Why AI Cites You for Some Queries but Ignores You for Others

A business owner reaches out: their HVAC company shows up when someone asks "best HVAC company in [city]" but disappears completely when someone asks "how much does HVAC installation cost in [city]." Same business, same week, same platforms.

This is not a glitch. It is not random. And it is not fixed by general content optimization.

Research published in the past 90 days -- including the largest controlled citation study accepted at SIGIR 2026, the premier academic venue for information retrieval -- explains exactly what is happening.

AI visibility is not a composite score

The standard framing for AI search performance is a score: a composite measure of how visible your business is to AI engines. That framing is useful for diagnosis but misleading as a description of real-world performance.

In November 2026, researchers at Nanyang Technological University and York University published arXiv:2611.04225, examining how GEO (generative engine optimization) effectiveness shifts across different user intent states. Their framework identifies two dimensions that govern AI citation probability: how specific the user's need is, and how far along in their decision process the user is.

A query like "HVAC company" is low-specificity and early-stage. A query like "how much does it cost to replace a furnace in Burlington" is high-specificity and late-stage. Their finding: optimization methods that improve citation rates for one intent state often produce no measurable gain for other intent states. Some query types respond; others remain unaffected regardless of content changes.

This is the problem with a composite score. If your HVAC company is well-positioned for category queries but completely absent for cost queries, an aggregate score may look acceptable while you are missing the entire late-decision buying window.

Four conditions that eliminate citation odds

Our November 9, 2026 investigation of arXiv:2605.25517 gives the mechanism. This study -- accepted at SIGIR '26 -- ran 252,000 controlled citation trials across six AI models. Each trial presented two candidate sources differing in exactly one factor. The question: which document wins the citation slot?

The result was not a ranking of factors to balance. It was four gatekeeper conditions. Failing any single one eliminated citation odds regardless of how the page performed on everything else.

**Topic mismatch.** A document that does not directly answer the specific query variant being asked loses regardless of its other strengths. This is not about general category relevance -- it is query-specific relevance. A service page titled "Plumbing Services Toronto" is correctly categorized. It is not an answer to "how fast can a plumber arrive for a burst pipe tonight." These look like the same topic to a business owner; they look like different documents to an AI retrieval system.

**Price not mentioned.** For queries where cost is a primary buyer concern, absent pricing information is disqualifying. The study maps this directly to service verticals: HVAC, plumbing, roofing, electrical, dental, legal, moving, pest control, home cleaning. Our 2026-11-10 methodology investigation found this in the data and stated it plainly: "Phase 1 fixes -- NAP corrections, directory presence, schema improvements -- will improve a business's position in the retrieval pool but cannot compensate for a gatekeeper failure at the content level." If your service pages contain no pricing information -- no ranges, no "starting at," no cost-factor explainers -- you are excluded from cost queries as a category. Not penalized. Excluded.

This is what explains the HVAC company's disappearance from cost queries. The page can have correct directories, fresh content, and appropriate vocabulary -- and still fail the price-absent gatekeeper because that gatekeeper only activates for cost queries. Schema improvements will not fix it. A new city page will not fix it. Adding pricing context to the page will.

**Old timestamp.** Outdated content loses to recent content. This is a gatekeeper, not a margin factor. A stale page does not get partially penalized -- it is knocked out before other factors apply. The SIGIR 2026 data converges with our existing content freshness findings: pages not updated within roughly 90 days face meaningful exclusion risk on competitive queries, independent of their other signals.

**Low list position.** The source's position in the AI engine's retrieval candidate pool determines citation probability at a structural level. Getting into the top of the retrieval pool requires the same signals that drive traditional organic search performance: entity recognition, directory presence, domain authority, relevant third-party mentions. This is a Phase 1 infrastructure problem, and no content optimization compensates for it. The SIGIR 2026 paper states it directly: a document at a lower list position loses to a document at a higher list position regardless of content quality.

Why diagnosing the right failure mode matters

A third piece of research from our November 9, 2026 investigation (arXiv:2603.09296) adds the diagnostic dimension.

This framework -- AgentGEO -- makes a point that changes how to think about fixes. Generic optimization rules are low-leverage because they do not diagnose which failure mode applies to a specific document. The researchers identify three failure stages:

**Parsing failures.** The retrieval system cannot correctly extract the document's content. Schema gaps, structured data errors, and format problems that prevent clean extraction live here.

**Fetching and context failures.** The document is retrieved but insufficient context matches the query. Thin content, vocabulary that does not match buyer language, missing specificity to the query variant being tested.

**Generation-stage failures.** The document is in the retrieval pool and readable, but the AI model selects other sources when constructing its final answer. Relevance, authority, and position signals determine the outcome here.

Applying content rewrites to a page with a parsing failure wastes effort. Applying schema fixes to a page with a generation-stage failure misses the point.

The gatekeeper framework from the SIGIR 2026 study tells you which gatekeeper you are failing. The AgentGEO failure taxonomy tells you what type of fix addresses it.

A price-absent failure is a fetching and context failure -- the relevant content does not exist on the page, so the page cannot match cost queries regardless of how well it is structured. Adding pricing information is the fix.

A low list position failure is a generation-stage failure that traces back to Phase 1 infrastructure. It requires directory presence, entity signal, and organic ranking improvements -- not content rewrites on the existing page.

What to check first

The first step is confirming which query types you are winning and which you are not. Generic AI visibility scores test a handful of category queries and return a composite. That composite can look fine while masking specific failures.

A Signal Check on Sourcepull tests your domain across multiple query types -- category queries, service-specific queries, local queries -- and returns per-query results. If your score looks acceptable overall but specific query types return consistently low or zero citations, that variance is the signal worth investigating.

If cost queries consistently return low scores: check your service pages for pricing content. The price-absent gatekeeper is the most common finding we see in service business audits, and it is the one fix that cannot be substituted -- improving directory presence or schema will not compensate for it.

If service-specific queries miss: check for vocabulary alignment. The gap between "professional plumbing services" (what the business wrote) and "how fast can someone get here for a burst pipe" (what the buyer asked the AI) is not obvious from the business side.

If your overall score is low across all query types: the problem is likely Phase 1 infrastructure -- entity recognition, directory presence, retrieval pool membership -- and content work comes after.

The intent state gap that showed up in the November 2026 research is real, but it is not mysterious. The gatekeeper framework gives you four specific conditions to check. Most businesses fail one or two of them, not all four. The diagnostic question is which one.