The SERP is an input.
Not the answer key.
“How does AI pull from rankings?” sounds like a request for a cutoff. The more useful question is: which retrieval job produced the candidate, and what happened to it next?
Imagine asking for a CRM that suits a 20-person team, stays within budget, connects to your existing tools, and meets a security requirement. A conventional result list might rank broad buying guides for that complete request. A generative system can split the work: look up seat pricing, check integrations, locate security documentation, compare migration support, then assemble a recommendation with qualifications.
Each search can produce its own ordering of pages. A page that barely appears for the original sentence might be an excellent match for one subquestion. Conversely, the first result for the original sentence might add nothing once its claims have been covered by clearer or more current sources. That is the central distinction behind query fan-out.
Google uses query fan-out for concurrent, related model-generated searches. Other platforms describe query rewriting, repeated search, multi-search, or iterative grounding. These behaviors are related, but their names do not establish an identical architecture. Google’s official explanation directly connects its generative features to the Search index and core ranking systems.
A defensible core statement
AI search may turn one prompt into several related searches, retrieve ranked candidates for each, select useful supporting material, and display a subset of sources. Traditional search remains an important upstream foundation. The exact-prompt top 10 is not the complete evidence universe.
What a provider discloses
A platform states that it rewrites queries, runs searches, exposes source metadata, or uses a particular retrieval component. The claim applies to the named surface and version.
What a study measures
Researchers compare source sets, rankings, or repeated answers. An association does not identify the proprietary mechanism or prove a ranking intervention caused a citation.
What helps explain it
A worked example or research architecture explains a plausible mechanism. It is not a reverse-engineered trace of a live commercial answer.
Throughout this guide, “derived” identifies a calculation from published figures; “illustrative” identifies invented teaching data. Neither is presented as a vendor disclosure.
Eight stages.
Several places to disappear.
This vendor-neutral process combines disclosed components with information-retrieval research. The exact order, number of iterations, and selection rules vary.
From a prompt to a source-backed answer
Follow the stages left to right, then continue on the next row. The illustration is not a measured conversion funnel.
Decide whether live or external information is needed.
Turn the task into one or more retrieval queries.
Rank results separately for each search.
Combine overlapping lists and remove duplication.
Assess passages for relevance, support and coverage.
Choose what information reaches the generator.
Compose claims using evidence and model knowledge.
Attach source links according to the product’s interface.
Search activation is the first gate
Not every AI answer is a live-web answer. A system may use model knowledge, conversation context, connected files, a structured data source, or external search. Freshness, complexity, the selected mode, and tool availability can affect that decision. If search does not activate, the current organic result list cannot directly supply that run through live retrieval.
Do not treat an unobserved search event as proof of no search. Some interfaces hide tool activity. In a dataset, distinguish search observed, search not used when explicitly reported, and unknown. OpenAI documents quick search, more iterative agentic search, and deeper research workflows; these are not interchangeable measurement strata. OpenAI tool documentation.
Rewrite, decomposition, and iteration solve different problems
A rewrite makes a request more retrieval-friendly. It can remove conversational phrasing, resolve a reference from earlier messages, or add a location. Decomposition separates a complex task into answerable subquestions. Fan-out broadens retrieval across related information needs, sometimes concurrently. Iteration means new searches are chosen after inspecting earlier evidence.
Parallelism and iteration can coexist. A first wave might find product names; a second wave might check the security claims of those products. The later queries depend on what the earlier searches found. An independently generated list of “likely fan-out keywords” is therefore a hypothesis, not a record of the model’s actual execution. ChatGPT Search’s documented rewrite and follow-up behavior.
Rank belongs to the retrieval query
A conventional organic position is scoped to a search engine, query, locale, device, and collection time. An internal retrieval rank is scoped to the system’s own query and candidate universe. Comparing a Perplexity citation with a Google ranking is an external overlap test, not a measurement of Perplexity’s internal rank.
Retrieval may combine lexical matching, which is sensitive to the actual terms, with semantic matching, which can identify related meanings. A named error code or model number can benefit from exact matching. A differently worded explanation can benefit from semantic retrieval. Perplexity’s published Search API architecture describes a hybrid candidate set, filtering, and progressively stronger rankers. That infrastructure disclosure does not publish every Pro Search citation rule. Perplexity’s architecture.
A candidate pool is not a single sorted SERP
Several lists can contain the same URL, different URLs carrying the same material, or complementary pages answering different constraints. A reconciliation stage can collapse duplicates, normalize references, or keep distinct evidence for different facets. Scores from different retrieval systems may not even be calibrated on the same scale.
Research systems demonstrate decomposition, merged candidate pools, and reranking. Commercial products do not disclose a common merge formula. In particular, the popularity of Reciprocal Rank Fusion in tutorials does not establish that Google or ChatGPT uses a fixed RRF configuration. We will use RRF below only as a transparent teaching example. Question Decomposition for Retrieval-Augmented Generation.
The generator needs usable evidence, not everything retrieved
The evidence budget is finite. A broad guide can be relevant without containing the current number, precise exception, or direct support a claim needs. A short official table can contribute more to one sentence than several long general pages. Passage selection is about information utility within the answer, not a universal preference for short content.
Coverage also matters collectively. Three individually relevant documents may repeat one another; a fourth may supply the missing limitation. RichRAG studies multi-faceted retrieval and listwise document selection. Uplift-RAG examines whether retrieved information improves generation beyond internal model knowledge. These research findings explain why ranking and answer utility can diverge; they do not reveal vendor weights. RichRAG · Uplift-RAG.
Claims and citations are another selection layer
The answer can combine a number from one source, a definition from another, and a caveat from a third. It can also use internal model knowledge. A source’s presence in the retrieved set does not guarantee its wording, brand, or link will appear in the response.
OpenAI’s web-search API distinguishes the broader consulted-source list from inline citations. This supports a specific API-level separation between consultation and displayed links. It does not establish that every ChatGPT interface exposes identical events. A citation should be checked against the claim it accompanies, not assumed to validate an entire paragraph. OpenAI’s sources documentation.
A weak original rank.
A strong evidence fit.
The following CRM example is invented to make the mechanics visible. It is not a trace from Google, ChatGPT, or any other provider.
Start with: “Which CRM fits a 20-person team, connects to our email platform, and offers single sign-on within our budget?” Three hypothetical searches explore pricing, integrations, and security. Each candidate’s rank changes with the question.
The same document can compete in several query families
Darker blue means a stronger rank in this invented example. Green outlines identify a possible evidence match, not a confirmed citation.
| Candidate page | Original prompt | Seat pricing | Integrations | SSO security |
|---|---|---|---|---|
| Broad CRM buying guide | 1 | 18 | 24 | Not found |
| Official pricing table | 38 | 1 | 40 | Not found |
| Integration documentation | Not found | Not found | 1 | 21 |
| Security & SSO reference | 62 | Not found | 32 | 2 |
| Older comparison article | 2 | 6 | 7 | 15 |
Why the first result can disappear
Its information may be redundant, too general, out of date for the claim, difficult to extract, or irrelevant to a later subquestion. It can also be consulted without becoming a visible citation. These are possible explanations, not a diagnosis of a specific answer.
Why a “non-ranking” page can appear
It may rank for another query, exist in another index or data source, cover a missing exception, or be the primary record for a fact. “Outside Google’s top 10 for the original prompt” does not mean “retrieved from nowhere.”
Do not reverse the inference
A citation outside the original top 10 is consistent with fan-out, but does not prove a particular hidden query ran. It could also reflect another source corpus, different extraction rules, query context, or later evidence selection. Only an exposed execution trace can establish the executed query.
Retrieval asks “relevant?”
Grounding asks “usable?”
A result can be relevant to the topic while providing weak support for the exact sentence the system needs to write.
Consider a claim that a CRM offers single sign-on on a specific plan. A product homepage establishes identity, a review establishes experience, and a security document establishes the plan-level feature. These are different evidentiary roles. A stronger answer may need all three, or only the security document, depending on the question.
Grounding adds concerns such as provenance, freshness, factual fidelity, coverage, and contradictions. A source that says “all plans” and a newer source that names one enterprise plan are not interchangeable. The model must either resolve the disagreement or qualify the answer. Microsoft’s engineering discussion describes this change in objective without disclosing a complete current Copilot pipeline. Bing on ranking pages versus supporting answers.
The unit of selection gets more specific
Relevance can be individual
A retriever can ask how well a document matches a query. Later rankers may inspect the query and document jointly, allowing a more expensive assessment. Cross-encoder ranking is one published example in Perplexity’s architecture.
Utility can be collective
The final set should cover the task without wasting context on repetition. Research into generator-friendly selection investigates this gap. It does not justify inventing a universal “GEO evidence score.” Retriever and LLM preference research.
What a rank-fusion formula can explain
Reciprocal Rank Fusion combines positions from several lists. This tiny example shows the mathematics only. It is not Google’s, OpenAI’s, or any named product’s disclosed citation algorithm.
Sum only across lists where the document appears.
A larger k makes small position differences matter less relative to appearing in multiple lists. Changing k never proves how a real answer engine merges candidates.
| Document | Query A rank | Query B rank | Query C rank | RRF score |
|---|---|---|---|---|
| Specialist | 1 | Not returned | Not returned | 0.016393 |
| Broad coverage | 4 | 3 | 5 | 0.046883 |
| Complementary source | 9 | 2 | Not returned | 0.030622 |
In this example, Broad coverage has the largest fused score. A later evidence-selection stage could still prefer another document.
The lesson is not “appear everywhere.” It is that a combined ordering can differ from the first result in any single list. Fusion also does not establish factual support: a frequently retrieved mistake is still a mistake.
Passage clarity is not a magic word count
Clear headings, direct answers, units, dates, and qualifications make information easier for people to interpret and for retrieval systems to identify. That does not create an official paragraph length, entity density, or answer-template rank factor. Google’s passage-ranking documentation describes understanding relevant sections, but does not say that the same component directly selects every AI-answer citation. Ranking systems guide.
There is no universal
AI citation formula.
Keep the product, surface, mode, API, and version attached to every technical claim. A developer API disclosure is not automatically a specification for the consumer app.
GGoogle AI Overviews & AI Mode
- Disclosed
- Related concurrent queries can gather information across subtopics and sources. Retrieval uses the Search index and core ranking systems. Supporting pages need the applicable index, snippet, and generative-feature eligibility.
- Important limit
- No public fixed fan-out count, per-query candidate cutoff, list-merging equation, or evidence-weight formula. The original-query first page is not the full disclosed retrieval scope.
- Current control
- Check that the site has not been excluded from Search generative AI features, including by an inherited Search Console setting. “Include” is the default; this is not a new requirement to manually opt in.
Optimization guide · AI features and websites · Inclusion control
CChatGPT Search & OpenAI web search
- Disclosed
- ChatGPT can rewrite a prompt into targeted searches and issue narrower follow-ups after reviewing initial results. Context such as approximate location and enabled memory can affect rewriting. OpenAI’s web-search API exposes consulted sources separately from inline citations.
- Important limit
- Do not reduce ChatGPT Search to “Bing’s top 10.” Third-party search and other sources can participate, while provider weights, candidate depths, merging rules, and citation scoring are not fully disclosed.
- Keep separate
- A tool response with structured source metadata and a consumer Sources drawer are different observability surfaces. Neither automatically reveals organic rank or the complete causal path for a displayed citation.
GGemini API with Google Search grounding
- Disclosed
- The model can produce one or multiple searches. Depending on the interface, executed queries and source-support metadata are returned. The current grounding guide illustrates queries inside the search call’s arguments and text annotations pointing to sources.
- Important limit
- Exposed queries improve auditability, but do not provide organic positions, every rejected candidate, or proprietary ranking scores. Consumer Gemini Apps should not be assumed to expose or implement the same interface.
Two metadata formats, not one interchangeable schema
In the grounding guide’s illustrated format, search steps have type: "google_search_call" and the queries are inside arguments.queries. GenerateContent separately documents groundingMetadata.webSearchQueries, groundingChunks, and groundingSupports. Keep these field paths associated with their own API schemas.
Illustrative field map, not an API request:
Search-call interface
steps[] → type = google_search_call
→ arguments.queries
GenerateContent interface
groundingMetadata → webSearchQueries
→ groundingChunks
→ groundingSupportsMBing & Copilot Search
- Disclosed
- Copilot Search uses Bing results for the user query and additional searches issued on the user’s behalf. Bing’s engineering writing describes grounding as an extension of search toward supportable information and evidence.
- Important limit
- The historical 2023 Prometheus architecture provides context, not a guarantee of every current implementation detail. A 2026 engineering discussion of grounding is not a published end-to-end Copilot citation formula.
- What to measure
- Bing’s AI Performance reporting gives aggregated citation information and sampled grounding phrases. It explicitly does not turn those counts into a ranking or placement measure.
Copilot Search · Bing grounding · AI Performance announcement
PPerplexity Pro & Search API
- Disclosed
- Pro Search conducts multiple searches. Perplexity’s Search API architecture combines lexical and semantic retrieval, a hybrid candidate pool, early filtering, stronger downstream rankers, and document plus subdocument scoring.
- Important limit
- This is published search infrastructure, not a complete guarantee about source selection for every Pro answer. Google rank is a comparison benchmark, not Perplexity’s own internal retrieval position.
AClaude web search
- Disclosed
- Claude can search repeatedly within a request. Supported API tool versions, including
web_search_20260209and later, can dynamically filter results before they enter model context. - Important limit
- Dynamic filtering should not be generalized to all older tools, hosting environments, or consumer Claude.ai runs. The upstream provider, rank weights, and citation-selection formula are not publicly specified in the cited documentation.
Similar words.
Different denominators.
“Overlap” can mean a share of citations, a share of queries, a conditional probability, or the similarity of two sets. Those are different questions.
The empirical record supports two ideas at once: conventional rank is associated with citation, and the original-query result list explains only part of the final source set. The magnitude depends on the product and measurement. A Google AI Overview study cannot be used as a universal ChatGPT ranking rule.
Every chart below identifies the unit, cutoff, dates, and source. The original studies are observational. Changing a website because a factor correlates with citation is not the same as demonstrating that the change caused an improvement.
Where cited AIO URLs appeared in organic results
Ahrefs • Published March 2, 2026
863K SERPs • 4M AI Overview URLs • Same original query
Blue links only, excluding ads and SERP features.
top 10
- Positions 1 to 1037.1%
- Positions 11 to 10026.2%
- Absent from the top 10036.7%
The denominator is the cited-URL set, not all queries and not all ranking pages. Absence from this same-query top 100 does not identify the alternative retrieval route.
Ahrefs primary study. The publication date is shown; the article does not specify an exact extraction date for this update.
Why the headline numbers are different
Ahrefs also analyzed mixed SERP blocks, including ads and features. That separate cut is 37.9%, 31.2%, and 31.0%, totaling 100.1% because of rounding. Do not replace one bucket from the blue-link distribution with a mixed-block number.
90% of queries is not 90% of citations
seoClarity • Data: October 12, 2025
362K US desktop queries that produced AIOs • 5.1M citations
Query-level overlap
At least one cited URL matched an organic top-10 result. Each square represents one percentage point, not a raw observation from a 100-query study.
Citation-level composition
Derived estimate
The study reported 56% of citations in the top 20. Of that subset, 73% ranked in positions 1 to 10. So 56% × 73% = 40.88%, or approximately 41%.
90% does not mean 90% of citations came from the top 10. A query can count as an overlap success with one matching URL and many citations from elsewhere.
View the denominators and calculation
| Metric | Value | Denominator |
|---|---|---|
| Any top-10 match | 90% | AIO-producing queries |
| Any top-20 match | 94% | AIO-producing queries |
| Top-20 citation share | 56% | All citations |
| Positions 1 to 3 within top 20 | 32% | Top-20 citations only |
| Positions 4 to 10 within top 20 | 41% | Top-20 citations only |
| Positions 11 to 20 within top 20 | 27% | Top-20 citations only |
| Derived top-10 share | 40.88% | All citations; 56% × (32% + 41%) |
| Derived positions 11 to 20 | 15.12% | All citations; 56% × 27% |
37.1% and approximately 41%: a better comparison, not a replication
Both now refer to a top-10 citation bucket. That makes the comparison more informative than setting 37.1% beside a top-20 figure of 56%. It does not align collection dates, query populations, extraction methods, or weighting. The results are directionally compatible, not a universal “40% rule.”
Higher organic positions were more often cited
For each reported organic position, how often was that position’s URL included in an AIO? This is a different denominator again.
Common horizontal scale: 0% to 50% inclusion
Only the four reported positions are drawn. There is no invented curve for positions 4 to 19. Shared factors such as relevance or source quality could influence both ranking and citation; this is not a causal estimate of a one-position ranking improvement.
seoClarity, October 12, 2025. Rank-conditioned inclusion, not share of the citation pool.
Original-prompt overlap varied across assistant series
Ahrefs • Collected early July 2025 • Published August 11, 2025
15,000 long-tail prompts • Exact cited/referenced URLs compared with Google’s top 10 for the original prompt
Common horizontal scale: 0% to 30% URL overlap
The often-quoted 11.9% is the unweighted mean of these five series. It is not a pooled citation share, not a four-product average, and not a Google AI Overview measurement. The two ChatGPT series distinguish different reference placements.
Ahrefs original study and chart. Historical product behavior, not a current league table of assistant quality.
When a line chart is appropriate
BrightEdge • Published February 12, 2026
Approximate monthly AIO-source overlap observations for February to July 2025. Both series come from the same report.
View all monthly values and limitations
| Observed month | Top 10 | Top 100 |
|---|---|---|
| February 2025 | 16.4% | 48.7% |
| March 2025 | 16.1% | 49.7% |
| April 2025 | 16.9% | 50.8% |
| May 2025 | 16.1% | 51.0% |
| June 2025 | 16.8% | 52.8% |
| July 2025 | 16.6% | 53.1% |
The report does not disclose a query count or a complete weighting specification. Values are approximate. Its rounded “about 17%” summary should not turn the February 2025 value of 16.4% into a February 2026 observation.
A line connects this report’s own dated observations. It does not connect BrightEdge to Ahrefs or seoClarity. A within-report series is more coherent than a cross-study collage, but it is still not a controlled causal experiment.
Different answer surfaces selected different source sets
Grossman et al. • SIGIR 2026 paper
Benchmark collection December 7 to 8, 2025 • Table 2 analysis conditioned on 7,439 queries with sources from all three systems
| Compared sources | Traditional Google | AI Overviews | Gemini 2.5 Flash |
|---|---|---|---|
| Traditional Google | 1.00 | 0.18 | 0.16 |
| AI Overviews | 0.18 | 1.00 | 0.11 |
| Gemini 2.5 Flash | 0.16 | 0.11 | 1.00 |
Jaccard divides the shared URLs by the combined unique set. A value of 0.18 does not mean 18% of citations ranked in the top 10. It answers a set-similarity question. The full benchmark and its time-sensitive additions are larger than this three-way-source subset.
Grossman et al., Table 2. Historical benchmark conditions, not a current measurement of all Gemini models.
The dates are part of the result
Ahrefs’ cross-assistant sample. Five citation/reference series compared with the original-prompt Google top 10.
seoClarity’s AIO query and citation analysis. Approximately 41% top-10 citation share is a derived result.
Ahrefs’ updated AIO analysis. Broader citation parsing and a distinct blue-link-only cut.
Ahrefs compares its earlier approximately 76% finding with a later approximately 38% finding, while also reporting improved parsing. Changes in extraction and systems can both matter. That comparison does not isolate the effect of time, nor prove that more aggressive fan-out alone caused the difference. Cross-study charts should keep such methodological changes visible instead of drawing a single smooth decline.
One overlap.
Three correct answers.
A citation share, a Jaccard score, and an any-match query result can describe the same source sets and produce very different numbers.
Change the shared sources
The organic result set contains 10 unique URLs. Choose the number of unique cited URLs and how many appear in both sets. This experiment concerns one answer, not a panel-level average.
2 shared ÷10 cited URLs.
2 shared ÷18 unique URLs in the union.
This one query counts as a match because at least one URL is shared.
With 10 cited URLs,10 organic URLs, and 2 shared URLs, the citation hit rate is 20%, Jaccard is 11.1%, and this single query counts as an overlap success. None of the numbers is wrong. The mistake is reporting one as another.
Pooled versus averaged changes the answer
A prompt with 20 citations contributes more observations to a pooled citation share than a prompt with 2 citations. An unweighted mean of per-prompt percentages treats those prompts equally. State which approach you use. Do not call an average of product-series percentages the share of all citations without the underlying weights.
Domains versus URLs changes the match
A domain-level match can count a homepage and a pricing page as equivalent. An exact-URL match does not. Canonicalization choices such as redirects, tracking parameters, protocols, and trailing slashes also affect the result. Preserve the raw URL and document the normalization rule.
Jaccard = |set A ∩ set B| / |set A ∪ set B|
Any-overlap query rate = queries with ≥1 match / eligible queries
A visibility report
is not a retrieval trace.
The absence of an observable event is not always the absence of the underlying event. Match each data source to the claim it can actually support.
Google Search Console
Google announced dedicated Generative AI performance reporting in June 2026 and added a worldwide-rollout note on August 31. Current documentation describes AI Overviews and AI Mode impressions, with page, country, device, and date dimensions. Property-level chart aggregation and page-level table aggregation are not interchangeable.
No hidden fan-out-query field is documented. For AI Overviews in standard position reporting, links share the containing element’s position. That does not create an ordering of individual citations, and should not be generalized to all AI Mode elements.
Bing Webmaster Tools
The AI Performance public preview describes citation activity for Microsoft Copilot, Bing AI summaries, and selected partner integrations. It includes cited pages and sampled grounding phrases, rather than a complete per-answer internal execution trace.
Counts are not rank, prominence, authority, clicks, or causation. A grounding phrase is useful evidence about the reported activity, but it should not be represented as every user’s exact prompt. Keep the report’s scope beside any trend you publish.
A crawler visit does not prove a citation.
Search crawlers, training crawlers, and user-triggered fetchers can serve different purposes. Verify the bot using the provider’s documented method, not just a spoofable user-agent string. Successful access is evidence of access, not a ranking boost or a conversion.
OpenAI crawlers · Anthropic crawler roles · Perplexity crawlers
Build an evidence ledger for a run
When an interface permits collection, store the prompt, timestamp, locale, mode, answer, citation URLs, and claim associations together. Add exposed queries when available. Leave unavailable fields explicitly unknown. Do not fill a missing execution trace with AI-generated guesses and then analyze those guesses as observed fan-out.
Illustrative run record, not a vendor response schema:
run_id: an internal stable identifier
surface_and_mode: the tested product and selected mode
prompt_and_context: exact wording plus allowed context
collected_at_and_locale: timestamp, country, language, device
search_activation: observed | explicitly absent | unknown
executed_queries: exposed values, otherwise unknown
cited_urls_raw: exact links shown by the product
cited_urls_normalized: links after documented normalization
claim_source_pairs: claim text + supporting URL
original_query_ranks: aligned organic snapshot
limitations: blocked, missing, or incomplete fieldsMeasure outcomes.
Keep the null results.
A single answer screenshot is a case study. A repeatable prompt panel can become a measurement system, if the sampling and denominators are designed first.
Choose the outcome before you collect answers
Decide whether the experiment concerns citations, brand mentions, favorable recommendations, impressions, referrals, or conversions. A neutral mention is not an endorsement. A citation to a third-party review is not a citation to your site. A brand can be visible in the answer while its own URL receives no traffic.
Write the inclusion rule and the primary endpoint in advance. For example, “the proportion of scheduled runs that cite this normalized URL” differs from “share of all cited URLs that belong to this domain.” Both can be useful. They should not be collapsed into one opaque AI-rank score.
Use a balanced prompt panel, not only favorable prompts
Include definitions, comparisons, procedures, constraints, current facts, troubleshooting, and relevant local or language variants. Separate clean sessions from conversational follow-ups. Log the surface, mode, and any visible model identity. Preserve a holdout prompt family that did not shape the content changes.
Natural paraphrases help assess sensitivity, but many paraphrases of the same task are not independent customers. Keep them nested inside a canonical prompt family. Otherwise a large-looking sample can mostly repeat one information need and produce artificially narrow uncertainty estimates.
Use aligned treatment and comparison periods
Treatment + control pages
Preserve comparison pages
Do not assume causation
Report uncertainty
Record instability rather than explaining it away
Repeat across runs and dates. Source sets can change with freshness, query context, sampling, and product updates. Academic work on generative search examines this variability and the effect of small query changes. A stable-looking single answer should not stand in for a distribution. Kirsten et al. · Grossman et al..
Predeclare an acceptable confidence-interval width or a minimum effect you care about. There is no universal minimum number of runs. When estimating uncertainty, account for clustering by prompt family, with paraphrases and repeated runs nested within families. Treating every response as independent can overstate precision.
| Metric | Numerator | Denominator | Interpretation limit |
|---|---|---|---|
| Observed activation rate | Runs with search or the feature observed | All scheduled runs | Unknown activation must remain labeled unknown |
| Conditional citation rate | Activated runs citing the target | Activated runs | Excludes the activation gate by definition |
| Unconditional citation rate | Runs citing the target | All scheduled runs | Retains no-search and no-citation outcomes |
| Citation share | Target citations | All observed citations in the same stratum | Influenced by citation-rich prompts |
| Mention rate | Answers naming the entity | All relevant answers | Can include neutral or negative mentions |
| Recommendation rate | Answers favoring the entity | Applicable recommendation answers | Needs a consistent human-checked coding rule |
| Citation fidelity | Reviewed claim-source pairs with adequate support | All reviewed pairs | A link alone does not establish support |
| Referral conversion rate | Attributed conversions | Attributed referral sessions | Does not measure zero-click answer exposure |
Never drop the invisible runs
Keep scheduled runs with no AIO, no observed web search, no citation, or no target mention. Dropping them turns a conditional success rate into an apparent overall visibility rate. Similarly, a blocked or failed collection is missing data, not a successful nonappearance observation.
Separate a site intervention from a product update
If citation prevalence changes after a content edit, plausible causes still include recency, an index update, a model change, a different query mix, seasonality, or altered source extraction. Check contemporaneous controls and retain collection logs. Report uncertainty and competing explanations instead of announcing that one formatting tweak “unlocked” AI rankings.
Make the evidence useful.
Make the process testable.
Understanding retrieval should improve the quality of the work, not produce a new collection of unsupported ranking rituals.
Map a task, not every possible phrase
Organize the information needs behind the user’s decision: requirements, constraints, comparisons, evidence, exceptions, and implementation. Use real customer questions and support issues. Generated query expansions can help brainstorm, but should remain labeled hypotheses.
One coherent page may answer several connected subquestions. A new URL is justified by a distinct intent, audience, or format, not simply by a different guessed fan-out phrase.
Publish information you can defend
Original tests, current specifications, methods, units, dates, and explicit limitations create evidence that is useful beyond surface-level keyword matching. Name the author or reviewer when relevant and keep update claims tied to substantive changes.
A citation to a fabricated statistic does not become evidence because it looks scholarly. Supportability is a property of the underlying information.
- Keep access and eligibility healthy. Check responses, canonicalization, internal links, relevant crawler access, snippet controls, and the platform-specific inclusion settings you actually use.
- Make context explicit. A number needs units, a date, a population, and a method. A product claim needs the relevant plan, version, region, or exception.
- Resolve contradictions visibly. Explain why an older value changed instead of leaving incompatible statements on different pages.
- Measure at the right level. Preserve URL-level evidence, aggregate to domains deliberately, and distinguish mentions from recommendations.
- Prefer changes you can attribute. Document one meaningful intervention and a credible comparison, then evaluate repeated outcomes.
Black-hat shortcuts confuse vulnerability with strategy
Query fan-out can be misread as an invitation to manufacture endless near-duplicate pages, fake authority, or manipulate model instructions. Those approaches range from unsupported to policy-violating. A system being susceptible to a tactic in an experiment does not establish durable benefit in a changing commercial product.
| Failure mode | Why it is a bad inference | Defensible alternative |
|---|---|---|
| Fan-out page spraying | Guessed queries do not prove demand or execution; near-duplicates can become doorway or scaled-content abuse. | Cover distinct needs with coherent, substantive resources. |
| Unnatural entity repetition | Semantic retrieval does not turn keyword stuffing into useful information. | Use accurate names, relationships and explanations naturally. |
| Fake research or credentials | Imitating evidence creates false support, not authority. | Show real methods, ownership and limitations. |
| Hidden prompt injection | Instructions in retrieved content try to cross the boundary between untrusted evidence and system behavior. | Treat retrieved text as data; audit claim support and instruction contamination. |
| Magic files or chunk sizes | No universal AI ranking boost follows from an invented file or arbitrary paragraph length. | Use supported platform controls and clear human-readable content. |
Policy and security context: Google spam policies · Bing Webmaster Guidelines · Indirect prompt-injection research · Research on product-visibility manipulation. These are risk explanations, not a recipe for manipulating an answer engine.
The terms that
cause the most confusion.
Is “AI fanning” the same thing as query fan-out?
“AI fanning” is informal marketing language. Query fan-out is the documented Google term for related concurrent searches. Other systems may rewrite, decompose, or iterate without using that label. Prefer the product’s own terminology when describing a disclosed behavior.
Does a low original-prompt overlap mean SEO no longer matters?
No. It means that the measured original-query list accounts for a limited share of the observed source set. Search access, indexing, quality, relevance, and retrieval can still be foundational. The unobserved query family and later selection stages can both affect the outcome.
Does the citation order reveal the retrieval order?
Not unless the platform explicitly documents that relationship. Citations may follow the sequence of claims, source-support mapping, or interface decisions. A first citation is not automatically a first-ranked candidate.
Can we discover the exact hidden fan-out queries?
Sometimes an API exposes executed queries. Consumer products often do not. A model asked to imagine its searches can produce plausible suggestions, but that is not an execution log. Label synthetic expansions as hypotheses and keep them separate from exposed metadata.
Are retrieval rank, rerank score, and citation probability interchangeable?
No. Retrieval rank orders candidates for a retrieval query. A reranker may reassess relevance or usefulness. Citation probability is a measured or modeled chance of visible inclusion under stated conditions. None is a universal cross-platform “AI rank.”
What is a passage, and is it different from a chunk?
A passage is a meaningful section of content. A chunk is often a system’s chosen segmentation unit, which may follow tokens, document structure, semantic boundaries, or another rule. A provider’s retrieval spans are not evidence that every publisher should write to one fixed chunk length.
What do “consulted,” “used,” “cited,” and “recommended” mean?
A consulted source was retrieved or inspected within the scope a product exposes. A used source contributed information. A cited source has a visible link or attribution. A recommended entity is favored in the answer. These can overlap, but observing one event does not prove all the others.
Why is the approximately 41% figure marked “derived”?
seoClarity reported 56% of citations in the top 20. Within that subset,32% were in positions 1 to 3 and 41% in positions 4 to 10. Multiplying 56% by 73% gives 40.88%. The underlying reported values are rounded, so approximately 41% is more appropriate than implying a precisely measured 40.88% statistic.
Should the three studies be plotted as one time trend?
No. Ahrefs, seoClarity, and BrightEdge differ in samples, definitions, extraction, weighting, and observation dates. Their numbers can be compared with explicit qualifiers, but connecting them with a line implies a continuity the evidence does not establish.
What is the most useful reporting sentence?
“In this specified prompt panel, product, mode, locale, and date range, the target appeared in this proportion of scheduled runs, with this denominator, normalization method, and uncertainty.” It is less dramatic than a universal AI rank and much more useful for deciding what to do next.
The source library.
Platform documentation establishes disclosed behavior. Original studies establish bounded observations. Research papers explain mechanisms without revealing a vendor’s proprietary implementation.
How this guide was assembled
Based on the supplied September 2026 research and the corrected companion infographic. Core platform claims and plotted datasets were rechecked against primary sources on September 13, 2026. Historical findings retain their collection dates and definitions. No simulated numbers are presented as a commercial platform measurement.
Platform behavior, eligibility, and reporting
- Google: Optimizing for generative AI Search featuresQuery fan-out, Search foundations, eligibility and unsupported tactics.
- Google: AI features and your websiteWebsite requirements and behavior of AI Overviews and AI Mode.
- Google: Search ranking systemsSearch relevance, quality, passage understanding and other systems.
- Google: Search generative AI inclusion controlDefault inclusion, exclusions and inherited property settings.
- Google: Generative AI performance announcementJune 3, 2026 announcement with August 31 rollout update.
- Google: Generative AI performance reportImpression dimensions, aggregation and reporting scope.
- Google: Impressions, position and clicksWhy AIO element position is not a citation ranking.
- OpenAI: Searching the web with ChatGPTQuery rewriting, follow-up searches and contextual inputs.
- OpenAI: Web-search toolSearch workflows, consulted URLs and inline citations.
- Google: Gemini Search groundingSearch-call arguments and source annotations.
- Google: GenerateContent schemaGrounding metadata, webSearchQueries, chunks and supports.
- Microsoft: Copilot SearchUser-query results and additional searches.
- Bing: From ranking pages to supporting answersMay 6, 2026 engineering discussion of grounding objectives.
- Bing: AI Performance public previewFebruary 10, 2026. Citation counts are not ranking or placement.
- Perplexity: What is Pro Search?Multiple searches and synthesis across sources.
- Perplexity: AI-first Search API architectureHybrid retrieval, filtering, stronger rankers and subdocument scoring.
- Anthropic: Claude web-search toolRepeated search and version-scoped dynamic filtering.
Original quantitative studies
- Ahrefs: AI Overview citations and top-10 resultsMarch 2, 2026 publication. Keep the blue-link cut separate from mixed SERP blocks.
- seoClarity: AIO and organic ranking overlapOctober 12, 2025 data. Query, citation, and rank-conditioned denominators.
- Ahrefs: Assistant citations and the original-prompt top 10Early July 2025 data; August 11 publication. Five reference series.
- BrightEdge: AI Overviews at the one-year markFebruary 12, 2026 report. Plotted monthly values are February to July 2025.
- Grossman et al.: How Generative AI Disrupts SearchSIGIR 2026. Table 2 gives the plotted mean URL-set Jaccard comparisons.
- Kirsten et al.: Characterizing Web Search in the Age of Generative AIFindings of ACL 2026. Source behavior and answer variability.
Retrieval research and integrity
- Ammann et al.: Question Decomposition for RAGResearch pipeline combining subquestions, retrieval and reranking.
- Wang et al.: RichRAGMulti-faceted queries and collective value in document selection.
- Qu et al.: Uplift-RAGRetrieved relevance versus usefulness for generation.
- Ke et al.: Bridging the Preference GapRetriever and generator preferences need not align.
- Google Search spam policiesScaled content abuse, doorway abuse, stuffing and manipulation.
- Bing Webmaster GuidelinesSearch quality and abusive practices.
- Greshake et al.: Indirect prompt-injection threatsSecurity risk from instructions embedded in retrieved content.
- Pfrommer et al.: Manipulating product visibilityExperimental manipulation research, not a durable ranking recommendation.
- OpenAI crawler documentationSearch, training, and user-triggered retrieval roles.
- Anthropic crawler documentationSite-owner controls and crawler roles.
- Perplexity crawler documentationDocumented bots, access controls and verification guidance.
What this guide does not claim
No universal fan-out count. No universal candidate depth. No proprietary evidence weights. No inferred execution trace disguised as observed data. No guaranteed citation boost from a format, bot visit, or keyword. The strongest conclusions survive those limits.