidate list comes from (and why Google is in the loop)
An AI answer engine has to start with a set of pages to consider. Some of that comes from its own crawling and its training data. But a meaningful chunk appears to come from live search results — and in at least one case, from Google’s specifically.
In October 2025, Reddit sued Perplexity along with three data-scraping firms (SerpApi, Oxylabs, and AWMProxy), alleging they harvested Reddit content by scraping Google’s search results rather than Reddit directly, sidestepping Reddit’s own anti-scraping controls (PBS NewsHour, Oct 22, 2025). Reddit’s most vivid claim: it says it planted a test post visible only to Google’s crawler, and the content surfaced in Perplexity soon after — which, in Reddit’s telling, means the data could only have come through Google’s index. These are allegations in an active suit, not proven facts. But the case is moving forward: on July 31, 2026, a Manhattan federal judge rejected most of Perplexity’s attempt to dismiss it (Reuters, Jul 31, 2026).
Set aside who’s right in court. The practical takeaway is simple. If search results help assemble the candidate pool, a page that ranks well in Google has a real shot. It gets into the room when the model picks who to cite. A page that ranks nowhere usually isn’t even a candidate. Google position isn’t the whole game — more on that below — but it’s the ticket to the shortlist.
How the model picks winners from that pile
Once an engine has its candidates, it runs something closer to a retrieval-and-ranking process than a beauty contest. Four things do most of the deciding.
How retrieval actually works
The engine breaks your page into small pieces: a passage, a paragraph, a definition. It converts each into an embedding, a numeric fingerprint of what that chunk means. When a question comes in, it turns the question into the same kind of fingerprint and grabs the chunks whose meaning sits closest. This is why a page can rank for a keyword yet never earn a citation. If your answer sits inside a rambling 300-word paragraph, the model finds no tidy chunk to retrieve. Write the answer as its own passage and it becomes retrievable.
Structure it can extract
Models reward content they can pull apart cleanly. A clear H1, a logical H2/H3 hierarchy, and a direct 40-to-60-word answer sitting right under the question it answers — that’s the format that gets quoted. It’s the same structure that helps a screen reader navigate a page, which is a useful tell: if your HTML is genuinely accessible, it’s usually AI-legible too. One job, two payoffs.
Signals that you’re a safe source
Engines favor sources they already have reason to trust. You build that trust the slow way. It starts with a recognizable entity the model can pin down. Add corroborating mentions across other sites the engine reads, and links from domains with real authority. This is where a brand-new site struggles most. The content may be perfectly good; the problem is that nothing out there yet vouches for it. Earned mentions do double duty. Aim for places AI engines lean on heavily, like Reddit, LinkedIn, and industry press. They help humans discover you, and they show the model that you exist and know your subject.
Freshness
For questions where recency matters, recently updated pages get an edge — the same instinct behind Google’s long-standing “query deserves freshness” behavior. A page with a visible, current date and genuinely refreshed content reads as more reliable for anything time-sensitive than an untouched post from three years ago.
So do I just need to rank #1 on Google?
No — and this is where a lot of advice oversells one half of the story. Ranking gets you into the candidate pool, but it doesn’t guarantee the citation. Several vendor analyses this year make a similar point. Many AI citations go to pages that do not sit at the top of Google results. They favor strong structure, clear entity signals, and useful specifics over pure position. Treat those specific percentages with caution. They come from individual tools measuring their own samples, not a neutral standard. Still, the direction fits what the retrieval process above would predict.
Hold both ideas at once, because both are true. Rank in Google to qualify for the shortlist. Then structure your page and earn your mentions so the model actually chooses you off that shortlist. Teams that only chase rankings get into the room and still don’t get quoted. Teams that only polish “AI formatting” but never rank aren’t in the room to begin with.
A playbook you can run across clients
If you manage sites for other businesses, this is squarely in your lane — it’s repeatable work that rewards a system, which is exactly what a good agency or freelancer is built to deliver. Here’s the sequence I’d run per page to shape how AI search engines choose sources:
1. Earn the ranking first. Target a real query, cover it properly, and get the page ranking in Google. That’s the entry ticket to the candidate pool. Nothing downstream matters if the page is invisible.
2. Chunk your answers. Put a direct 40-to-60-word answer under each question or subhead. One idea per passage. Make it trivially easy for a model to lift a clean quote without dragging in three unrelated sentences.
3. Nail entity clarity. Be consistent about who you are — same business name, same author bylines, an about page a model can parse, and where it fits, schema markup. You’re helping the engine pin down the entity behind the content.
4. Earn outside mentions. Pursue genuinely relevant mentions and links, and be useful in the communities AI engines read. A single editorial mention on a site you’d be proud to be named on beats a pile of filler links that only inflate a report.
5. Keep it fresh and measure. Revisit important pages on a schedule, update them, and show the date. Then actually check whether AI engines are citing you, so you’re optimizing against reality instead of a hunch.
Then tie it back to SEO
Treat this as an extension of good SEO rather than a separate program. The same fundamentals — rank, structure, authority, freshness — feed both the blue links and the AI answers, which is genuinely good news: you maintain one program that pays off in two places. When an agency is honestly the right hands for that work, it is; this is detailed, ongoing, multi-client optimization, and doing it properly at scale is a real discipline. If you’d rather not do it by hand, that’s the gap we built Hepteon to close.
Where this fits with what we’re building
Hepteon runs a website end to end with seven AI agents, each owning one discipline. The Strategist sets the plan, the Connector wires up the data, the Technical agent keeps the site fast and crawlable, the Writer produces the chunked, citable content, the Amplifier earns the outside mentions, the Results agent measures what’s actually landing in search and AI answers, and the Publisher ships it all. The name comes from the heptathlon — seven events, one athlete — because getting cited by AI comes down to rank, structure, trust, and freshness working together, maintained continuously. That’s the whole point of the earlier sections: the engines reward the site that does all of it, not the one that does one of it loudly.
Want to go deeper on how these systems differ? Our breakdown of SEO vs GEO vs AEO lays out where each one starts and stops.
Frequently asked questions about how AI search engines choose sources
They build a shortlist of candidate pages, often drawing on existing search results, then rank them by how easily they can trust and quote each one. Pages that rank in Google, use clear structure, show entity trust signals and stay fresh are the ones that get picked.
Yes. Many engines lean on the index Google already built, so a page that ranks well has a real chance of entering the candidate pool. A page that ranks nowhere usually is not even considered, though ranking alone does not guarantee the citation.
Retrieval works on small passages, not whole pages. If your answer to a specific question is buried in a long paragraph, there is no clean chunk to lift. Write each answer as its own 40-to-60-word passage and it becomes retrievable.
A recognizable entity the model can identify, corroborating mentions across other sites it reads, and links from domains with real authority. New sites struggle because nothing external vouches for them yet, so earned mentions do real work.
No. Ranking gets you onto the shortlist, but structure, entity clarity and outside mentions decide whether the model actually quotes you. Chase both: rank to qualify, then make the page easy to trust and easy to quote.