Search Is Evolving from “Providing Answers” to “Completing Tasks”: What Comes After SEO?

In my previous Yunqi conference field report, I laid out the three stages of search evolution using this trajectory: Query → Results → Question → Answer → Goal → Plan → Search → Reason → Search Again → Tool → Action → Verify. After publishing, the three questions clients asked most frequently were: What does this trajectory mean for existing content assets? How should SEO and GEO be repositioned? And what engineering actions can we start implementing right away?

This article breaks down all three questions.

1. The Difference Between the Three Stages Lies Not in Capabilities, but in Deliverables

A quick recap: First-generation search gives you a list of links, and the user does the comparison work themselves. Third-generation search delivers a procurement recommendation report with supporting evidence, a booking draft, or a proactive strategy brief.

This distinction is better understood through concrete examples than abstract descriptions. Let’s explore three real-world work scenarios:

First, three cloud provider comparisons. First-generation search gives you links to three official websites and several review pages; second-generation search organizes public information into a summary, telling you the differences in performance, price, service, and ecosystem; third-generation search first asks about your business type, traffic, and compliance requirements, then calls the cloud provider’s public pricing API, scrapes the latest discount policies, compares them against your actual usage, and delivers a procurement proposal with supporting evidence. Tavily and EXA acknowledge in their public comparison (Exa vs Tavily) that Exa scores 81% on WebWalker multi-hop retrieval benchmarks versus Tavily’s 71%, with 1.4 seconds latency versus 4.5 seconds—a gap that directly determines whether an Agent can complete multiple rounds of supplemental searches within the few seconds a user waits.

Second, hotel booking. First-generation returns several booking platform links; second-generation recommends hotels based on your destination and dates; third-generation filters candidates based on group size, budget, whether you need meeting rooms, dining preferences, and accumulated loyalty points, then pulls three options, calls the hotel API for real-time pricing and room availability, calculates commute time to the client office via a maps API, and finally generates a booking draft with a calendar invitation.

Third, competitive monitoring. In the first generation, you’d search “competitor X’s pricing” and get some news articles; in the second generation, you’d get a summary of changes; the third generation continuously monitors competitors’ official websites, job postings, version releases, and user communities, proactively notifying you when significant changes occur and explaining “whether this 30% price cut was a response to your release last week, and what it means for your pricing strategy.”

The difference between “finding an answer” and “getting things done” is exactly this.

2. Research Loops Matter More Than First Recall

Traditional search systems place great emphasis on Recall, Precision, and Ranking. Agentic Search still needs these capabilities, but the evaluation criteria need to evolve.

A good Research Agent shouldn’t search only once. It should generate a second round of queries when evidence is insufficient; seek third-party evidence when two sources conflict; call another API when real-time pricing is needed; and reorganize the search plan when it discovers that the user actually wants to compare total cost of ownership (TCO).

Learn AI Slowly #025

In S1-DeepResearch’s 2026 paper survey, standard LLMs without multi-turn retrieval achieved less than 10% accuracy on multi-step research benchmarks. Deep Research Agents—systems that iterate through search-read-synthesize-recheck loops—reached over 50%. That fivefold gap isn’t about mastering a single technique; it’s about the power of a closed research loop.

So what makes a closed loop? One approach spans eight stages: Goal → Plan → Search → Reason → Search Again → Tool → Action → Verify. Another compresses to five: Plan → Search → Reason → Refine → Conclude. I use both depending on the task—the eight-stage version for complex, long-horizon work like enterprise research or cross-domain due diligence, the five-stage for everyday research cycles. The catch is that any error in the middle compounds on everything that follows.

This also means the search infrastructure and the upper-layer Agent should be understood separately. Tavily, EXA, Elasticsearch, OpenSearch, Brave Search, and Browser Search can all serve as Retrieval Providers. The actual product layer is responsible for Intent, Planning, Source Strategy, Reasoning, Evidence evaluation, Evaluation, and Action execution.

In our hands-on experience, we’ve encountered a typical anti-pattern: a financial institution interpreted “search augmentation” as “switching to a smarter search engine.” As a result, the Agent got an apparently authoritative document with outdated policy in the first round of recall, failed to trigger a follow-up search, and ultimately cited a risk control rule from two years ago that had already been abolished. The system appeared to be running smoothly, and the decision-makers didn’t catch the mistake. It wasn’t until an audit surfaced the citation source that the entire Research Loop was found lacking three critical layers of judgment: Evidence Age, Conflict Resolution, and Source Authority.

Extending search quality assessment from “are the results relevant” to “can we trust the entire research cycle” — that’s the point I’ve driven home most consistently with clients over the past year.

3. Memory Determines Whether an Agent Can Run the Distance

At the Agentic Search forum, Memory was singled out for dedicated discussion, covering Long-term Memory, Task Memory, and Context Compression.

That makes perfect sense. Once a task evolves from a simple question-and-answer exchange into a multi-step research process spanning dozens of steps, the system immediately faces a challenge: what has already been searched, verified, or debunked, and can we reliably hold onto it?

MemGPT introduced the “LLM as OS” paradigm in 2023, with the core idea being layered memory architecture: core memory sitting within the context window, a storage layer outside the conversation, and retrievable archival memory. Letta engineered this approach in 2024, and DeepLearning.AI turned it into a short course in 2026. There’s a solid lineage of prior work to cite here — this isn’t terminology we invented ourselves.

When a research task keeps re-searching the same information, it wastes substantial costs. What’s more dangerous is that the system forgets it previously found a source unreliable, only to cite that same source again later.

Therefore, Task Memory should be recorded in a structured way. When each research task concludes, the system should persist the following fields:

Query Plan: the sub‑questions into which this task was broken down, and the objective of each sub‑question.
Source List: the sources retrieved for each sub‑question, together with each source’s authority level, publication date, and deprecation status.
Evidence Snapshot: the key facts extracted from each source, with the original source citation and the verbatim excerpt.
Conflict Tags: any inconsistencies among sources, what the disagreement concerns, which source the system chose, and why.
Interim Conclusions: the system’s provisional judgment on the current question after each round of retrieval.
Failure Paths: the retrieval routes that yielded no results, why they failed, and whether they are worth retrying next time.
Final Verdict: whether the user accepted or rejected this conclusion, and the reasoning behind that decision.

This way, when a similar task arises, the system can directly reuse the structured experience instead of starting from scratch.

Many teams in the past used vector databases (Embedding + Retrieval, i.e., converting text into vectors and searching by similarity) for Memory, storing all chat logs as Embeddings and retrieving relevant snippets on the next turn. This approach carries two hidden risks: First, “relevant snippets” may not actually be relevant, increasing the likelihood that the model gets distracted by noise; second, large numbers of snippets lack structured tags, making it impossible to manage versions, sources, or expiration dates. What Memory truly needs is designed precipitation, evaluation, and structuring—otherwise, as Context accumulates, the next decision becomes even less certain.

Four, the truly valuable part of self-evolution: precipitating successful experiences into Skills

The session also showcased Agent Swarm (multi-agent collaboration) and a self-evolution闭环.

This framing easily evokes images of models training themselves. A more accurate understanding is that after an Agent completes a task, it precipitates valuable methods into Memory, Knowledge, and Skill through evaluation, human feedback, and result validation.

For instance, after conducting 20 consecutive product opportunity researches, a system can gradually form Product Opportunity Research Skill. This Skill is not merely a prompt; it should specify:

Trigger: What types of tasks will invoke this Skill
Goal: What defines success for this research
Context Schema: What enterprise context needs to be pre-loaded (brand, category, target market)
Constraints: Which sources are unreliable, which data cannot be cited
Tools: Which search sources, APIs, and databases to call in sequence
Workflow: Step sequence (first identify market entry points → then locate main competitors → then review pricing, traffic, and user reviews → then dig into Reddit and community complaints)
Source Priority: Authoritative sources first (financial reports/annual reports), community second, blogs last
Verification: What evidence is sufficient to support “demand exists”
Output Schema: What fields will be in the final output, and what format should those fields follow

Here we need to distinguish for the reader: the Skill fields above are not OpenAI Function Calling (which registers external functions as JSON Schema interfaces that the model can invoke) — that belongs to the tool‑layer protocol; nor are they Anthropic Tool Use (similar to Function Calling, but Anthropic uses the finer‑grained tool_use/tool_result message types) — again, it stays at the tool layer. AutoGen’s Agent Spec standardizes an agent’s capabilities into a serializable object, which only partially overlaps with the Skill here. The real point of a Skill is “the workflow template that a team has refined through dozens of real‑world engagements”, bundling the four aspects of Tools, Workflow, Constraints, and Verification — something none of the OpenAI/Anthropic/AutoGen protocol layers cover.

The 21st time you conduct a similar survey, you no longer need to reinvent the methodology. That is the true compounding value of Skill.

慢慢学AI<039> — How Skills and Evaluation Drive Self-Evolution in AI Systems

When I was coaching an AI transformation for a retail chain brand, I saw this process unfold with my own eyes. In the first week, as the Agent ran competitive research, every product manager had to re-instruct it on search paths, source prioritization, and judgment criteria. By week three, we had distilled each round of effective research paths into a Skill, while explicitly flagging the missteps—such as over-reliance on a single data source. By week eight, a new product manager could simply feed objectives into the system, and the resulting research reports consistently outperformed the initial drafts produced by our most senior team member back in week two. Note: The three-phase evolution described here is illustrative; actual timelines vary by project.

This is the real value of Skills: they transform the tacit methods scattered across individual brains within a team into shared, inheritable, and improvable engineering assets.

True self-evolution therefore hinges on Evaluation. Without explicit outcome assessment, erroneous experiences get preserved too, and the system only grows more confident in repeating its mistakes.

SEO Hasn’t Disappeared—But the Funnel Has Grown a Step

The classic SEO funnel used to be Ranking → Impression → Click → Signup → Paid. With generative search, users can now get answers directly within search results, ChatGPT, Perplexity, or other Agents—without clicking through to individual sources.

In March 2025, Pew Research Center tracked 68,879 real Google searches from 900 American adults and found that when AI summaries appeared in search results, traditional links received clicks only 8% of the time—compared to 15% when no AI summary was present. Nearly half of all clicks essentially got squeezed out.

This shifts content’s value proposition from “driving traffic” to “becoming the kind of evidence AI models are willing to cite and rely on.” Operations metrics will gradually incorporate a new funnel: AI Visibility → Citation / Mention → AI Referral → Qualified → Signup → Paid.

This funnel operates across four levels: being seen, being cited, being clicked, and converting. AI Visibility forms the foundation—whether your brand or product appears in AI answers. Citation/Mention tracks how many times you’re referenced, in what context, and whether the reference is accurate. AI Referral measures whether users follow those references to your site and how much high-quality traffic they drive. Qualified tracks how many of those visitors go on to register, start a trial, or make an inquiry. Ultimately, it all flows back to Signup and Paid conversions.

The baseline figures from GEO industry research are: 97% citation rate on Perplexity, 34% on Google AI Overviews, and 16% on ChatGPT. AI-driven traffic currently accounts for about 1.08% of total website traffic, but Perplexity referrals convert at 3.1 to 4.4 times the rate of Google organic search, with session durations 4.7 times longer. Both of these figures represent order-of-magnitude differences—exact numbers vary by site and industry, but the trend is consistent: AI traffic is small in volume but high in quality.

The biggest difference from the traditional SEO funnel is that it adds two new stages in the middle—Citation/Mention and AI Referral—and that citation quality matters far more than citation quantity. When an AI answer cites your product as “one alternative option” versus “one of the three products most worth evaluating for this use case,” the Qualified traffic you get is worlds apart. Ahrefs’ data scientists found that Google AIO sources change by about 45% per round—which means simply “being cited once” isn’t enough; what matters is Citation Persistence.

Here’s a misconception to avoid. Citations themselves can also become a Vanity Metric. If a brand gets mentioned across a large number of AI answers but generates no high-quality visits, branded searches, sign-ups, or revenue, raw citation frequency has limited commercial value.

So GEO still has to ultimately tie back to the full Funnel. But the measurement criteria in the middle have shifted: it’s no longer about “being found through search” but about “being trusted by AI enough to be quoted.”

Six. Content Building Starts Establishing Reliable Evidence for the Problem Space, Not Just Writing for Keywords

Traditional SEO easily builds a content matrix around keywords—one page per keyword, with the goal of ranking, capturing long-tail traffic, and earning backlinks.

With Agentic Search, content still needs to cover search intent, but the structure becomes closer to the problem space. An agent conducting research may consecutively pose multiple sub-questions and compare multiple sources. It demands information that is clear, verifiable, structurally consistent, and sourced unambiguously.

This means content building requires establishing stability across five dimensions beyond just keywords: clear Entity, verifiable Fact, dated Date annotations, traceable Source, and complete Topical Coverage. In other words, the old approach of piling low-density pages to chase rankings will simply break down once agents arrive.

Agents prioritize pages with clear entities, verifiable facts, explicit sources, and complete topics during research—not pages that stuff the title with the same keyword three times. The patterns from GEO-Bench (the benchmark used in FeatGEO, a 2026 ACL paper) show that document-level content attributes—structure, substance, and language—have far greater impact on citation rates than scattered keyword tweaks. Adding source citations to every factual claim, attaching time windows to every number, and explicitly linking thematic relationships between pages—these practices that traditional SEO overlooked have become core moves in the GEO era.

A site that consistently produces AI-trustworthy content in a given domain becomes more valuable in both search and the Agent era.

I validated this pattern while working on content strategy with several B2B manufacturing clients: one industrial automation client reorganized three years of product manuals, industry papers, and white papers across four dimensions—entity, fact, date, and source—while adding site-wide topic maps. After six months, their product pages saw citation rates in mainstream AI answers increase nearly threefold, with Qualified Leads rising by roughly 40%. Note: These figures are directional indicators from a confidential project review; specific multiples vary by industry, baseline, and execution depth, and are not publicly reproducible precise data.

Seven, Search APIs Are Increasingly Looking Like the Infrastructure Layer for Agents

For developers, this shift carries another engineering implication.

Going forward, “search” should no longer be built in isolation as a standalone product capability. A more sensible architecture divides into three layers:

Layer One: Search Infrastructure

This layer serves as a provider-agnostic retrieval foundation that manages keys, quotas, costs, stability, caching, routing, and fallback strategies across different providers (Tavily, EXA, Brave, Google, Bing, OpenSearch, Elasticsearch, and self-built retrieval systems). Its interface exposes a unified “retrieve and return evidence” contract—not a specific provider’s SDK.

This layer sits parallel to the five-tier architecture of classic RAG (Retrieval-Augmented Generation)—Document Store, Retriever, Generator, Reranker, and Prompting Strategy—but the alignment isn’t exact. RAG follows a “question-and-answer in one shot” paradigm, whereas Search Infrastructure operates as “multiple agent calls with dynamic assembly.” Mixing the two concepts often leads to confusion around Reranker and Source Priority.

Tier 2: Research Agent. This tier handles Planning, Query Expansion, Retrieval, Reasoning, Evidence Assessment, Verification, and Action Orchestration. Rather than calling Providers directly, it obtains candidate evidence through the unified abstraction from Tier 1, then determines which evidence is trustworthy and which conflicts require additional searching to resolve.

Layer 3: Domain Skill. For different tasks like SEO Research, Competitor Research, Academic Research, Product Research, and Legal Research, this layer accumulates specialized methods, source prioritization rules, validation criteria, and output schemas. Skill invokes Agent, and Agent invokes Infrastructure.

This layered architecture offers an asymmetric benefit: the underlying Provider can be swapped out—if Tavily encounters issues, you can temporarily switch to EXA, and the business side won’t need to rewrite the Research Loop just because the underlying provider changed. Meanwhile, the upper-layer capabilities keep accumulating, so the business side doesn’t need to rewrite Skills when switching Providers. This is the true value of layering—decoupling.

A financial services firm’s AI team experimented with this architecture last year, restructuring their search capabilities from “engineering teams each integrating different APIs independently” into these three layers. The result: engineering teams were no longer derailed by “a sudden rate limit from a Provider,” and business teams could directly describe what research they needed using Skill definitions without worrying about which underlying service was being called. Within three months, cross-team research task delivery time was cut by roughly half. Note: The “cut in half” figure comes from a desensitized directional review of the project; actual multiples vary based on team size and pre-existing structure.

VIII. The Real Change: “Information Acquisition” Is Integrated into Task Completion

If you only think of AI Search as “the search box got smarter,” you’re underestimating the scale of this shift.

Once Search, Memory, Tool Use, and Action converge, what users actually request will increasingly resemble high-level objectives, with Queries retreating into the system’s internals.

“Find me some materials on this topic” will become “Research this market for me, compare several options, and give me evidence-backed recommendations.” “Help me search for hotels” will become “Find me suitable hotels given these constraints, compare total costs, and get everything ready for booking.” “Check out our competitors” will become “Continuously monitor our competitors, alert me when something significant changes, and explain whether it affects our current strategy.”

This shift carries different implications across four industries.

In telecom, the Agent moves beyond answering “Which 5G plan is cheaper?” Instead, it analyzes the user’s calling patterns, data usage, and roaming habits, evaluates whether their current plan is cost-effective, and proactively presents three options—renew, switch plans, or switch carriers—before the contract expires, sending the comparison results directly to the user.

In financial services, the Agent moves beyond explaining “What is an ETF?” With the user’s authorization, it reviews their asset allocation, market conditions, and regulatory policy changes, then proactively recommends portfolio adjustments, complete with evidence sources for each suggestion.

In manufacturing scenarios, an Agent is no longer limited to “look up the fault code for this machine.” Instead, it cross-examines equipment logs, sensor data, and recent maintenance records to diagnose the root cause, then presents three-tier maintenance plans—“today / tomorrow / this weekend”—while automatically generating work orders with spare parts lists and time estimates. This category of use case is the most compelling in 2026 industrial Agent deployments: vibration data from a CNC machine combined with three months of maintenance records and real-time sensor alerts—where manual troubleshooting would take 4 hours, a multi-round Agentic retrieval with evidence chains delivers three actionable plans in 8 minutes, each backed by cited sources.

In e-commerce scenarios, an Agent is no longer limited to “look up competitor pricing.” Instead, it continuously monitors pricing, inventory, and promotional rhythms across the entire platform, proactively pushing alerts when a SKU of user interest drops “20% below its 30-day historical average,” along with recommendations on whether to adjust your own pricing strategy.

Search still exists—it has simply retreated to a larger task orchestration system.

For content teams, SEO practitioners, and AI product builders, this shift is what truly matters: whoever can continuously produce trustworthy evidence, whoever can turn retrieval into citations, and whoever can convert citations into actionable outcomes.

Implications for Decision-Makers

If you’re a C-suite executive or VP responsible for digital transformation, the real question for Agentic Search deployment isn’t “which search API should we switch to?”—it’s about locking down three things:

Learn AI Slowly <001>

The Three Things That Actually Determine Whether Your AI Can Take Off

Evidence Infrastructure — Is your product documentation, industry reports, white papers, and customer service records structured along the four dimensions of “entity—fact—date—source”? This is the prerequisite for whether an Agent can cite your content over the long term, not something an SEO tool can simply swap out.
Skill Assets — Do the tacit experiences of your most seasoned employees—those mental playbooks for “when X happens, do Y”—exist as reusable, improvable Skills? If not, every AI initiative starts from scratch.
Evaluation Loop — What do you use to determine whether AI outputs are right or wrong? Self-improvement without Evaluation just means getting increasingly confident about repeating mistakes.

What these three things have in common: none of them appear on procurement lists. They’re all inside the organization.

You Might Be Wondering

Q1: With Agentic Search emerging, do we still need traditional SEO?

No. SEO is the foundation of GEO. If a website can’t rank in traditional search, its chances of being cited by AI are even lower. SEO handles “being findable.” GEO handles “being trusted enough by AI to be quoted.” These are two different things with a complementary relationship, not a substitution.

Q2: Citations are up, so why haven’t Qualified Leads increased?

It’s likely a citation quality issue. Being “just another option” versus “one of the top three worth evaluating” in an AI-generated answer yields entirely different traffic. The former is mere browsing; the latter is genuine inquiries. Examining where and in what context your citations appear in AI answers is far more valuable than looking at raw numbers alone.

Q3: Should you build a Research Agent now?

First, define the problem scope. If your research tasks require proprietary data (customer profiles, internal product manuals, compliance records) and involve compliance reconciliation or regulatory interpretation, building your own is necessary. If research tasks primarily involve publicly available information, start by making the most of existing tools (Tavily / EXA / Perplexity / Alibaba Cloud OpenSearch Agentic Search, etc.) and observe for 3-6 months before deciding whether to invest in custom development.

Reverse Audit (Avoiding Self-Deception)

Treating “being cited by AI” as a KPI rather than a means—if the citations don’t funnel back to registrations or inquiries, it’s a vanity metric.
Treating Skill accumulation as a one-time documentation effort—if a Skill has no Evaluation, you’ll only lock in mistakes.
Treating the search API as an engineering problem—assessing search quality is fundamentally about the credibility of the research loop, not just choosing an interface.
Treating “how much AI did” as evidence—what really matters is “what the system changed as a result.”

If you’re evaluating how your enterprise search or SEO team can embrace Agentic Search, which GEO metrics are worth tracking, and which content assets are likely to become targets for AI citations, let’s talk. We specialize in enterprise AI transformation consulting—from search architecture and content strategy to GEO metrics, helping you turn “pages getting discovered” into “being trusted by AI enough to be quoted.”

Enterprise in-house training—we deliver search architecture, content assets, and GEO measurement as hands-on workshops your team can apply (2–3 days, fundamentals +实战).
Targeted consulting—we diagnose your current search infrastructure, Skill accumulation roadmap, and Citation quality, then map out a practical execution path.
Executive briefings and industry talks—we bring frameworks like the three stages of search, content as a problem space, and the GEO funnel to your industry conferences or leadership meetings.

Contact: [email protected]

Further Reading: AI Transformation: A Seven-Step Framework, a systematic guide to the complete path for enterprise AI adoption.

About This Series

“Cloud Trek Observations” is an industry field series from IAIUSE, launched from the 2026 Cloud Trek Conference. We take a researcher’s lens to deconstruct the real changes unfolding in the AI industry—not chasing headlines, but examining the directions being bet on and the strength of the evidence behind them.

The series covers topics including the system layer above foundation models, agent deployment, context assets, enterprise AI organizational design, and the shift in AI product competitive units, totaling approximately 10 articles.

I bring nearly eight years of experience in large enterprise consulting and business analysis, having worked at IBM on projects across telecommunications, finance, insurance, and manufacturing. I then continued working on the front lines—at an operator’s product division, in internet products, and in AI application development—handling requirements analysis, product design, and cross-team implementation. This publication is actually run by a small team: myself and one to two long-term collaborators, handling AI coding tool research, organizational governance case studies, and coaching conversations respectively. Most of the projects mentioned as “helping enterprises navigate through” were delivered collaboratively by our team.

The insights in this series come from my on-the-ground observations and cross-industry validation, reflecting a clear authorial perspective and not representing the views of any vendor.

Localization Key Points (Multi-Language Translation Reference, IAIUSE Multi-Language Strategy · 2026-08-09 Guidelines)

When translating into 19 languages, replace the following content with target-language market localization while keeping structure and visual elements unchanged:

中文稿内容	英文版	日文版	德文版	阿拉伯版
阿里云 OpenSearch	Alibaba Cloud OpenSearch(保留)	アリババクラウド OpenSearch	Alibaba Cloud OpenSearch	OpenSearch علي بابا كلاود
Tavily / EXA	Tavily / EXA(全球性产品保留)	Tavily / EXA	Tavily / EXA	Tavily / EXA
ChatGPT / Perplexity	ChatGPT / Perplexity	ChatGPT / Perplexity	ChatGPT / Perplexity	ChatGPT / Perplexity
百度 / Google	Google	Yahoo! JAPAN / Google	Google	Google

China Telecom / China Mobile / China Unicom	AT&T / Verizon / T-Mobile	NTT / KDDI / SoftBank	Deutsche Telekom / Vodafone	STC / Etisalat
飞书 / 钉钉	Slack / Teams	Slack / Teams / Lark	Slack / Teams	Microsoft Teams
Tavily / Exa case localization	Tavily / Exa (original case)	Tavily / Exa (original case)	Tavily / Exa (original case)	Tavily / Exa (original case)
China Merchants Bank / ICBC	JPMorgan Chase / Bank of America	Mitsubishi UFJ / Sumitomo Mitsui	Deutsche Bank / Commerzbank	National Commercial Bank (Saudi) / QNB

Retail chain brand (anonymized)	Target / Best Buy (anonymized)	Aeon / Seven & i (anonymized)	Lidl / Aldi (anonymized)	Panda / Al Othaim (anonymized)
Industrial automation client (anonymized)	Honeywell / GE (anonymized)	Fanuc / Yaskawa (anonymized)	Siemens / Bosch (anonymized)	SABIC / Aramco (anonymized)
Financial client (anonymized)	JPMorgan / Goldman (anonymized)	Mitsubishi UFJ / SMBC (anonymized)	Deutsche Bank (anonymized)	NCB / QNB (anonymized)

Note: Apart from the above localization items, the global products/concepts (Research Agent, Task Memory, Context Provider, Citation / Mention, AI Visibility, Long-term Memory, MemGPT, Letta, RAG, Tavily, EXA) remain in their original form and are not translated. The remaining 15 languages follow the IAIUSE three‑tier approach: the five priority languages (Chinese/English/German/Japanese/Arabic) are localized per the table above; the nine secondary languages (Spanish/French/Portuguese/Korean/Russian/Italian/Dutch/Polish/Turkish) retain the original OpenSearch / Tavily / EXA names and replace the local representative enterprises; the five optional languages (Swedish/Thai/Vietnamese/Ukrainian/Indonesian) keep the original names as placeholders.

Citation notes (itemized, 2026-08-09 agreement · Gate 1 must be checked)

#	Quote in Text	Source	Publication Date	Evidence Level	Stance Label
1	“Exa 81% / Tavily 71% on WebWalker multi-hop benchmark; Exa 1.4s / Tavily 4.5s p95 latency”	exa.ai/versus/tavily (Exa Labs official comparison page)	2026-02-12	Verified fact (vendor-hosted)	Exa’s own page, carries Exa’s stance; WebWalker third-party benchmark can be independently verified
2	“Alibaba Cloud OpenSearch Agentic Search commercialized starting 2026-08-31, previously free during public beta”	alibabacloud.com/help/doc-detail/3053142.html (Alibaba Cloud OpenSearch official documentation)	Effective 2026-08-31	Verified fact (official documentation)	Alibaba Cloud, vendor stance

| 3 | “Standard LLMs achieve less than 10% accuracy on multi-turn retrieval tasks, while Deep Research Agents exceed 50%” | tianpan.co/blog/2026/04/12/deep-research-agents… (Industry Analysis); See also arxiv.org/html/2606.15367v1 (S1-DeepResearch Survey) | 2026-04 | Industry Watch (Analyst Survey) | Tianpan, independent analyst; peer-reviewed by S1-DeepResearch paper |
| 4 | “MindDR achieves 45.7% on BrowseComp-ZH / 52.5% on DeepResearch Bench” | arxiv.org/html/2604.14518v1 (Mind Auto Mind DeepResearch Technical Report) | 2026-04-14 | Verified Facts (Paper) | Mind Auto in-house model, vendor perspective |

| 5 | “DRBench: 100 enterprise deep research tasks, 1093 sub-questions, 10 domains” | arxiv.org/pdf/2510.00172(ServiceNow Research) | 2025-10 | Verified facts (paper) | ServiceNow’s own research |
| 6 | “MemGPT 2023 paper ‘LLM as OS’ layered paradigm; Letta 2024 engineering; DeepLearning.AI 2026 short course” | blog.stackademic.com/letta-platform… ;letta.com/blog/benchmarking-ai-agent-memory;linkedin.com/posts/deeplearningai… | 2023-2026 | Verified facts (technical review) | MemGPT/Letta’s own blogs, with tool-side perspective |

| 7 | “Pew Research: 900 US adults, 68,879 Google searches; 8% click rate with AI summaries, 15% without AI summaries” | instituteforpr.org/do-ai-summaries-reduce-clicks-on-google (Pew Research overview) | 2025-07 | Verified facts (independent research) | Pew Research independent institution |

| 8 | “Perplexity citation rate: 97%; Google AIO: 34%; ChatGPT: 16%; AI-driven traffic accounts for ~1.08% of total website traffic, with conversion rates 3.1–4.4× higher than Google organic search and session durations 4.7× longer” | cite.solutions/generative-engine-optimization(2026-05-02);omnius.so/blog/generative-engine-optimization-kpis-and-metrics(2026-08-18);trycited.app/generative-engine-optimization(2026-08-17) | 2026-05/08 | Industry observation (multiple GEO tool vendors); MarGen 2026 citations | Data from GEO tool vendors, with tool-side bias |

| 12 | “RAG Architecture Deep Dive: Document Store / Retriever / Generator / Reranker / Prompting Strategy” | medium.com/@angelosorte1/rag-architectures-every-ai-developer-must-know-in-2026(Angelo Sorte 综述);levelop.dev/blog/…/agent-rag-architecture-five-layer-retrieval-stack(2026-07-23);braintrust.dev/articles/best-vector-databases-for-rag-2026 | 2026 | Industry Watch (Technical Survey) | Engineering Practice Survey |

| 13 | “Anthropic Agent Skills: filesystem-level resources, on-demand Skill loading, composability” | docs.anthropic.com/en/docs/agents-and-tools/agent-skills/overview (Anthropic official documentation) | 2026 | Verified facts (official documentation) | Anthropic, vendor position |

Customer cases marked as “de-identified/illustrative” in the text (chain retail brand evolution, B2B industrial automation customer with 3× citation rate / 40% increase in qualified leads, financial customer with delivery time reduced by half): derived from sanitized project retrospectives, not referring to any specific customer, with figures provided as directional indicators only.