SEO · November 21, 2026

Stop Optimizing for Imagined AI Search Queries

By Morgan Ross · Senior Technical Lead

A researcher comparing branching search paths on a digital screen

Photo by Nicolas Arnold on Unsplash

Many AI-search audits start by taking one target keyword, asking a language model to invent the related searches an answer engine might run, and scoring a page against that list. The result looks scientific: a coverage percentage, a gap report, and a backlog of headings to add.

The problem is that the list is usually not observed. It is a model’s prediction of a model’s behavior.

That distinction matters because Google’s own description of query fan-out is deliberately qualified. AI Overviews and AI Mode may issue related searches across subtopics and data sources, but Google also says the two experiences can use different models and techniques. A page only needs to be indexed and eligible for a normal search snippet to be eligible as a supporting link; Google does not prescribe a special fan-out markup or technical requirement (Google’s AI features documentation).

A new pre-registered study from Laurelin Labs gives SEO teams a more useful warning. Across 235 non-branded keywords, the researchers observed roughly 2.35 clusters of sub-queries per keyword, while a model asked to guess the likely searches produced 16.1. The guessed list was nearly seven times larger. More importantly, the share of those observed sub-queries a page answered did not improve out-of-sample prediction of citation after core-query rank and page length were accounted for (the Laurelin Labs study).

My conclusion is not that topical depth is irrelevant. It is that imagined fan-out coverage is a poor unit of SEO work. Teams should build useful topic coverage for people, then measure whether the resulting pages earn retrieval and citations. They should not manufacture dozens of hypothetical searches and mistake the resulting score for an algorithmic signal.

What the study actually tested

The Laurelin study is valuable because it tests the commercial premise behind fan-out scoring tools rather than merely counting citations. It used a pre-registered observational design, fixed decision rules before outcome data was pulled, and compared candidate pages within each query.

The sample began with 1,795 keyword rows where Ahrefs reported an AI Overview citation from a domain in the researchers’ audit set. After removing branded and navigational rows, the researchers sampled 235 keyword-country pairs across 46 domains. For each keyword, a candidate was either an organic top-ten page or a page cited in the AI Overview. That design does not represent every search, but it gives the comparison a clear denominator: cited pages versus plausible competitors for the same query.

The researchers then captured fan-out in two different ways:

  • Gemini was called with Google Search grounding three times per keyword, and its searches were read from grounding metadata.
  • Claude was called with web search twice per keyword, and its tool calls were recorded.
  • A separate model was asked to guess the searches an AI system might run, approximating the workflow used by many fan-out tools.

The observed systems did not behave like the imagined-search workflow. Across the analysis frame, Gemini averaged 0.88 searches per answer and ran no search in 331 of 705 calls—47%. The combined observed data produced 2.35 sub-query clusters per keyword. The guessed lists produced 16.1 after merging similar suggestions.

That is not a minor calibration error. It changes the economics of content planning. If a tool generates 16 plausible subtopics for every keyword, a team can spend weeks expanding a page to satisfy an inventory that the target system may never have used. The audit can become self-fulfilling: the more hypothetical gaps it reports, the more content it recommends.

The study also found an important boundary around its claim. Better organic rank for the head query was strongly associated with citation, with an odds ratio of 0.43 per standard deviation of log rank. Longer pages were also associated with citation, at 1.30 per standard deviation. But adding heading similarity, top-ten rank for observed sub-queries, or the share of observed sub-queries answered did not meet the study’s pre-set out-of-sample improvement threshold. The model’s fair top-ten-only baseline had an AUC of 0.741; adding observed coverage changed it to 0.739.

Those figures do not prove that adding a useful section can never help. They show something narrower and more actionable: coverage scores did not add dependable predictive information in this test once stronger retrieval signals were present.

Why guessed fan-out expands so quickly

A model asked to brainstorm related searches is rewarded for completeness and plausibility. A search system answering a user is rewarded for usefulness under latency, retrieval quality, and response constraints. Those are different objectives.

Consider the question “which inventory system is best for a liquidation retailer?” A brainstorming model may produce searches about barcode scanning, warehouse bins, auction workflows, Shopify integrations, eBay fees, purchase orders, mobile offline mode, analytics, returns, and dozens of variants. All are reasonable editorial considerations. They are not evidence that Google issued all of them for a particular answer.

The distinction can be represented as two separate sets:

# Editorial planning: broad and intentional
coverage_plan = {
    "job_to_be_done": ["compare systems for liquidation retail"],
    "useful_subtopics": ["unit tracking", "auction workflow", "channel sync"]
}

# Measurement: only claim what you observed or can reproduce
observed_searches = capture_grounding_queries(answer_id)

if not observed_searches:
    report("fan-out not observed; do not infer a coverage score")
else:
    report(compare_page_to_observed_queries(page, observed_searches))

The first object is a content strategist’s model of user needs. The second is an empirical observation with a provenance trail. Both are useful, but they answer different questions. The first asks, “What should this page help someone do?” The second asks, “What evidence do we have about this retrieval event?” A dashboard that combines them under “fan-out coverage” hides the difference.

Google’s official generative-AI guide points in the same direction. It explains that RAG and query fan-out may help retrieve pages for an answer, but it warns against creating separate pages for every variation of a query. Google says that producing many pages primarily to manipulate rankings or generative responses can violate its scaled-content-abuse policy. Its recommendation is to focus on unique, useful, people-first content rather than reverse-engineering every possible wording (Google’s generative AI optimization guide).

The practical reading is simple: use related questions to test whether a page solves a real user problem, not to manufacture a larger page because a tool says the model might search for it.

The strongest signal is still the page’s main job

The Laurelin result is especially useful for teams that have replaced keyword rank with a fan-out score. It does not return SEO to a 2019 keyword-only model. It says that the main retrieval problem still matters.

A page must first be a strong answer to a coherent job. That means clear intent, useful evidence, technical accessibility, and enough detail to satisfy the person who arrives. If the page is weak for the main question, adding speculative sections for twelve imagined sub-queries is usually a poor trade.

Use this priority order:

PriorityQuestionEvidence to collect
1. EligibilityCan search systems crawl, index, and show the page with a snippet?Indexing status, robots directives, rendered text, canonical URL
2. Core retrievalDoes the page earn visibility for the primary job?Head-query impressions, rank distribution, qualified clicks
3. SatisfactionDoes the page resolve the user’s decision or task?Engagement quality, conversions, assisted revenue, user research
4. Supporting coverageAre important adjacent questions answered naturally?Internal search, support logs, sales calls, query clusters
5. AI citationIs the page selected as supporting evidence?Search Console generative-AI report, cited URL samples

This sequence also protects against a measurement trap. A page can have excellent “coverage” in an audit and still fail to rank, load, explain the product, or convert. A broad content score cannot compensate for a broken canonical, inaccessible text, or a page that answers a different intent than the title promises.

Google’s technical guidance says there are no additional technical requirements for AI Overviews or AI Mode beyond being indexed and eligible for a normal Search snippet. It recommends familiar fundamentals: crawlability, internal links, textual content, page experience, and structured data that matches visible text. That makes technical SEO a prerequisite, not a fan-out optimization layer (Google’s AI-feature guidance).

Related questions still belong in editorial planning. The change is to treat them as evidence about the user’s task, not as a list of algorithmic boxes to tick.

Start with a decision, not a keyword expansion

Write one sentence describing what the visitor must decide or accomplish. “Choose inventory software for a liquidation operation with multiple sales channels” is more useful than “rank for inventory software.” Every proposed subsection must help that decision. If it does not, cut it—even if a tool labels it an uncovered fan-out query.

Source subtopics from your own audience

Use customer interviews, support tickets, internal site search, sales objections, product analytics, and post-purchase questions. These sources are imperfect, but they describe real information needs. They are better editorial inputs than a model inventing a comprehensive list because the list sounds plausible.

Consolidate when the job is shared

Do not create a page for every wording variation. If five questions require the same explanation, answer them in one well-structured section. Create a separate URL only when the audience, task, evidence, or desired conversion is meaningfully different. This follows Google’s warning against multiplying pages primarily to manipulate AI responses.

Attach claims to inspectable evidence

A page is more useful when its important claims can be checked: a test procedure, a configuration, a calculation, a screenshot, a limitation, or a named source. The goal is not to make a paragraph “AI-shaped.” It is to make the page credible to a person and easy for a retrieval system to interpret.

Measure citations as outcomes, not promises

Google now documents a Generative AI performance report in Search Console and says generative features are also represented in the overall Web search type. Its new control lets site owners include or exclude links and content from AI Overviews, AI Mode, and generative features in Discover; the control is not a ranking signal for other Search surfaces (Search generative AI control documentation).

That creates a more honest operating loop:

  1. Publish a page for a defined user job.
  2. Verify crawlability, indexing, and snippet eligibility.
  3. Track ordinary Search performance and business outcomes.
  4. Inspect generative-AI visibility where Search Console exposes it.
  5. Sample cited URLs and compare them with the page’s actual evidence.
  6. Revise when the user’s problem, facts, or product changes—not because a guessed query list grew.

The loop is slower than generating a score, but it produces decisions that can be defended.

A measurement stack that does not pretend to see inside Google

Google does not expose every internal retrieval step behind an AI answer. A third-party tool cannot honestly claim otherwise. Google’s own guidance warns site owners to be cautious with vendors promising ranking success or access to internal AI metrics (Google’s optimization guide).

A practical dashboard should therefore label each field by its evidence level:

Observed:       Search Console impressions, clicks, indexed status, cited URL
Reproducible:   Search result snapshots, controlled prompt runs, page changes
Inferred:       Likely subtopics from users, sales, support, and query data
Speculative:    Model-generated guesses about hidden fan-out queries

Observed and reproducible fields can support performance claims. Inferred fields can support editorial hypotheses. Speculative fields can inspire questions, but should never become a KPI or be presented as a ranking factor.

This complements—not replaces—the asset-level measurement approach described in SEO visibility needs an asset-level measurement stack. The important addition is provenance: every visibility number should say whether it came from a platform report, a repeatable experiment, a first-party audience signal, or a model’s imagination.

The recommendation: build for users, audit for evidence

Query fan-out is a real technique that Google says its AI features may use. But the existence of fan-out does not validate every tool that guesses its contents, and it does not turn speculative coverage into a ranking requirement.

The Laurelin study’s result is a useful corrective: observed fan-out was small, guessed fan-out was much larger, and coverage of the observed set did not improve citation prediction beyond core signals in the tested sample. The study has limits. It is observational, its candidate domains came from an audit set, and Gemini and Claude are proxies rather than Google’s private systems. Those limits make the conclusion narrower, not useless.

For SEO teams, the operating rule should be:

  • Do not create pages or sections solely to satisfy a guessed fan-out inventory.
  • Do use real audience evidence to decide which adjacent questions deserve coverage.
  • Do not report a model-generated coverage percentage as an algorithmic fact.
  • Do separate observed citations, Search performance, and editorial hypotheses.
  • Do not optimize for a hidden query list you cannot inspect.
  • Do make one strong page answer a coherent job with accessible, verifiable evidence.

The best AI-search strategy is not to predict every search an answer engine might make. It is to publish something worth retrieving, make it technically reachable, and measure what actually happens. That is less theatrical than a fan-out score—and far more useful when the score stops matching reality.

Ready To Scale
Your Ecommerce?

Let Decipher Tech handle your ecommerce operations end-to-end. Our ECOMAS model means you get a full team of specialists without the overhead of hiring in-house.