The Visibility Fork — why AI answers cite a different web than search results

AI-search visibility has forked. A classic rank, an AI Overview citation, a ChatGPT search source, and a user-preferred publication are now different branches of discovery. Treat them as one metric and the dashboard lies. Treat them as separate branches and the work becomes measurable.

Why did one search result become several source pools?

For years, marketers used one mental model: a query produced a ranked page of results, and the job was to climb that page. That model still exists. It is just no longer the whole surface.

Google's generative AI optimization guide says its RAG process relies on core Search ranking systems to retrieve up-to-date pages from the Search index, and defines query fan-out as concurrent related queries generated by the model.[3] That means classic Search still matters. It is the reservoir. But the reservoir is no longer the same thing as the answer.

Google also says AI Overviews and AI Mode may use query fan-out, and that AI Mode and AI Overviews may use different models and techniques, so the responses and links they show can vary.[1] Read that as an architecture warning. The same underlying index can feed different answer surfaces, and those surfaces can show different links.

Buyer prompt

"Which issue tracker should an engineering team use if it ships every day?"

The old model assumed one ranked list. The new model asks which branch the engine enters, what sources each branch is allowed to read, and which source survives the branch-specific filter.

fan-outrankcitationpreference

Branch 01

Classic SERP

rank · click · impression

Ten links, snippets, local packs

source pool · Indexed pages ranked for the visible query

Branch 02

AI answer

citation · support link · claim fidelity

AI Overview, AI Mode, answer cards

source pool · Fan-out results plus supporting links

Branch 03

Assistant search

engine mention · cited URL · bot log

ChatGPT, Claude, Perplexity

source pool · Crawler policy plus provider search stack

Branch 04

Preference layer

preference prompt · repeat sample · segment

Saved sources, history, app context

source pool · Trusted domains plus user-specific context

Framework: Martech LLC · branches map to the Google, OpenAI, Anthropic, and Perplexity sources cited inline

This is the Visibility Fork. A buyer asks one question. Classic Search ranks pages. AI Overviews assemble support links. Assistant search decides whether to search, what providers or partners to consult, and what citations to show. Preference layers can bias which sources are highlighted for a user who has already selected a publication.

The commercial implication is blunt: the question "do we rank?" is no longer enough. The real question is "which branch did we enter, and what did that branch do with us?"

How different are AI citations from classic search results?

The strongest reason to take the fork seriously is not theory. It is measurement.

Grossman, Liu, Chen, Smith, Borcea, and Chen introduced an 11,500-query benchmark and found that Google AI Overviews appeared on 51.5% of representative real-user queries.[4] Their result is not just that AIOs are common. It is that the source pools diverge.

In the same study, the total AIO/SERP Jaccard similarity was 0.18 and the AIO/Gemini Jaccard similarity was 0.11.[5] In plain English: for many queries, the set of URLs shown by the AI answer is mostly not the same set as the classic results. The pages are not merely re-ordered. They are different candidate pools.

Another 2026 audit sharpened the point. Xu, Iqbal, and Montgomery issued 55,393 trending queries across 19 categories over March 13-April 21, 2026, finding 13.7% overall AIO activation and 64.7% activation for question-form queries.[7] They found that 29.8% of AIO-cited domains did not appear in the co-displayed first-page results.[8]

The user sees a conclusion, not a list of candidates.

That sentence from Xu, Iqbal, and Montgomery is the best description of the shift. A ranked list asks the user to choose. An AI answer chooses first, then exposes a few sources as evidence.

Representative real-user queries that generated AIOs

51.5%

Grossman et al. ORCAS query set

Average AIO/SERP shared URL set

0.18

Jaccard similarity; lower means a larger fork

AIO-cited domains absent from first-page results

29.8%

Xu, Iqbal, Montgomery longitudinal audit

AIO atomic claims unsupported by cited pages

11.0%

98,020 extracted claims

Sources: Grossman et al. 2026 · Xu, Iqbal & Montgomery 2026 · figures cited inline

This does not prove that rank is dead. It proves rank is not a complete measurement surface. A page can rank and not be cited. A domain can be cited and not appear on page one. A brand can be mentioned without a click. A source can be used by one branch and absent from another.

Which queries open the fork?

The fork is not equally wide on every query. It opens most often when the user asks a question that needs synthesis.

Google's own wording points in that direction: AI Mode is described as useful for exploration, reasoning, and complex comparisons. The 2026 measurements show the same shape. Question-form trending queries triggered AIOs far more often than non-question queries in Xu, Iqbal, and Montgomery's crawl; ELI5-style informational queries triggered AIOs far more often than retail keyword queries in Grossman et al.

Question-form trending queries

AIO activation in Xu et al.

64.7%

Non-question trending queries

same crawl window

9.5%

ELI5 benchmark queries

AIO activation in Grossman et al.

94.6%

Amazon retail keyword queries

same benchmark

17.4%

Sources: Xu, Iqbal & Montgomery 2026 · Grossman et al. 2026

This is why old keyword portfolios under-measure the shift. A head term like "issue tracker" may still behave like search. A buyer question like "which issue tracker should an engineering team use if it ships every day?" enters the answer branch. It carries constraints, comparison logic, and a reason to synthesize. That is where AI-search visibility becomes commercially meaningful.

For brands, the lesson is not to publish a separate page for every fan-out variation. Google says there are no additional technical requirements for AI Overviews or AI Mode, but supporting-link eligibility requires a page to be indexed and eligible for a Search snippet; Google also says no special AI text file or schema.org markup is required.[2] The work is to build pages that survive the branch: crawlable, useful, textually available, clearly authored, and specific enough to support answer spans.

Why does SEO still matter, but not enough?

The most common bad interpretation of AI search is that SEO has been replaced. It has not. The better interpretation is that SEO has become one branch in a larger retrieval system.

Google's public docs keep the foundation intact. A page still needs to be crawlable, indexable, eligible for snippets, and useful. Search systems still assess relevance, quality, and context. In the classic branch, ranking remains the clearest measurable outcome.

But the AI branch adds more gates. The page has to answer adjacent fan-out queries. The passage has to be self-contained enough to use. The claim has to be easy to verify. The source must not be blocked by the crawler the assistant uses. The answer may show one source even when many pages rank.

That is why C-SEO Bench is such a useful corrective: it found that most current conversational-SEO methods were largely ineffective and often negative, while traditional SEO strategies that improve the source's ranking in the LLM context were significantly more effective.[11] The answer is not gimmick markup. The answer is better evidence, better source eligibility, better page quality, and repeated measurement.

01

Eligibility

Can this branch fetch, index, and show the page?

Measure with

robots · snippet eligibility · crawler logs

02

Source pool

Does the page appear in the branch-specific candidate set?

Measure with

SERP rank · AIO support link · assistant citation

03

Claim fit

Can the answer use one sentence without unsupported glue?

Measure with

claim ledger · evidence anchors · freshness date

04

Preference

Does the user, locale, or source preference bias the branch?

Measure with

repeat samples · locale splits · preferred-source prompt

05

Outcome

Did the branch cite, mention, refer, or hand off action?

Measure with

confidence interval · receipt · conversion segment

Framework: Martech LLC · controls trace to official platform docs and measurement papers cited inline

The control panel is branch-specific. In classic search, you track rank and clicks. In AI Overviews, you track support links and cited claims. In assistant search, you verify crawler policy and cited URLs. In preference systems, you measure repeat visibility for users who have selected or trusted a source. In every branch, you need receipts.

Why will the fork get harder to measure?

Classic search had a powerful feedback loop. A user searched, clicked a page, stayed or returned, and those document-level signals could be studied. Generative search breaks that loop.

The NExT-Search paper argues that generative AI search has a feedback-loop disconnect because its pipeline spans query decomposition, retrieval, and answer generation while usually receiving only coarse feedback on the final answer.[12] The user may like or dislike the answer, but the publisher does not know whether the weak point was fan-out, retrieval, ranking, synthesis, citation, or UI.

Classic web search

01

query

02

ranked page

03

click

04

dwell / return

05

ranking update

The click lands on a document, so the feedback can be attached to a document.

Generative search

01

query decomposition

02

retrieval

03

rerank

04

answer synthesis

05

coarse answer feedback

The user judges the final answer, but the weak link may live in decomposition, retrieval, rerank, or synthesis. That is the measurement gap.

Source: Dai et al. 2025 NExT-Search · interpreted as a publisher measurement model

The assistant branch adds another split. OpenAI says ChatGPT search can choose to search the web based on the user's question, includes links to sources, and exposes a Sources sidebar with references.[13] OpenAI also says ChatGPT search uses third-party search providers and partner content, listing partners including Associated Press, Financial Times, Le Monde, News Corp, Reuters, The Atlantic, Time, and Vox Media.[14]

Crawler policy becomes a measurement input, not an infrastructure detail. OpenAI says OAI-SearchBot is used to surface websites in ChatGPT search results, and that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.[15] A brand can allow ordinary search crawling, block a search-specific assistant crawler, and then wonder why one branch disappears.

Preference adds the next layer. Google Preferred Sources can appear in AI Mode and AI Overviews, and eligibility is domain-level or subdomain-level rather than subdirectory-level.[16] This is not a universal ranking factor. It is a user-selected branch. But it marks where the surface is going: source visibility can depend on the person asking, not only the page answering.

The Visibility Fork, run on one buyer question

The point of the model is to make one query inspectable. Take a common B2B software question:

The query

Which issue tracker should an engineering team use if it ships every day?

Five branch checks

1. Branch 01· you cover

   Classic Search asks for a ranked page.

   tests · Is the page indexed, relevant, and trusted for the visible query?

   Operator lever

   Keep the canonical page crawlable, internally linked, fast, and clearly authored.

2. Branch 02· you cover

   AI Overview asks for supporting pages across fan-out queries.

   tests · Can the page support one answer span and adjacent subtopics?

   Operator lever

   Write self-contained claim atoms with dates, evidence, and source links.

3. Branch 03· you absent

   Assistant search asks whether its crawler may surface the page.

   tests · Do crawler rules allow the search bot, not only training bots?

   Operator lever

   Separate search-crawler access from training access and verify server logs.

4. Branch 04· you absent

   Preference systems ask whether the user already trusts a source.

   tests · Does an audience repeatedly select, cite, or prefer the publication?

   Operator lever

   Build direct audience routes and ask repeat users to mark preferred sources where supported.

5. Branch 05· you cover

   Measurement asks whether the branch changed beyond noise.

   tests · Was the brand cited across repeated prompts, engines, and dates?

   Operator lever

   Report intervals and receipts, not one prompt screenshot.

3/5steps covered

A page can rank and still fail the AI branch because the support sentence is not easy to verify, or because the assistant search crawler is blocked. The practical work is branch-by-branch: eligibility, support, access, preference, and repeated measurement.

Illustrative run: based on the branch mechanics cited in the article, not a measured score

The example is intentionally simple. It shows why one screenshot is not a measurement. You need the classic rank, the AI support link, the assistant citation, the crawler receipt, and the repeated sample. Without all five, you do not know whether the brand lost because it lacked authority, lacked passage fit, blocked the search bot, or merely fell inside normal answer variance.

What brands should do now

Start with a branch map, not an optimization list.

First, build a query panel around the questions buyers actually ask. Include broad head terms, comparison queries, "best for" questions, pricing questions, implementation questions, and objection questions. The point is not volume. The point is branch coverage.

Second, sample each query repeatedly across surfaces: Google classic Search, AI Overviews or AI Mode where available, ChatGPT search, Claude search, Perplexity, and any category-specific answer engine that matters to the market. Log the answer, cited URLs, brand mentions, position, date, locale, and whether the result changed beyond noise.

Third, inspect crawler access separately. Search visibility and training access are not the same policy. A robots rule that feels principled can remove the brand from a branch that uses a different user agent. The correct answer is not always "allow everything." The correct answer is to know which branch the rule controls.

Fourth, make every important page an evidence object. This does not mean special AI markup. It means visible authorship, clear dates, specific claims, external evidence, internal links, source-backed FAQ answers, and machine-readable schema that matches visible text. The older generative-search verifiability audit found that only 51.5% of generated sentences were fully supported by citations and only 74.5% of citations supported the associated sentence across four engines.[18] Verification is still fragile. Make it cheap.

Finally, treat AI visibility as an interval. Grossman et al. reported that generative search was less consistent across repeated runs and less robust to device, location, and query-syntax changes than traditional search.[6] A single prompt result is a weather report, not a climate model.

What the evidence does NOT prove

This evidence does not prove that classic ranking is obsolete. Google's own docs say AI features rely on core Search systems, and C-SEO Bench suggests the source's rank inside the LLM context still matters. The fork is additive, not a funeral.

It also does not prove that every assistant uses the same mechanics. Google AI Overviews, AI Mode, ChatGPT search, Claude search, Perplexity, and API search tools expose different crawlers, source pools, and citation interfaces. The Visibility Fork is an observed measurement model, not a claim that every vendor has the same internal architecture.

The papers are snapshots. They study defined crawl windows, query sets, devices, locales, and model versions. AIO behavior in March 2026 is not the same thing as every answer surface forever. The durable finding is the divergence pattern: source pools differ, answer activation varies by query form, and citation fidelity remains imperfect.

There is also a quality risk. Allaham and Diakopoulos audited 712 queries across ChatGPT, Copilot, Gemini, and Perplexity and found evidence that about 16% of cited sources were AI-generated.[17] More citation does not automatically mean better evidence. A serious GEO program should measure not only whether a brand is cited, but who else is cited, whether those sources are primary, and whether the answer is supportable.

The same caution applies inside Google AI Overviews: Xu, Iqbal, and Montgomery decomposed AIO responses into 98,020 atomic claims and found that 11.0% were unsupported by the cited pages.[9] Khosravi and Yoganarasimhan separately estimated that AI Overview exposure reduced daily traffic to exposed English Wikipedia articles by approximately 15% across 161,382 matched article-language pairs.[10] The fork changes visibility, evidence, and economics at once.

That is why the right posture is sober measurement. Rank the page. Probe the answer. Verify the citation. Read the source. Store the receipt. Repeat until you know whether movement cleared noise.

— Sundar Ramesh Kumar · martech.llc