How AI Search Engines Choose Sources, and How to Be One

First published: 4 July 2026 · Last updated: 4 July 2026

1

Crawl and index

AI engines maintain their own indexes (ChatGPT via OAI-SearchBot, Perplexity via PerplexityBot, Google AI via Googlebot, Bing via Bingbot). Pages must be crawlable, indexable and parseable. Robots.txt blocks, JS-rendering issues and slow LCP all kill source candidacy here.

2

Retrieve semantically relevant passages

For a given query, the engine pulls candidate passages, not whole pages. Retrieval scores semantic match (vector similarity to the query intent), not keyword density. The unit of competition is the passage, usually 100 to 400 words around a clear answer.

3

Rerank by E-E-A-T plus GEO signals

From the candidate passage set, the engine selects 2 to 5 to actually cite. Reranking weights: source authority (cross-web mentions, link graph), passage credibility (cited statistics, named expert quotes), structural clarity (clean question + answer, schema markup) and freshness.

The honest research on how AI search engines pick sources is much narrower and more practical than most "GEO best practice" content suggests. The foundational paper, Pranjal Aggarwal et al's "GEO: Generative Engine Optimization" (ACM SIGKDD 2024, with collaborators from IIT Delhi, Princeton, Georgia Tech and the Allen Institute for AI), did the work that everyone keeps quoting badly. They built a benchmark of 10,000 queries, tested nine optimisation strategies on real generative engines, and measured which strategies actually moved citations. Three strategies moved the needle (each by 30 to 40 percent on a position-adjusted citation metric): adding inline citations to authoritative sources, adding statistics with concrete numbers, and adding expert quotations with attribution. Several strategies that intuitively sound like they should work moved the needle by approximately zero, including keyword stuffing, adjective insertion, and authority-claim language without supporting evidence. This piece walks through what the evidence and engine-side documentation tell us about the actual source-selection pipeline, what each stage is selecting on, and what the practical SG playbook is to get cited. We covered the broader GEO landscape in our generative engine optimization Singapore guide; this is the deeper "why" piece on the mechanics behind it.

Stage 1: Crawl and Index, the Quiet Eliminator

Before an AI engine can pick your page as a source, it needs to be in that engine's index. Most SG sites assume that being indexed by Google means they are indexed by every AI engine. They are not. Each major AI engine maintains its own index, populated by its own crawler:

• OpenAI / ChatGPT Search: OAI-SearchBot crawls and indexes for ChatGPT search results. GPTBot crawls for training data; the two are separate. Block one and not the other if you want training-data exclusion without losing search visibility.
• Perplexity: PerplexityBot crawls and indexes Perplexity's own retrieval index. Perplexity also retrieves from third-party indexes (notably Bing) as a supplement.
• Google AI Overviews: Uses Google's main index via Googlebot. If you are indexed in Google Search, you are eligible for AI Overviews. There is no separate "AI Overviews bot."
• Bing Copilot: Uses Bing's index via Bingbot. Same eligibility logic as Google AI Overviews.
• Anthropic / Claude: ClaudeBot crawls for Claude's web access. Claude search is newer and the index is smaller.

The first practical implication: check your robots.txt for any of these crawlers being blocked. We see SG sites blocking GPTBot accidentally (intending to stop training-data scraping) and inadvertently blocking OAI-SearchBot too. Result: invisible to ChatGPT search. The second implication: pages need to render server-side or be fully crawlable. Heavily JS-rendered SPAs that work fine in a browser often fail to be indexed cleanly by AI engine crawlers. If your page only assembles its content after JavaScript execution, AI engines often store an empty or partial version. We covered the technical detection side of this in AI search engine optimization Singapore. The third implication: speed and uptime matter. AI engine crawlers throttle slow sites and skip frequently-down sites. Core Web Vitals are not directly used for ranking by AI engines but a slow site is a site that gets crawled less frequently, which means staler content in the engine's index.

Stage 2: Retrieval as Passage-Level, Not Page-Level Competition

The biggest mental model shift compared with classic SEO: AI engines do not retrieve whole pages, they retrieve passages. The unit of competition is a 100 to 400 word block of text around a clear answer, not the whole page. This is because AI engines use retrieval-augmented generation (RAG). For a given query, the engine vectorises the query into a semantic embedding, then matches it against vectorised passages from its index. The top-scoring passages (typically 5 to 20) are retrieved as candidates. Only those candidates are sent to the language model that composes the answer. Practical implications for SG content: **1. Each section of your page competes independently.** A page with one strong section and three weak sections is one passage in retrieval, not four. Front-load the strongest answer to the highest-priority query. **2. The first 100 to 200 words of any section matter disproportionately.** Retrieval truncates. A section with a clear definitional opening and dense supporting context in the first 150 words ranks higher in retrieval than a section with throat-clearing introduction. **3. Passage-level signals beat page-level signals.** A clear H2 question + 40 to 60 word answer + 200 word supporting paragraph is the unit AI engines retrieve and lift. Massive prose paragraphs without internal structure are harder to retrieve cleanly. **4. Semantic match beats keyword match.** Retrieval vectorises both query and content. A section that addresses the query intent in natural language ranks higher than a section that stuffs the exact keyword into every sentence. This is why our AEO content structure guide is built around the H2-question + 40-to-60-word-answer pattern. It is not a stylistic preference, it is the structural shape that matches how AI engines retrieve.

Stage 3: The Rerank Layer Where Authority Decides

After retrieval surfaces candidate passages, the rerank layer decides which 2 to 5 actually get cited in the generated answer. This is where E-E-A-T signals plus GEO-specific signals do their work. E-E-A-T signals the engines weight (this is well-documented in Google's own guidance plus reverse-engineered behaviour from ChatGPT and Perplexity):

• Cross-web mentions of the source domain. How many other authoritative sites talk about your brand. Both linked and unlinked mentions count, with unlinked mentions weighted more heavily by AI engines than by classic Google.
• Author-level signals. Named author with bio, credentials, and presence elsewhere on the web. Anonymous content is systematically downranked.
• Domain-level authority signals. Aggregated trust score broadly correlated with classic Domain Rating but not identical.
• Topical consistency. A site that publishes 50 posts on a topic is treated as more authoritative on that topic than a site that publishes one.

GEO-specific signals from the Aggarwal et al research:

• Inline citations to authoritative sources. Linking to studies, primary sources, government data. Direct measured impact on citation rates.
• Specific statistics with numbers. "73 percent" beats "most." "Increased from SGD 4,200 to SGD 5,800 in 90 days" beats "increased significantly."
• Named expert quotations with attribution. "According to Dr Vanessa Tan, Head of Family Medicine at Angsana Medical, ..." beats "experts say."
• Structured definitions. Clear "X is Y because Z" patterns. Especially valuable for being lifted in answer engines.

Three signals that intuitively feel like they should help but do not, per the same research:

• Keyword stuffing. Approximately zero impact on AI engine citations.
• Authority-claim language without supporting evidence. "Industry-leading", "world-class", "best-in-class" without the data to back it. Approximately zero impact.
• Adjective and adverb density. Stylistic enhancement that classic content marketers were taught to add. Approximately zero impact on AI citation.

The clear takeaway: if you are deciding what to add to your SG content to improve AI citation rates, prioritise stats, named expert quotes and inline citations to authoritative sources. Skip the adjective polishing.

Signal	Weight	What to do on your SG site
Cross-web brand mentions (linked + unlinked)	High	PR, podcast appearances, guest posts, directory listings, journalist quotes
Inline citations to authoritative sources	High	Link to .gov.sg, .edu.sg, peer-reviewed studies, primary data
Specific statistics with numbers	High	Replace "many" with specific %, replace "increased" with from/to numbers
Named expert quotations	High	Add 1 to 2 named-expert quotes per long-form piece, with credentials and attribution
Author bio with credentials	Medium-high	Author byline, bio, link to LinkedIn, credentials visible
Topical consistency on the domain	Medium-high	Publish a cluster of pieces on the topic, not just one
Structured definitions and clear answers	Medium-high	H2 question + 40 to 60 word answer immediately under each heading
Schema markup (Article, FAQ, HowTo)	Medium	Implement on every blog post and money page
Domain authority (general)	Medium	Slow build through PR, partnerships, mentions; faster via guest content
Freshness (recent dates, updates)	Medium	"Last updated" dates, refresh content on a quarterly cadence
Keyword density	Negligible	Do not stuff. Natural mention 2 to 4 times across the piece is plenty.
Adjective density	Negligible	Cut "leading", "innovative", "best-in-class" without supporting data

How the Different Engines Diverge in Source Selection

While the three-stage pipeline is broadly consistent across major AI engines, each one weights signals differently. The single biggest finding from cross-platform GEO research is that only about 11 percent of cited domains overlap between ChatGPT and Perplexity. The rest are platform-specific. Optimising for one engine does not automatically lift you in another.

ChatGPT Search. Heavy weighting on long-form structured content, named expert quotes, and topical authority. ChatGPT is also heavily influenced by content seeded on Reddit (its training data and live retrieval both lean heavily on Reddit). For consumer-facing SG categories, getting mentioned in active Reddit threads is one of the most underrated ChatGPT visibility tactics.

Perplexity. Heavy weighting on freshness, primary-source citations, and direct factual accuracy. Perplexity is the most "SEO-aware" of the AI engines, retrieving from indexes that overlap with classic search rankings. Strong on statistics and data-rich content.

Google AI Overviews (Gemini). Heavy weighting on existing Google search rankings, schema markup, Google Business Profile signals for local queries, and YouTube video integration. The closest of the AI engines to "SEO works the same way" but with a passage-level retrieval twist.

Bing Copilot. Heavy weighting on Bing's index, with strong consumer commerce integration. For SG e-commerce sites, Bing Copilot citation correlates fairly closely with Bing organic ranking.

Claude (Anthropic). Smaller live web index than the others. Tends to weight authoritative source domains more heavily than smaller specialist sites. Worth being on but not yet a primary visibility battleground for most SG sites.

For deeper AI engine specifics, our piece on AI answer engines vs Google covers the higher-level competitive landscape, and how to rank in Google AI Overviews Singapore drills into Google's specific behaviour.

What Happens to a Page That Gets Picked

Useful to think about the downstream economics of being cited, since they are different from being a top-3 organic result.

When ChatGPT, Perplexity or AI Overviews cite your page, the user usually reads the AI's synthesised answer and either acts on it directly or clicks through to one of the cited sources. Click-through rates on AI engine citations vary widely (Perplexity is generally highest at 5 to 15 percent depending on category, AI Overviews lower at 1 to 5 percent, ChatGPT in between).

Lower CTR than classic SEO. But three reasons being cited is still high-value:

1. Brand impression value. Even users who do not click see your brand cited as a credible source. For SG B2B and considered-purchase categories this builds awareness over time.

2. Higher-intent click-through. Users who do click from an AI engine citation have already pre-qualified themselves through the AI's answer. They convert at meaningfully higher rates than classic organic search clicks.

3. Long-tail compounding. A single well-crafted piece can be cited across hundreds of conversational query variations. Cumulative impression and click value compounds far beyond what classic ranking-for-one-query economics suggest.

The SG Playbook for Becoming a Cited Source

Combining the research and the engine-specific behaviour, the practical SG playbook:

1. Get crawlable. Audit your robots.txt for OAI-SearchBot, PerplexityBot and ClaudeBot. Confirm none are blocked. Confirm pages render server-side or work in JS-rendering tests. Audit indexability in Google Search Console as a proxy.

2. Restructure pages around passage retrieval. Each high-priority page gets H2 questions with 40 to 60 word direct answers, followed by 200 to 400 word expansion paragraphs. Schema markup on every article (Article + FAQPage where applicable).

3. Add the three Aggarwal-validated tactics to every long-form piece. At least 1 to 2 inline citations to authoritative external sources. At least 3 specific statistics with concrete numbers. At least 1 named-expert quotation with attribution.

4. Build cross-web brand mentions. PR placements, podcast appearances, guest content, directory listings (SG-specific only, per our SG SOP), Reddit and HardwareZone organic engagement, journalist quotes. Both linked and unlinked mentions count.

5. Build topical depth. Publish a cluster of 10 to 30 pieces on the topic where you want to be cited. AI engines weight topical authority. One isolated post will rarely earn citations against a domain with 30 posts on the same topic.

6. Refresh and update. Add visible "Last updated" dates. Refresh content on a quarterly cadence. AI engines use freshness as a meaningful signal.

7. Measure citation share. Manually query your top 30 to 50 priority queries through ChatGPT, Perplexity and AI Overviews quarterly. Track which cite your brand vs competitors. Use the gap list to brief content updates.

For broader GEO measurement guidance, our measuring GEO performance tools piece covers the tracking infrastructure side.

Common Mistakes SG Sites Make Trying to Get Cited

Mistake 1: Optimising for retrieval and ignoring trust. Beautiful passage structure, no cross-web mentions. The engine can find your content but does not consider you authoritative enough to cite.

Mistake 2: Blocking AI engine crawlers accidentally. Robots.txt directives that aim at GPTBot also catch OAI-SearchBot. Or blanket "disallow all bots" rules that include PerplexityBot.

Mistake 3: Treating the page as the unit of competition. It is the passage. Restructure pages so each section can stand alone as a retrievable answer.

Mistake 4: Skipping the three validated tactics. Adding stats, expert quotes and inline citations is not stylistic preference, it is the empirically validated lever. Most SG content has none of the three.

Mistake 5: Single-piece topical coverage. One post on a topic competes against domains with 30. Topical depth on the cluster level is what AI engines weight.

Mistake 6: Optimising for one engine and assuming cross-platform transfer. Only about 11 percent of cited domains overlap across major AI engines. Each engine needs its own audit.

Mistake 7: Ignoring author-level signals. Anonymous "team" bylines kill AI engine trust. Named author with credentials and presence elsewhere on the web is the baseline.

Frequently Asked Questions

How do AI search engines actually choose which sources to cite?

Three stages: crawl and index the web much like classic search engines, retrieve the most semantically relevant passages for a given query (not whole pages), then rerank those candidate passages by E-E-A-T signals (cross-web brand mentions, author credentials, domain authority) plus GEO-specific signals validated by the Aggarwal et al research (inline citations, specific statistics, named expert quotations). Out of the rerank set, the engine cites 2 to 5 sources in the generated answer.

Does keyword stuffing help my page get cited by ChatGPT or Perplexity?

No. The Aggarwal et al GEO research measured this directly and found approximately zero impact from keyword density manipulation on AI engine citation rates. The signals that do move citations are inline citations to authoritative sources, specific statistics with concrete numbers, and named expert quotations with attribution. Each of these lifted citation rates by 30 to 40 percent in controlled testing.

What is the unit of competition in AI search, the page or the passage?

The passage, not the page. AI engines use retrieval-augmented generation: for a given query, they retrieve candidate passages of typically 100 to 400 words, then rerank those passages and select 2 to 5 to cite. This means each section of your page competes independently. A page with one strong section and three weak sections is effectively one passage in retrieval. Front-load the strongest answer.

How do I know which AI engine crawlers are accessing my SG site?

Three approaches. First, check server logs for user-agent strings: OAI-SearchBot (ChatGPT search), GPTBot (ChatGPT training), PerplexityBot (Perplexity), ClaudeBot (Claude), Bingbot (Bing Copilot), Googlebot (Google AI Overviews uses the main Googlebot). Second, audit robots.txt for any of these being blocked. Third, manually query each engine for your brand and key pages, see whether you are cited. The combination tells you what is being indexed and what is being trusted.

Do I need to optimise separately for each AI engine?

Largely yes. Cross-platform research shows only about 11 percent of cited domains overlap between ChatGPT and Perplexity. Each engine has its own preferred signals: ChatGPT weighs Reddit and topical authority heavily, Perplexity weighs freshness and primary-source citations, Google AI Overviews tracks closely with Google search rankings, Bing Copilot tracks Bing rankings. The good news: the underlying validated tactics (cite, stat, quote, structure, brand mentions) lift visibility across all of them. You optimise once at the content level and audit per-engine for gaps.

How long does it take to start being cited by AI search engines?

Faster than classic SEO ranking but not instant. Pages that are crawlable, well-structured for passage retrieval and contain the three validated GEO tactics typically start appearing in Perplexity and ChatGPT within 2 to 6 weeks of publishing. Google AI Overviews lags slightly longer (4 to 8 weeks). The trust layer (cross-web brand mentions, topical depth) takes longer to build, on the order of months. Most SG sites we work with see meaningful citation lift within 60 to 90 days of implementing the playbook.

Related reading

• What is generative engine optimization Singapore
• AI search engine optimization Singapore
• How to rank in Google AI Overviews Singapore