When ChatGPT searches the web for you, it generates internal search queries called fan-out queries. These are the actual strings sent to Bing. They determine which pages get fetched, which enter the AI's context window, and which get cited in the response.
Several companies now sell tools that capture these fan-out queries. The pitch: "See exactly what ChatGPT searches for, then optimize for those exact strings."
We tested whether that approach works. We submitted the same 180 queries to ChatGPT three times, under identical conditions (same model, same API settings, same day). Then we compared the fan-out queries generated each time.
98% of the fan-out query strings were completely different across runs. The AI generated fresh search strings every time. Zero overlap.
But here is the part that changes the strategy: the type of search stayed the same 65% of the time. When ChatGPT decided to inject brand names on the first run, it injected brand names on the second run too -- just different brand names, in different combinations, with different phrasing. When it compressed to keywords, it compressed again -- just to different keywords.
The search strings are random. The search strategy is not.
This means tracking specific fan-out query strings -- what most tools capture -- is chasing a moving target. By the time you optimize for "best password manager Bitwarden 1Password comparison 2026," ChatGPT has already moved on to "top password managers NordPass Dashlane security review." Different words, same type of search.
The actionable target is the type, not the string.
What We Found
This comes from a study of 1,323 fan-out queries across ChatGPT, Gemini, and Perplexity. 540 parent queries across 10 commercial verticals and 5 intent types. The full methodology is in our research paper, but here is what matters for practitioners.
The Two Layers of AI Search
AI search is not one decision. It is two.
Layer 1: Does the AI search the web at all? This decision is highly stable. Across three identical runs on ChatGPT, 91.7% of queries produced the same search/no-search decision every time. If ChatGPT searched the web for "best CRM software" on run one, it searched again on run two and run three. If it answered "how does photosynthesis work" from memory on run one, it answered from memory every time.
This layer is driven by intent. Discovery queries ("best X for Y") trigger web search 98% of the time on ChatGPT. Informational queries ("how does X work") trigger search only 12% of the time. The AI is confident it knows how things work. It is less confident it knows what to recommend.
Layer 2: What does the AI search for? When it decides to search, the AI generates fan-out queries. The specific strings change every time (98% zero overlap across runs). But the structural type of search is moderately stable (65% of the time, the dominant type stays the same).
This two-layer structure means there are two distinct optimization problems. Layer 1 is about training data presence -- if the AI never searches, your web content does not exist. Layer 2 is about matching your content to the right type of search.
The Seven Types of AI Search
We classified every fan-out query into one of seven structural types based on what the AI is doing when it generates the search.
| Type | % of All Fan-Outs | What the AI Is Doing |
|---|---|---|
| Tangential | 17.2% | Searching for an adjacent topic the user did not ask about |
| Expansion | 16.0% | Looking up background context ("marketing budget," "skincare ingredients") |
| Evidence Seeking | 15.6% | Searching for reviews, studies, benchmarks, or proof |
| Entity Injection | 15.6% | Adding brand names the user never mentioned |
| Compression | 14.5% | Stripping the conversational question down to keywords |
| Narrowing | 9.6% | Adding specifics: a year, a price range, an audience segment |
| Reformulation | 7.0% | Rephrasing the same question in different words |
| Price/Availability | 4.5% | Checking prices or purchase options |
No single type dominates. The distribution is more even than our earlier Study 2 found, which was dominated by compression (37.9%). The broader query set in this study -- covering 10 verticals and 5 intent types instead of 3 client verticals -- produces a more diverse retrieval pattern.
Intent Determines the Type
The type of search the AI runs depends heavily on what the user is trying to do (chi-squared X2=299.6, p<0.001).
When someone is shopping (discovery intent), the AI's dominant behavior is entity injection -- it decides which brands matter and searches for them by name. Entity injection rates for purchase-intent queries: 19.8%. For knowledge-intent queries: 6.0%. That is a 3.3x difference.
When someone is learning (informational intent), the AI compresses the question to keywords and searches for explanatory content. It does not inject brand names.
When someone wants opinions (review-seeking intent), the AI searches for evidence -- reviews, studies, proof.
When someone is comparing (comparison intent), the AI checks prices and availability at the highest rate (4.8% price/availability, vs near-zero for other intents).
This pattern was statistically significant (p<0.001) and is the strongest finding in the study.
How to Optimize for Each Type
Here is the practical part. For each fan-out type, what content strategy matches it?
Entity Injection (15.6% of fan-outs) -- This Is PR Work
When ChatGPT decides to search for brand names, it pulls those names from its training data. Not from your website. Not from your schema markup. From the web content it was trained on.
If your brand appears in enough third-party sources -- review sites, industry roundups, directory listings, press coverage, Reddit threads -- the AI associates your brand name with your product category. Then when someone asks "best [your category]," the AI injects your brand into its search queries and goes looking for pages about you.
If your brand is not in that training-data category map, the AI never searches for you. Your pages could rank #1 on Google for every relevant keyword, and ChatGPT would never see them -- because it generates a fan-out query for "Bitwarden review 2026" instead of "YourBrand review 2026."
The strategy: Get mentioned on other people's sites. Review sites, comparison articles, industry lists, Reddit communities, YouTube reviews. This is reputation and PR work, not SEO. Our brand citation research found that 93.4% of brand-query AI citations come from third-party sources.
ChatGPT is the most aggressive entity injector. 32% of its fan-outs inject brands (vs 9.8% for Perplexity and 4.0% for Gemini). The smaller model tiers inject even more: gpt-5.4-nano injects entities on 44.8% of fan-outs. If your customers use ChatGPT, training-data presence is your top priority.
Compression (14.5% of fan-outs) -- This Is Classic SEO
The AI takes "how do I know if my gut bacteria are out of balance and what should I do about it" and turns it into "gut microbiome imbalance symptoms." A 15-word conversational prompt becomes a 4-word keyword search.
Your existing keyword research mostly covers this. The key is to make sure you are targeting the compressed form, not the conversational phrasing. If your page title is "Everything You Need to Know About Gut Health and How to Improve It," the AI may be searching for "gut microbiome imbalance symptoms" -- and your page does not match.
The strategy: For every target topic, identify the 4-6 word compressed keyword form. Put it in your title, H1, and first paragraph. This is traditional SEO with one adjustment: you are optimizing for what the AI compresses to, not what the human types.
Compression has the highest citation yield. In our Perplexity citation data, compression fan-outs produced 148% citation yield (1.48 citations per fan-out). They map directly to keyword-ranked pages. This is where traditional SEO directly translates to AI citation.
Evidence Seeking (15.6% of fan-outs) -- This Is Content Marketing
The AI searches for "[your product] review," "[your category] case study," "[your product] vs [competitor]." It literally uses these phrases.
The strategy: Create content that matches evidence-related search terms. The content needs to actually exist, rank, and contain real evidence -- not just marketing copy with "review" in the title. A page titled "Why Our Product Is the Best" will not match an evidence-seeking fan-out. A page titled "CRM Software Comparison: 6 Platforms Tested Over 90 Days" will.
Perplexity and ChatGPT lead in evidence seeking (20.5% and 23.1% respectively). If your customers use these platforms for research before buying, evidence pages are high priority.
Evidence-seeking fan-outs have the second-highest citation yield at 128%. When the AI searches for evidence, it cites what it finds.
Narrowing (9.6% of fan-outs) -- This Is On-Page Optimization
The AI adds "2026," "for small business," "under $50," "for beginners," "enterprise." These qualifiers come from the AI's interpretation of the user's context, not from the user's actual words.
The strategy: Make sure your pages include the qualifiers AI platforms add. Year mentions in titles and headings. Audience segments ("for small business," "for enterprise"). Price tiers. Experience levels. If the AI narrows to "best CRM for small business 2026" and your page title is just "Best CRM Software," you might not match the narrowed query.
Narrowing fan-outs produce 129% citation yield -- the second-highest after compression. When the AI narrows, it is getting specific, and specific queries produce direct citations.
Expansion (16.0% of fan-outs) -- You Cannot Optimize for This
Expansion fan-outs are background context lookups. The AI searches for "marketing budget" or "skincare ingredients" or "ecommerce supply chain" -- short, broad terms that build the AI's understanding of the topic before it answers.
These produce the lowest citation yield at 48%. The AI reads this content to inform its thinking, but it does not cite it. The content shapes the answer without appearing in the sources.
The strategy: There is no direct optimization for expansion fan-outs. They are the AI's homework. Your content may be read as background without ever being cited. This is useful to know because it explains why some high-traffic, high-ranking pages never appear in AI citations -- they serve as expansion context, not citable sources.
Tangential (17.2% of fan-outs) -- Cover Adjacent Topics
The AI wanders. It searches for topics related to but not directly about the user's question. Someone asks about CRM software and the AI searches for "sales pipeline management" or "customer retention strategy."
The strategy: Build content clusters that cover adjacent topics, not just your core product. If you sell CRM software, also cover sales methodology, customer success metrics, pipeline management frameworks. When the AI tangentially searches for these topics and finds your content, you enter its context window -- and you may appear as a citation for the main query even though the fan-out was about an adjacent topic.
Tangential fan-outs produce 67% citation yield. Not as high as compression or evidence, but not negligible. They are a path to citation through topical authority rather than direct keyword matching.
Price/Availability (4.5% of fan-outs) -- Almost Never Happens
Despite the fact that our entire query set was commercial in nature, the AI rarely checks prices or availability. When it does, it is most common on comparison queries (4.8% of comparison fan-outs).
The strategy: If you sell products online, make sure your pricing is visible in the HTML (not hidden behind JavaScript or "contact for pricing"). When the AI does search for pricing, it needs to find it on the page. But do not invest heavily in optimizing for price fan-outs -- they account for less than 5% of retrieval behavior.
Perplexity never generated a price fan-out in our data (0.0%). ChatGPT and Gemini accounted for all of them. If your customers compare prices on Perplexity, the AI is not checking your prices for them.
Each Platform Searches Differently
The three platforms have distinct retrieval personalities (chi-squared X2=386.9, p<0.001, Cramer's V=0.38 -- a large effect).
| Type | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| Entity Injection | 32.0% | 4.0% | 9.8% |
| Evidence Seeking | 23.1% | 5.7% | 20.5% |
| Expansion | 4.9% | 27.2% | 13.7% |
| Compression | 7.5% | 18.7% | 18.9% |
| Tangential | 12.4% | 21.3% | 16.1% |
| Narrowing | 9.5% | 7.3% | 13.1% |
| Price/Availability | 4.9% | 7.9% | 0.0% |
ChatGPT injects brands and seeks evidence. If your customers use ChatGPT, your priority is training-data presence (get mentioned on third-party sites) and evidence content (reviews, case studies, comparisons).
Gemini explores broadly. It casts a wide net with expansion and tangential searches. If your customers use Gemini, topical authority and content breadth matter more than any single page.
Perplexity compresses and seeks evidence. It is the most keyword-oriented platform. Traditional SEO (ranking for compressed keyword queries) and evidence content are both high priority for Perplexity.
The Model Tier That Changes Everything
An unplanned finding: ChatGPT's decision to search the web at all depends on which model is running.
| Model | Queries | Searched the Web | Search Rate |
|---|---|---|---|
| gpt-5.4-nano | 44 | 44 | 100% |
| gpt-5.4-mini | 46 | 46 | 100% |
| gpt-5.4 (flagship) | 90 | 26 | 29% |
The flagship model (gpt-5.4) only searched the web 29% of the time. The smaller models searched every time. The bigger the model, the more confident it is in answering from training data alone.
This means your web content may be invisible to the most capable version of ChatGPT -- not because it cannot find you, but because it never tries. It answers from memory.
The intent effect makes this worse. Informational queries trigger web search only 12% of the time on ChatGPT (across all model tiers). If someone asks the flagship model an informational question, the probability of any web search happening is close to zero.
The practical implication: For informational content targeting ChatGPT, training-data presence is more important than web ranking. For discovery and comparison content, web ranking still matters because those intents trigger search 72-98% of the time.
Why Tracking Specific Fan-Out Strings Does Not Work
This is the central finding.
We submitted the same 180 queries to ChatGPT (gpt-5.4-mini) three times. Same model, same settings, same day.
| What We Measured | Result |
|---|---|
| Search trigger agreement | 91.7% (same search/no-search decision each time) |
| Fan-out string overlap | 0.012 Jaccard (98% of pairs share zero exact strings) |
| Top fan-out type match | 65.1% (same dominant search type 2 out of 3 times) |
| Citation domain overlap | 0.339 Jaccard (roughly one-third of cited domains the same) |
The search decision is deterministic. The search strings are random. The search types are moderately stable.
A tool that captures fan-out query strings gives you a snapshot of one random sample. By the next session, 98% of those strings are gone. But the structural pattern -- whether the AI injects brands, compresses to keywords, or seeks evidence -- persists.
This is ChatGPT-specific data (gpt-5.4-mini). Perplexity and Gemini were not replicated in this study. The pattern likely generalizes -- there is no reason to think other platforms would be deterministic at the string level -- but we can only confirm the 91.7% / 98% / 65% figures for ChatGPT.
What This Means for Your Strategy
Step 1: Identify which intent types your customers use
Are they shopping (discovery)? Learning (informational)? Comparing? Seeking reviews? Each intent triggers a different fan-out type distribution.
- Discovery triggers entity injection. Invest in third-party brand presence.
- Informational triggers compression. Invest in keyword-optimized educational content.
- Review-seeking triggers evidence searching. Invest in reviews, case studies, and comparison content.
- Comparison triggers price checking and narrowing. Invest in comparison tables and pricing transparency.
Step 2: Match your content to the fan-out types, not strings
Do not try to optimize for specific fan-out strings like "Bitwarden 1Password NordPass comparison 2026." That string will not exist next session.
Instead, ensure you have content that matches each type:
- For entity injection: Brand mentions across third-party sites (PR, not SEO)
- For compression: Pages that rank for 4-6 word keyword queries (SEO)
- For evidence seeking: Review pages, case studies, comparison tests (content marketing)
- For narrowing: Pages with year, audience, and price qualifiers in titles and headings (on-page optimization)
Step 3: Know which platforms your customers use
ChatGPT is an entity injector. Gemini is an explorer. Perplexity is an evidence seeker. Optimizing for ChatGPT (brand presence + evidence) is a different strategy than optimizing for Perplexity (keyword ranking + evidence) or Gemini (topical breadth).
The short version:
| Fan-Out Type | Strategy | Discipline |
|---|---|---|
| Entity Injection | Get mentioned on third-party sites | PR |
| Compression | Rank for 4-6 word keyword queries | SEO |
| Evidence Seeking | Create reviews, case studies, comparisons | Content Marketing |
| Narrowing | Add year, audience, price qualifiers to pages | On-Page SEO |
| Tangential | Build content clusters on adjacent topics | Content Strategy |
| Expansion | Cannot directly optimize (background context) | N/A |
| Price/Availability | Make pricing visible in HTML | Product Page UX |
Limitations
Replicate analysis is ChatGPT-only. The 91.7% determinism, 98% stochasticity, and 65% type stability are from gpt-5.4-mini. Perplexity and Gemini were not replicated.
Rule-based classification at ~88% accuracy. Approximately 12% of fan-outs may be misclassified, primarily at the compression/tangential boundary.
Mixed ChatGPT model tiers. The ChatGPT fan-out distribution is a composite of three model tiers (gpt-5.4, gpt-5.4-mini, gpt-5.4-nano). Per-model breakdowns show entity injection ranges from 17% to 45% depending on model size.
Citation yield is Perplexity-only. The citation-to-fan-out linkage data comes exclusively from Perplexity. ChatGPT and Gemini yield rates may differ.
Single time period. All data from April 2026. Platform behavior changes with model updates.
Frequently Asked Questions
What is a fan-out query?
When you ask an AI platform a question, it does not search the web with your exact words. It generates multiple internal search queries -- called fan-out queries -- that are the actual strings sent to web search engines. These determine which pages get fetched and can be cited in the response.
Why can I not just optimize for the fan-out strings tools capture?
Because 98% of fan-out strings change between identical runs of the same query. The AI generates fresh search strings every time. A captured string is a snapshot of one random sample. By the next session, ChatGPT is searching for completely different words.
What should I optimize for instead?
The structural type of search, not the specific strings. When the AI decides to inject brand names, it injects different brands each time -- but it is still doing entity injection. When it compresses to keywords, it compresses to different keywords -- but it is still compressing. Match your content strategy to the type: PR for entity injection, SEO for compression, content marketing for evidence seeking, on-page optimization for narrowing.
Does ChatGPT always search the web?
No. The flagship model (gpt-5.4) only searches 29% of the time. Smaller models (gpt-5.4-mini, gpt-5.4-nano) search on every query. Informational questions trigger search only 12% of the time. If your content targets informational queries on the flagship model, the AI may never search for it at all.
Do different AI platforms search differently?
Yes. ChatGPT injects brands from training data (32% of fan-outs). Gemini explores broadly with expansion and tangential searches (48% combined). Perplexity compresses to keywords and seeks evidence (39% combined). Each platform needs a different optimization approach.
Methodology
- 540 parent queries (180 queries x 3 platforms: ChatGPT, Gemini, Perplexity)
- 180 queries: 10 verticals x 5 intent types x 3 queries per cell, plus 30 situation-first reformulations
- 1,323 classified fan-out queries after cleaning
- ChatGPT captured via OpenAI Responses API (3 model tiers: gpt-5.4, gpt-5.4-mini, gpt-5.4-nano)
- Gemini captured via Google GenAI API with Google Search grounding
- Perplexity captured via browser-level SSE interception (Pro account)
- Replicate analysis: 180 queries submitted 3 times to ChatGPT gpt-5.4-mini
- Classification: rule-based taxonomy at ~88% accuracy (9 types, priority-ordered)
- Statistical tests: chi-squared with Cramer's V, Kruskal-Wallis, Mann-Whitney U, log-linear models
References
- Lee, A. (2026). "The Hidden Search Queries AI Runs Before It Answers You." Blog
- Lee, A. (2026). "I Rank on Page 1 -- What Gets Me Cited by AI?" aiXiv
- Semrush (2026). "ChatGPT traffic analysis: Insights from 17 months of clickstream data." Blog
- AIVO Evidentia (2026). "Decision-Path Analysis across AI Recommendation Systems." Zenodo