How AI Search Engines Choose Which Websites to Mention

The 30-second answer

In reality, AI search engines pick sources in two stages. A retrieval system searches an index for pages that semantically match the question, then the model filters those candidates by signals like information gain, entity clarity, and freshness, and finally quotes or links the clearest sources. In reality, Structure and clarity decide who gets mentioned.

What you’ll learn

The sections follow one question through the machine, then turn the machinery around into actions your site can take this month.

What happens when you ask an AI search engine a question

A founder I advise showed me two screenshots last month. She had asked two different AI tools the same buying question about her category, and both answers named the same three competitors. Her company appeared in neither, despite ranking second on Google for the exact phrase. Second on Google. Meanwhile, Invisible to the machines. Her question to me was better than any conference talk on this subject: what does the machine actually see?

What the machine sees is a pipeline, and AI search engines run it in about the time it takes you to read your own question, which means every optimization decision you make has to land inside a window measured in seconds. Five steps, five places where a website can fall out of the race entirely.

If this is your first look inside the category, our consumer-level explainer, What Is AI Search?, covers what the user sees; the rest of this section covers what the site experiences. Walk the path once and the mystery mostly evaporates.

The pipeline, step by step

1. The question gets parsed and expanded. The engine rewrites your sentence into internal search queries and decides what kinds of sources it needs. A page that only matches its own brand vocabulary can fall out right here.
2. Retrieval pulls candidates from an index. AI search engines query their own crawl or a partner index and gather pages that look semantically relevant. If your page isn’t crawled, it was never in the race.
3. Candidates get ranked for semantic relevance. Meaning match, not string match: each passage is scored on how well it actually addresses the question, the way modern retrieval pipelines are documented by specialists like ziptie.dev.
4. The model filters for information gain and trust. Ten passages saying the same thing collapse into one. Sources that add a new fact, a number, or a clearly identified entity survive; vague repetition doesn’t.
5. Generation writes the cited answer. Only a short list of sources survives into the final text, citations attached. Everyone else did the work of existing and got nothing.

Every step has a cut, and the brutal one is step four, because getting retrieved only gets you read by a machine while getting mentioned requires being the source the model trusts enough to quote. Those are different games. AI search engines don’t owe your site anything; they owe their users an answer. Your job is being the best raw material for that answer.

Three engines, three indexes

Where do the candidates come from? Each engine assembles its own raw material, and AI search engines differ more than people assume here: ChatGPT leans on its own crawler plus a partner index, Perplexity runs its own retrieval stack, and Google sits on the web’s largest index of its own making.

The differences show up at the edges, on fresh news and narrow niches, which is why the same question can cite different sites on different engines. You’re not competing in one game. You’re competing in three that share a scoreboard.

The five signals that decide who gets mentioned

No engine publishes its scoring sheet, so treat this as the closest thing to a checklist: five signals that keep showing up in the research literature, in documented retrieval pipelines, and in the observed behavior of the products themselves, quarter after quarter. Skip one and the funnel leaks at that stage. All five together is how a small site out-cites a big one.

Signal by signal

1. Semantic relevance. Your page must address the question in meaning, not just contain the words. Pages organized around one clear topic per section make this easy. Pages padded with every variation of a keyword make it impossible.
2. Information gain. Engines prefer sources that add something the other candidates don’t: an original number, a first-hand detail, a sharper explanation. If your page paraphrases the top five results on Google, it adds nothing worth quoting.
3. Entity clarity. The engine has to know who you are: your name, what you do, and how you connect to other known entities. Consistent descriptions, a real About page, and structured data do this work. Vague brands stay vague to machines.
4. Quotability. The research here is blunt. Adding citations, quotations, and statistics to content lifted visibility in generative answers by up to 40%, with quotation-heavy content gaining roughly 37-41% on some measurements (Aggarwal et al., arXiv, November 2023). Write sentences a machine could lift intact.
5. Freshness and crawlability. Visible dates, honest update notes, clean sitemaps, and fast pings all matter. IndexNow, the protocol backed by Bing and Yandex, lets sites announce changes immediately instead of waiting for a re-crawl.

None of these five are exotic, and AI search engines reward exactly this kind of unglamorous work: being findable and quotable at the same time. The signals also compound. A clear entity publishing pages with genuine gain, quotable phrasing, and fast crawl coverage scores on all five without any single trick.

Old fundamentals, new filter

Notice the overlap with classic SEO rather than the replacement. Retrieval still leans on crawlability and rankings to build its candidate pool, which means the fundamentals keep paying rent. What’s new is the second filter, the one that decides whether a retrieved page actually gets spoken aloud. Most sites have spent years optimizing for the first filter and almost no time on the second. That imbalance is the opportunity, and it belongs to whoever moves first.

Why Wikipedia and Reddit dominate citations (and when they don’t)

Start with the obvious pair. One site. Nearly eight percent of everything ChatGPT cited, per Profound’s June 2025 study, which put Wikipedia first at 7.8% of the total, with Reddit close behind in many categories. Why do AI search engines lean on the same handful of sites? The Wikipedia reasons are boring and instructive: it’s structured, regularly updated, and written in the neutral, definitional register that models find easy to reuse, with a citation already attached to nearly every claim, which is exactly the format a quote-hungry machine prefers.

Reddit earns its place differently. Threads carry the phrasing real people actually use, complete with the vocabulary of someone who has the problem. When a model wants to explain what buyers experience rather than what a brochure promises, a forum thread often beats a polished product page. It answers the emotional version of the question too.

The part most commentary misses

Here’s the part most commentary misses: the dominance is gravity, not law. Semrush’s three-month AI-citation study measured, by November 2025, both Reddit’s and Wikipedia’s citation share dropping from earlier 2025 levels. Search Engine Land’s March 2026 analysis explained the mechanism. Giant sites win because a million-topic footprint means something relevant always exists. Which means niche topical depth can beat them on specific queries, and the citation pools AI search engines draw from keep rearranging.

What should you do with that? Don’t game the giants; that’s a losing sport with a moderation team on the other side. Study what they have in common instead, because the lesson transfers. Copy the bones. Structure, specificity, and citations made Wikipedia quotable. Lived experience and plain speech made Reddit quotable. Build your own pages with the same bones, and AI search engines get the same payoff from quoting you that they get from quoting the giants, without the anonymity problem.

What the research says about getting cited

The best evidence base comes from a paper. Aggarwal et al. published “GEO: Generative Engine Optimization” on arXiv in November 2023, with authors from Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi. The team tested how generative engines choose sources and what publishers could do about it, running controlled experiments rather than arguing from anecdote. The headline still holds: applying GEO methods boosted visibility in generative answers by up to 40%. Forty percent. From editing choices, not ad budgets.

The tactic-level findings matter more than the headline: adding citations, quotations, and statistics drove the biggest gains, with quotation-heavy content lifting visibility roughly 37-41% and statistics adding 22-30% on some queries. The classic tricks, keyword stuffing first among them, showed little effect. The machines grade for evidence, not vocabulary.

Two 2026 datapoints tell you the window is still open. Digital Applied reported in January 2026 that 47% of brands have no GEO strategy at all. That’s the opening. Omnibound sized the US generative engine optimization market at roughly $365.4 million for 2026, compounding near 42.9% a year. We turned the full tactic list into plain English in What Is GEO? Generative Engine Optimization Explained. Early movers get to define their category’s answers.

You can replicate the spirit of that research on your own category in an afternoon. Gather the ten questions your buyers actually ask, run them through three assistants, and grade every answer for what it cited: definitions, numbers, named sources, or pure vibes. Grade hard. The gaps you find, the questions with vague answers and weak citations, are your content roadmap for the quarter. The pattern we see across categories is consistent: the wins come from being specific where everyone else is generic.

Entities: the new moat

Here’s the idea that turns the last two sections from tactics into strategy: entities. An entity is a thing a machine recognizes as distinct, your brand, your product, your founder. When AI search engines resolve your name into a known entity with attributes and relationships, every question in your category gets easier to answer with you in it, because the machine no longer has to guess whether two mentions refer to the same thing. When they can’t resolve you, you’re just words on a page. That’s the problem to fix.

Building that recognition is unglamorous work. One consistent description of what you do, repeated across your site, your profiles, and the directories that matter. An About page that states facts plainly. Organization and FAQ structured data so machines don’t have to guess. External corroboration from press, partners, and review platforms. None of it is clever. All of it compounds.

Ask the machines directly

How do you know whether the work is landing? Ask. Type your brand into a few assistants and see whether they describe you accurately, then ask what you do, who you serve, and who you compete with. When AI search engines can answer those cleanly, the entity has resolved; when they hedge or hallucinate, the signals aren’t consistent yet. Log the answers monthly. Entity repair is one of the few SEO tasks where the diagnosis is free.

Why call it a moat? Because entities survive reorganizations. Citation pools shift, models get swapped, and the sites that keep getting mentioned are the ones the machine can reliably identify, which is why entity work pays off across every future model change rather than just the current one. We took up the doom-flavored version of this question in Will AI Replace SEO? The Honest 2026 Answer. The short version: the work survives, it just changed shape.

A 30-day plan to become citable

Research is lovely; a calendar is better. Here’s the one-month version I’d run for any site that wants to become citable, in order, with no step skipped.

Table 1. The 30-day citability plan

Set two expectations before you start. This works in weeks and months, not hours, because entity recognition and citation share move at the speed of re-crawling and model updates, and no dashboard on earth can hurry either one. And it compounds quietly, which is precisely why most competitors won’t finish week two. Boring, methodical, and early beats clever here.

Day 30 should leave you with three artifacts: a baseline scorecard from week 1, a set of pages that quote cleanly, and a monthly habit of asking the machines what they know. If the score hasn’t moved, don’t panic and don’t quit; the engines update on their schedule, not yours. Give the program a full quarter before you judge it. Then repeat monthly, and let the compounding do what it does.

Frequently asked questions

How do AI search engines choose sources?

In two stages. Retrieval pulls candidate pages from an index by semantic relevance to your question. The model then filters those candidates on signals like information gain, entity clarity, and freshness, and writes one cited answer from the survivors. Pages that state facts cleanly and identify themselves clearly survive the filter most often.

Do backlinks still matter for AI mentions?

Less directly than in classic rankings, but they haven’t disappeared. Authority and crawl priority still decide what gets retrieved in the first place, and links from trusted pages help engines discover you. The newer lever is citability: pages that state facts cleanly, with named sources and quotable lines, get mentioned even with modest backlink profiles.

Why does Reddit show up in AI answers so often?

Scale and usefulness. Search Engine Land explained in March 2026 that Reddit, Wikipedia, and YouTube are cited heavily because they’re massive sites with a million-topic footprint, so nearly any question finds a relevant thread. The mix shifts, though: Semrush’s November 2025 study measured Reddit’s and Wikipedia’s citation share falling from earlier in 2025. Niche depth can beat them.

Does my Google ranking affect AI mentions?

It helps, but it isn’t the whole game. High rankings feed retrieval, since engines favor pages that already rank and get crawled often. Selection then adds different tests: entity clarity, information gain, and quotability. A well-structured page that answers one question precisely can get cited over a stronger domain that buries the answer.

How do I track whether AI mentions my site?

Start manually: ask ChatGPT, Perplexity, and Google’s AI Mode the ten questions a customer would ask, and log which brands get named. Then watch analytics for referral traffic from AI tools and check server logs for AI crawler user agents. Dedicated GEO trackers appeared through 2025 and 2026; treat their dashboards as directional until their methods are published.

The bottom line

AI search engines choose their mentions the way a strict editor picks quotes: they take the source that is relevant, adds something new, is clearly who it claims to be, and is easy to cite intact. None of that is out of reach for a small site.

Wikipedia and Reddit carry the load on volume, but Search Engine Land’s March 2026 analysis showed the openings that niche depth leaves, and the GEO research turned citability into a checklist you can actually work through. The checklist asks for editing discipline, not budget.

So run the 30-day plan, honestly and in order: baseline what the engines say today, clean up your entity, make your best pages quotable, then measure. Do it before your competitors work out that ranking and being mentioned are different sports. It’s the same discipline we apply with clients at HelpingHandAI, and the full playbook, including the SEO foundation that still feeds all of this, is in Will AI Replace SEO? The Honest 2026 Answer. The machines are already choosing. Make their choice easy.

Sources and further reading

• Aggarwal et al., “GEO: Generative Engine Optimization”, arXiv 2311.09735 (November 2023)
• Profound, study of ChatGPT citation sources putting Wikipedia first at 7.8% (June 2025)
• Semrush, three-month AI citation study tracking Reddit and Wikipedia share (November 2025)
• Search Engine Land, analysis of why LLMs cite Reddit, Wikipedia, and YouTube, and when niche depth wins (March 27, 2026)
• Ziptie.dev, explainer on how retrieval (RAG) pipelines select sources by relevance, information gain, and entity clarity (2026)
• Digital Applied, survey finding 47% of brands lack a GEO strategy (January 2026)
• Omnibound, US generative engine optimization market estimate of $365.4 million at 42.9% CAGR (May 2026)
• IndexNow.org, protocol documentation supported by Bing and Yandex