Subtitle:
Multi-round verification, time slicing, safety-first gating, and multimodal evidence are entering the picture—the changes across the three platforms are rewriting the underlying rules of GEO.
Introduction
Over the past year, many people have reduced GEO to one sentence: Make it easier for AI to find me, cite me, and recommend me.
But when the recent changes from DeepSeek, Kimi, and Doubao are viewed together, the problem is clearly no longer as simple as “being crawled.”
On the surface, DeepSeek’s changes look like more citations, more diverse sources, and longer answers; in essence, the answer mechanism is shifting from a single-round process of “search once, answer once” to an agentic search workflow that “decomposes questions, splits directions, searches in parallel, and verifies centrally.” On Moonshot’s official side, web search, search/crawl, multi-step tool use, and the agent cluster collaboration of K2.5 have been written into its products and documentation; this means Kimi does not lack the ability to access the web, and the front-end observation of “stopping in 2025” is more like the platform becoming more cautious about “what to trust first.” As for Doubao, the higher refusal rate, stronger risk avoidance, and more obvious commercialization front-loading on consumer, marketing, and recommendation questions after 315 are not isolated events either, but are happening together with AI-generated advertising, citation advertising, platform responsibility, and data-poisoning governance.
First, let’s set a boundary. Some of the following content is front-end observable phenomena, some is secondary industry interpretation, and not all of it equals official platform announcements. Statements such as “a single answer’s citations expand to about 50,” “authoritative sources carry 6 times the weight of self-media,” “DeepSeek launched ‘source backtracking’ on a certain day,” “Kimi adopts ‘time slicing’,” and “Doubao has a 60% refusal rate” are better treated as industry samples and product observations rather than official positions. What can truly be publicly verified is these three points: first, DeepSeek officially continues to strengthen web search, multi-step reasoning for complex search, and thinking in tool-use; second, Moonshot officially documents web search, official tools, multi-step tool calling, and agent collaboration in its product documentation; third, the 315 exposé and the 2026 advertising regulation documents have indeed pushed “AI poisoning,” “AI-generated advertising,” “citation advertising,” and “platform compliance responsibility” into the foreground.
For this reason, the changes across the three platforms cannot be understood merely as product fluctuations. They collectively point to this: AI search platforms are moving from “answer as much as possible” to “answer trustworthily first,” and from “broad indexing” to a composite system of “evidence orchestration + source governance + risk control + commercial distribution.”
1. DeepSeek: It Raised the “Evidence Gate”
If anyone has most clearly exposed the upgrade of AI search’s underlying mechanisms in this round of change, it is DeepSeek.
According to official public information, when DeepSeek launched web search on the web at the end of 2024, it already stated clearly: for complex questions, the model will automatically extract multiple keywords and search them in parallel; the 2025 V3.1 release notes directly emphasized stronger multi-step reasoning for complex search tasks and stronger agent capabilities; by V3.2, the official documentation further integrated thinking directly into tool-use. More importantly, DeepSeek’s publicly available Web/App search prompt explicitly requires: answers should synthesize multiple relevant pages, citations should be placed near the corresponding sentences rather than piled up at the end, and the same page should not be cited repeatedly. In other words, DeepSeek’s web-grounded answers are clearly no longer single-pass / single-query linear retrieval generation, but closer to a kind of layered agentic search orchestration.
This also corresponds to that judgment in your social feed:
It is not simply “search once, answer once,” but more like first performing query understanding / intent decomposition, then planner / router distributes the subtasks to parallel retrieval workers / search sub-agents, and finally the orchestrator / synthesizer / verifier handles deduplication, conflict resolution, cross-verification, and citation orchestration.
Put more plainly: It used to be like one person looking things up; now it is more like a small team working in parallel.
This has a very direct impact on GEO. In the past, what was optimized was a page under a keyword, competing for “position”; now the competition is: when the model breaks a question into multiple subtasks—definition, comparison, background, cases, price, pros and cons, risk boundaries, and so on—can your content enter more subqueries, more evidence chains, and more verification nodes? The GEO paper itself points out that generative search does not simply return a list of links, but synthesizes multiple sources into a single answer; source visibility depends on whether it is selected, displayed, and cited by the model, not just on ranking. The paper also shows that an appropriate GEO strategy can increase content visibility by up to 40%. By 2026, IF-GEO further framed this problem as a multi-query, conflict-aware optimization problem: not optimizing a single query, but maintaining stable visibility across multiple heterogeneous queries.
This means what DeepSeek’s change truly brings is not the shallow layer of “more citations,” but a rewrite of GEO’s competitive unit from “single page” to “evidence chain.”
“GEO is no longer a content-volume game played alongside search engines, but infrastructure work that provides agents with verifiable evidence assets.” A more rigorous phrasing would be: GEO’s optimization target is migrating from single-point hits in single-turn Q&A to evidence coverage, citation eligibility, and cross-verification pass rates in multi-subtask retrieval.
2. Kimi: It Tightened the “Trust Time Gate”
The second change comes from Kimi.
The piece you shared, “When Kimi Stopped in 2025,” is most valuable not for “which year it stopped at,” but for accurately grasping a deeper problem: platforms are beginning to redefine “what information deserves to enter the cognitive layer.”
Because according to Moonshot’s official documentation, Kimi does not have a “cannot access the web” problem. The official tool documentation explicitly lists web-search as an official tool and states that its role is to call the latest information, cite authoritative sources, and make results verifiable; the official web search example demonstrates a search/crawl tool-calling flow; after 2025, Moonshot continued to document kimi-k2-thinking and kimi-k2.5’s multi-step tool calling, deep reasoning, and agent capabilities in its documentation; in 2026 it even described an “agent cluster collaboration model” in the K2.5 notes. This shows Kimi’s issue is not “a lack of search and tool capabilities,” but rather more conservative adoption and refresh of the latest content on the front end.
Why would the platform do this? It becomes easier to understand when the post-315 data-poisoning context is brought in. Xinhua News Agency’s client directly documented the GEO gray industry’s language of “paying for placements and feeding content, poisoning, and influencing model citations” in its 315 exposé; Outlook (Liaowang) subsequently stated clearly that the abuse of GEO techniques is pushing mass pseudo-original advertorials, low-quality content, and false information into AI’s retrieval and citation chains. Meanwhile, Anthropic’s research shows that in its setup, 250 carefully constructed poisoned documents are enough to stably implant backdoors across models of different sizes. In other words, if platforms do not raise the bar for the newest, bulk-generated, low-credibility, hard-to-trace information, they will remain exposed to the risk that “even a small number of samples can cause contamination.”
So a more reasonable explanation for front-end phenomena like “Kimi stopped in 2025” is not product regression, but that the platform is doing some kind of trusted time-window management: it does not first ask “is this the latest,” but “is this trustworthy enough to include now.”
This is also the real meaning of “brand GEO should return to mind management.” For brands, competition in the AI search era is not just about publishing information faster, but about getting the platform to form, in its long-term cognition, a clear, stable, credible entity perception with time semantics. If your 2026 new technology, new product, or new strategy is merely published but never enters the platform’s trusted time window, then for users it effectively did not happen.
This directly rewrites a core GEO action: time governance.
Content teams used to care more about “whether it was published” and “how fast it was published”; now they must care more about: which content is evergreen fact, which is a version update, which needs updated timestamps, which needs historical archiving, and which must be taken offline, revised, or redirected. Otherwise, the brand’s image in AI will fall into two extremes: either stuck in an old year, or old and new information mixed together and contradicting each other.
3. Doubao: It Moved the “Safety Gate” and “Commercial Gate” Up Front
The third change comes from Doubao.
After 315, what matters most about Doubao is not any specific percentage, but the three recurring signals: first, some consumer/marketing questions are more likely to be refused or answered generically; second, the platform’s interpretation of users’ implicit intent and vague expressions has become more conservative; third, the presence of advertising or commercial exposure in the answer space is rising.
To be clear: these are not officially confirmed product metrics from Doubao, but front-end observation samples.
But even treated only as phenomena, they reveal a new reality: what shapes AI answers today is no longer just “what retrieval can find,” but also “how far risk control allows the answer to go” and “how much answer space commercialization occupies.”
This actually aligns with 315 and the 2026 regulatory logic. In the 315 exposé, the GEO gray industry was directly tied to “manipulating model citations and recommendations through paid placements and poisoning”; the 2026 National Advertising Regulation Work Priorities explicitly lists livestream e-commerce ads, citation ads, and AI-generated ads as key difficulties in internet advertising regulation; another regulatory notice specifically noted that if an ad uses an AI-fabricated persona for recommendation or endorsement without clear disclosure, it may be deemed false advertising; it also imposed stricter requirements on citation ads, misleading large-print-small-print tactics, and absolute claims such as “first/best/leading.”
This means platforms like Doubao being more conservative on consumer decisions, marketing Q&A, and recommendation questions is not accidental but driven by a full set of external pressures:
platforms must avoid being manipulated by “data poisoning” and fake authority, reduce compliance risks from confusing AI recommendations with advertising, and manage their own commercial distribution logic. The result is that some questions become safer, but less useful; some questions become more standardized, but less able to understand what users really want to ask; and in some commercial scenarios, the phenomenon of “answer space being squeezed by commercial placements” will become increasingly obvious.
From a GEO perspective, this change is highly instructive.
Many practitioners used to focus only on “how to get AI to mention me”; now they must add another question: even if my content is retrieved on this platform, can it pass the safety, compliance, and commercial gates?
If not, then the GEO problem is not merely “insufficient optimization”—you simply do not have a license to enter the answer.
4. Looking at the Three Platforms Together, GEO Sees Five New Gates
If the three sets of changes from DeepSeek, Kimi, and Doubao are put together, a new entry mechanism for AI search becomes visible.
The first gate is the retrieval gate.
Can you be found?
The second gate is the evidence gate.
After being found, can you survive multi-source comparison and multi-round verification?
The third gate is the trust time gate.
Is your information trustworthy enough, and is it within the refresh window the platform is willing to include?
The fourth gate is the safety gate.
Can your phrasing, question context, industry risk, and answer content pass platform risk control?
The fifth gate is the commercial gate.
In some scenarios, will the platform prioritize part of the answer space for ads, recommendation slots, partner ecosystems, or its own distribution logic?
Much of past GEO was in fact only fighting at the first gate: as long as AI crawled and mentioned you, it counted as success.
But that is no longer enough. DeepSeek raised the evidence gate, Kimi tightened the trust time gate, and Doubao moved the safety and commercial gates forward. GEO is therefore no longer just “indexing competition,” but has entered a more complex trusted cognition competition.
More notably, platforms giving more citations does not automatically equal a higher density of truth. Research shows that as long as answers contain citations, users’ subjective trust in the answer rises significantly, even if those citations may be random; and the EBU/BBC research reported by Reuters shows that AI assistants still make large-scale errors in news Q&A, with nearly half of answers containing major problems and about one-third having serious source attribution errors. That is, more citations first raise the “feeling of credibility” and do not necessarily raise the “factual accuracy rate.” This is why GEO must move one step beyond “competing to be cited” into “competing to be cited correctly and verifiably.”
5. Five New Trends in GEO
1. From “Single-Round RAG” to “Agentic Search”
The official documentation of both DeepSeek and Kimi is evolving in one direction: multi-step tool calling, complex search reasoning, and the fusion of tool-use and thinking. For GEO, this means the optimization target is no longer a single query hit, but multi-semantic coverage after question decomposition.
2. From “Page Rankings” to “Evidence Chains and Citation Eligibility”
The GEO paper and the 2026 IF-GEO both emphasize one thing: visibility in generative search fundamentally depends on whether content enters the candidate sources, whether it is selected and cited, and whether it stays stable in a multi-query environment. What will be more valuable in the future is the evidence chain, not just the page.
3. From “Chasing Hot Topics With Press Releases” to “Time Governance”
What Kimi reveals is not temporary conservatism, but that “the latest information no longer automatically equals prioritized information.” Brands must begin to manage the temporal semantics of content: versions, changes, archiving, expiration, redirects, and update timestamps. Otherwise, what ages first in AI is not your pages, but your brand cognition.
4. From “Plain-Text Optimization” to “Multimodal Evidence Assets”
Another frequently mentioned judgment about DeepSeek is “video enters the picture.” Even without treating claims such as “Douyin videos are already integrated” as official announcements, it reveals at least one trend: GEO’s evidence assets are expanding from plain text to multimodal. Relevant 2026 research already describes AI-native search as synthesizing multimodal evidence, and incorporates VLMs and agents into the GEO framework. In the future, tutorial, demonstration, step-by-step, and specification videos that can be parsed by ASR, OCR, and visual understanding will look increasingly like “evidence blocks callable by models” rather than mere traffic material.
5. From “Traffic KPIs” to “Answer KPIs”
Bain research shows that 80% of consumers already rely on AI-generated results in at least 40% of their searches, which has taken a 15% to 25% hit on organic traffic. In other words, users are increasingly completing comparison, judgment, and initial screening “before they ever click into your website.” GEO teams should track in the future not just sessions and CTR, but: mention rate, citation rate, first-position answer appearance rate, source-layer coverage, competitor co-occurrence rate, outdated-information rate, refusal rate, ad-slot squeeze rate, and AI-assisted conversion.
6. A Practical Guide for Brands: First Upgrade “Visible” to “Trustworthy and Usable”
First, the official website must upgrade from a display page to a “fact foundation.”
What will be most valuable to AI is not vague brand copy, but pages such as: brand definition, product capabilities, version/pricing, FAQ, case library, comparison pages, specification pages, changelogs, and historical version notes.
Each page should have as much as possible: a one-sentence definition, structured specifications, applicable boundaries, timestamps, and source notes.
This is not for “word count,” but so the model can grab, during multi-round verification, atomic evidence blocks.
Second, do not interpret “the return of the official website” as “third parties are useless.”
There is a judgment that “the official website returns while external endorsements depreciate.” I would revise it more precisely:
What truly depreciates is not high-credibility third-party endorsement, but low-quality, untraceable, mass-produced third-party advertorials.
What truly appreciates is the combination of a brand-owned fact foundation plus high-credibility third-party verification.
Third, upgrade “press-release cadence” to “time governance.”
For new product launches, price changes, version updates, and service changes, do not just issue a press release.
Be sure to sync them to the official fact pages, FAQ, changelog, and version notes, and handle redirects, archiving, and expiration labels for old content.
The biggest noise in AI cognition is often not “no information,” but “multiple versions of information existing at the same time.”
Fourth, treat video as an evidence asset, not just traffic material.
Make the point in the first 3 seconds, keep the spoken structure clear, put key data on screen, ensure audio and visuals agree, and make steps repeatable.
Such videos are friendlier to AI because they are easier to break into structured facts.
Tutorial, comparison, demonstration, and Q&A short videos in particular will increasingly resemble a brand’s “multimodal manual.”
Fifth, high-risk industries should prioritize “strongly traceable content.”
In high-risk fields such as finance, law, healthcare, and education, it is even more necessary to write clearly about qualifications, licenses, standards, regulatory definitions, raw data, test reports, expert bylines, publication times, and applicable boundaries.
The more platforms emphasize risk control, the more they prefer such verifiable materials.
Sixth, stop all “feeding-style” and “black-hat” GEO.
315 has already made this clear: mass press releases, fabricated authority, invented reviews, fake products, and data poisoning are not “smarter optimization,” but an overdraft on a brand’s future credibility and compliance room. Xinhua News Agency’s client and Outlook (Liaowang) have directly linked such practices to “poisoning” and “gray industry chains,” and regulators are also including AI-generated ads and citation ads in priority governance.
7. A Practical Guide for GEO Practitioners: Stop Doing Only “Content Optimization,” Do “Answer Engineering”
1. Build a Query Universe First, Then a Keyword Library
Break core business questions into: definition, price, who it suits, how to choose, pros and cons, cases, risks, comparisons, alternatives, industry standards, and common misconceptions.
The unit of GEO is no longer the keyword, but user tasks and semantic facets.
2. Build an Evidence Layer Map
Audit the brand’s content assets across four layers:
the fact layer, the explanation layer, the proof layer, and the time layer.
You will quickly find that many brands do not really “lack content”; rather, they lack content that AI can use as evidence.
3. Build a Citation Block on Every Page
A one-sentence definition, three to five core conclusions, specifications/data, comparison tables, timestamps, scope, and sources.
Research in 2026 has begun to push GEO from “generic rewriting” toward “citation failure diagnosis” and “structure engineering”: in other words, the future is not about whether you can write, but about why you were not cited and how to fix it in a targeted way. At the same time, structural features themselves affect citation behavior.
4. Do Differentiated GEO by Platform
For DeepSeek, emphasize multi-source evidence, cross-subquery coverage, page structuring, and citation eligibility.
For Kimi, emphasize time governance, authoritative sources, update cadence, and consistency of brand mindshare.
For Doubao, emphasize compliant semantics, authentic user-intent expression in consumer scenarios, and avoiding content that platforms misclassify as advertising, exaggerated, or manipulative.
5. Monitor “Answers,” Not Just “Rankings”
For DeepSeek in particular, the official documentation already states that the API version and the App/Web version are not fully identical.
So GEO’s real observation point must return to the front end:
Is the brand mentioned?
Where does it appear in the answer?
Who vouches for it?
Who does it appear alongside?
Has it been replaced by outdated information?
Is it blocked by the safety gate?
Is it crowded out by ad slots?
These are the metrics closer to business results today.
6. Manage “Cognition Assets,” Not “Content Volume”
The essence of GEO must return to mind management.
Truly effective long-term optimization is not suppressing an answer with large amounts of low-quality content, but getting the brand to form, in AI’s long-term cognition:
who you are, what problem you solve, how you differ from others, what evidence proves you, what the latest version is, and which old information has expired.
Conclusion
If the changes across these three platforms must be condensed into one sentence, my judgment is:
GEO is not over; what is over is “feeding-style GEO.”
DeepSeek tells us that the answer mechanism is moving from single-round retrieval to multi-round verification and evidence orchestration;
Kimi tells us that platforms now put “what to trust first” before “what to ingest first”;
Doubao tells us that safety governance, platform risk control, and commercial distribution have truly entered the answer-generation process.
So the next stage of GEO is no longer just “let me be seen by AI,” but:
let me be found,
let me be trusted,
let me pass the platform’s safety and compliance gates,
and let me enter the answer in a correct, verifiable, and sustainable way.
The truly competitive brands of the future will not be those that issue the most press releases or stack the most keywords, but those that possess clear entity definitions, stable evidence assets, trusted time governance, multimodal explanation capabilities, and cross-platform consistent cognition.
Ultimately, in the AI search era, the competition is not about who is better at “feeding,”
but about who is more worthy of being treated by AI as part of the answer.